Python cuda.select_device方法代码示例

# 需要导入模块: from numba import cuda [as 别名]
# 或者: from numba.cuda import select_device [as 别名]
def rotate_iou_gpu_eval(boxes, query_boxes, criterion=-1, device_id=0):
    """rotated box iou running in gpu. 500x faster than cpu version
    (take 5ms in one example with numba.cuda code).
    convert from [this project](
        https://github.com/hongzhenwang/RRPN-revise/tree/master/lib/rotation).
    
    Args:
        boxes (float tensor: [N, 5]): rbboxes. format: centers, dims, 
            angles(clockwise when positive)
        query_boxes (float tensor: [K, 5]): [description]
        device_id (int, optional): Defaults to 0. [description]
    
    Returns:
        [type]: [description]
    """
    box_dtype = boxes.dtype
    boxes = boxes.astype(np.float32)
    query_boxes = query_boxes.astype(np.float32)
    N = boxes.shape[0]
    K = query_boxes.shape[0]
    iou = np.zeros((N, K), dtype=np.float32)
    if N == 0 or K == 0:
        return iou
    threadsPerBlock = 8 * 8
    cuda.select_device(device_id)
    blockspergrid = (div_up(N, threadsPerBlock), div_up(K, threadsPerBlock))
    
    stream = cuda.stream()
    with stream.auto_synchronize():
        boxes_dev = cuda.to_device(boxes.reshape([-1]), stream)
        query_boxes_dev = cuda.to_device(query_boxes.reshape([-1]), stream)
        iou_dev = cuda.to_device(iou.reshape([-1]), stream)
        rotate_iou_kernel_eval[blockspergrid, threadsPerBlock, stream](
            N, K, boxes_dev, query_boxes_dev, iou_dev, criterion)
        iou_dev.copy_to_host(iou.reshape([-1]), stream=stream)
    return iou.astype(boxes.dtype) 

# 需要导入模块: from numba import cuda [as 别名]
# 或者: from numba.cuda import select_device [as 别名]
def rotate_iou_gpu_eval(boxes, query_boxes, criterion=-1, device_id=0):
    """rotated box iou running in gpu. 500x faster than cpu version
    (take 5ms in one example with numba.cuda code).
    convert from [this project](
        https://github.com/hongzhenwang/RRPN-revise/tree/master/lib/rotation).
    
    Args:
        boxes (float tensor: [N, 5]): rbboxes. format: centers, dims, 
            angles(clockwise when positive)
        query_boxes (float tensor: [K, 5]): [description]
        device_id (int, optional): Defaults to 0. [description]
    
    Returns:
        [type]: [description]
    """
    boxes = boxes.astype(np.float32)
    query_boxes = query_boxes.astype(np.float32)
    N = boxes.shape[0]
    K = query_boxes.shape[0]
    iou = np.zeros((N, K), dtype=np.float32)
    if N == 0 or K == 0:
        return iou
    threadsPerBlock = 8 * 8
    cuda.select_device(device_id)
    blockspergrid = (div_up(N, threadsPerBlock), div_up(K, threadsPerBlock))

    stream = cuda.stream()
    with stream.auto_synchronize():
        boxes_dev = cuda.to_device(boxes.reshape([-1]), stream)
        query_boxes_dev = cuda.to_device(query_boxes.reshape([-1]), stream)
        iou_dev = cuda.to_device(iou.reshape([-1]), stream)
        rotate_iou_kernel_eval[blockspergrid, threadsPerBlock, stream](
            N, K, boxes_dev, query_boxes_dev, iou_dev, criterion)
        iou_dev.copy_to_host(iou.reshape([-1]), stream=stream)
    return iou.astype(boxes.dtype) 

# 需要导入模块: from numba import cuda [as 别名]
# 或者: from numba.cuda import select_device [as 别名]
def rotate_iou_gpu_eval(boxes, query_boxes, criterion=-1, device_id=0):
    """
    rotated box iou running in gpu. 8x faster than cpu version (take 5ms in one example with numba.cuda code).
    convert from [this project](https://github.com/hongzhenwang/RRPN-revise/tree/master/lib/rotation).
    :param boxes: rbboxes, format: centers, dims, angles(clockwise when positive), FloatTensor[N, 5]
    :param query_boxes: FloatTensor[K, 5]
    :param criterion: optional, default: -1
    :param device_id: int, optional, default: 0
    :return:
    """
    boxes = boxes.astype(np.float32)
    query_boxes = query_boxes.astype(np.float32)
    N = boxes.shape[0]
    K = query_boxes.shape[0]
    iou = np.zeros((N, K), dtype=np.float32)
    if N == 0 or K == 0:
        return iou
    threads_per_block = 8 * 8
    cuda.select_device(device_id)
    blocks_per_grid = (div_up(N, threads_per_block), div_up(K, threads_per_block))

    stream = cuda.stream()
    with stream.auto_synchronize():
        boxes_dev = cuda.to_device(boxes.reshape([-1]), stream)
        query_boxes_dev = cuda.to_device(query_boxes.reshape([-1]), stream)
        iou_dev = cuda.to_device(iou.reshape([-1]), stream)
        rotate_iou_kernel_eval[blocks_per_grid, threads_per_block, stream](N, K, boxes_dev, query_boxes_dev,
                                                                           iou_dev, criterion)
        iou_dev.copy_to_host(iou.reshape([-1]), stream=stream)
    return iou.astype(boxes.dtype) 

# 需要导入模块: from numba import cuda [as 别名]
# 或者: from numba.cuda import select_device [as 别名]
def no_gpu():
    """Check for the required GPU dependencies"""
    try:
        from numba import cuda
        import cudf  # noqa

        try:
            cuda.select_device(0)
        except cuda.cudadrv.error.CudaDriverError:
            return True
    except ImportError:
        return True
    return False 

# 需要导入模块: from numba import cuda [as 别名]
# 或者: from numba.cuda import select_device [as 别名]
def nms_gpu(dets, nms_overlap_thresh, device_id=0):
    """nms in gpu. 
    
    Args:
        dets ([type]): [description]
        nms_overlap_thresh ([type]): [description]
        device_id ([type], optional): Defaults to 0. [description]
    
    Returns:
        [type]: [description]
    """

    boxes_num = dets.shape[0]
    keep_out = np.zeros([boxes_num], dtype=np.int32)
    scores = dets[:, 4]
    order = scores.argsort()[::-1].astype(np.int32)
    boxes_host = dets[order, :]

    threadsPerBlock = 8 * 8
    col_blocks = div_up(boxes_num, threadsPerBlock)
    cuda.select_device(device_id)
    mask_host = np.zeros((boxes_num * col_blocks, ), dtype=np.uint64)
    blockspergrid = (div_up(boxes_num, threadsPerBlock),
                     div_up(boxes_num, threadsPerBlock))
    stream = cuda.stream()
    with stream.auto_synchronize():
        boxes_dev = cuda.to_device(boxes_host.reshape([-1]), stream)
        mask_dev = cuda.to_device(mask_host, stream)
        nms_kernel[blockspergrid, threadsPerBlock, stream](
            boxes_num, nms_overlap_thresh, boxes_dev, mask_dev)
        mask_dev.copy_to_host(mask_host, stream=stream)
    # stream.synchronize()
    num_out = nms_postprocess(keep_out, mask_host, boxes_num)
    keep = keep_out[:num_out]
    return list(order[keep]) 

# 需要导入模块: from numba import cuda [as 别名]
# 或者: from numba.cuda import select_device [as 别名]
def rotate_nms_gpu(dets, nms_overlap_thresh, device_id=0):
    """nms in gpu. WARNING: this function can provide right result 
    but its performance isn't be tested
    
    Args:
        dets ([type]): [description]
        nms_overlap_thresh ([type]): [description]
        device_id ([type], optional): Defaults to 0. [description]
    
    Returns:
        [type]: [description]
    """
    dets = dets.astype(np.float32)
    boxes_num = dets.shape[0]
    keep_out = np.zeros([boxes_num], dtype=np.int32)
    scores = dets[:, 5]
    order = scores.argsort()[::-1].astype(np.int32)
    boxes_host = dets[order, :]

    threadsPerBlock = 8 * 8
    col_blocks = div_up(boxes_num, threadsPerBlock)
    cuda.select_device(device_id)
    # mask_host shape: boxes_num * col_blocks * sizeof(np.uint64)
    mask_host = np.zeros((boxes_num * col_blocks, ), dtype=np.uint64)
    blockspergrid = (div_up(boxes_num, threadsPerBlock),
                     div_up(boxes_num, threadsPerBlock))
    stream = cuda.stream()
    with stream.auto_synchronize():
        boxes_dev = cuda.to_device(boxes_host.reshape([-1]), stream)
        mask_dev = cuda.to_device(mask_host, stream)
        rotate_nms_kernel[blockspergrid, threadsPerBlock, stream](
            boxes_num, nms_overlap_thresh, boxes_dev, mask_dev)
        mask_dev.copy_to_host(mask_host, stream=stream)
    num_out = nms_postprocess(keep_out, mask_host, boxes_num)
    keep = keep_out[:num_out]
    return list(order[keep]) 

# 需要导入模块: from numba import cuda [as 别名]
# 或者: from numba.cuda import select_device [as 别名]
def rotate_iou_gpu(boxes, query_boxes, device_id=0):
    """rotated box iou running in gpu. 500x faster than cpu version
    (take 5ms in one example with numba.cuda code).
    convert from [this project](
        https://github.com/hongzhenwang/RRPN-revise/tree/master/lib/rotation).
    
    Args:
        boxes (float tensor: [N, 5]): rbboxes. format: centers, dims, 
            angles(clockwise when positive)
        query_boxes (float tensor: [K, 5]): [description]
        device_id (int, optional): Defaults to 0. [description]
    
    Returns:
        [type]: [description]
    """
    box_dtype = boxes.dtype
    boxes = boxes.astype(np.float32)
    query_boxes = query_boxes.astype(np.float32)
    N = boxes.shape[0]
    K = query_boxes.shape[0]
    iou = np.zeros((N, K), dtype=np.float32)
    if N == 0 or K == 0:
        return iou
    threadsPerBlock = 8 * 8
    cuda.select_device(device_id)
    blockspergrid = (div_up(N, threadsPerBlock), div_up(K, threadsPerBlock))

    stream = cuda.stream()
    with stream.auto_synchronize():
        boxes_dev = cuda.to_device(boxes.reshape([-1]), stream)
        query_boxes_dev = cuda.to_device(query_boxes.reshape([-1]), stream)
        iou_dev = cuda.to_device(iou.reshape([-1]), stream)
        rotate_iou_kernel[blockspergrid, threadsPerBlock, stream](
            N, K, boxes_dev, query_boxes_dev, iou_dev)
        iou_dev.copy_to_host(iou.reshape([-1]), stream=stream)
    return iou.astype(boxes.dtype) 

# 需要导入模块: from numba import cuda [as 别名]
# 或者: from numba.cuda import select_device [as 别名]
def rotate_iou_gpu_eval(boxes, query_boxes, criterion=-1, device_id=0):
    """rotated box iou running in gpu. 8x faster than cpu version
    (take 5ms in one example with numba.cuda code).
    convert from [this project](
        https://github.com/hongzhenwang/RRPN-revise/tree/master/lib/rotation).
    
    Args:
        boxes (float tensor: [N, 5]): rbboxes. format: centers, dims, 
            angles(clockwise when positive)
        query_boxes (float tensor: [K, 5]): [description]
        device_id (int, optional): Defaults to 0. [description]
    
    Returns:
        [type]: [description]
    """
    box_dtype = boxes.dtype
    boxes = boxes.astype(np.float32)
    query_boxes = query_boxes.astype(np.float32)
    N = boxes.shape[0]
    K = query_boxes.shape[0]
    iou = np.zeros((N, K), dtype=np.float32)
    if N == 0 or K == 0:
        return iou
    threadsPerBlock = 8 * 8
    cuda.select_device(device_id)
    blockspergrid = (div_up(N, threadsPerBlock), div_up(K, threadsPerBlock))

    stream = cuda.stream()
    with stream.auto_synchronize():
        boxes_dev = cuda.to_device(boxes.reshape([-1]), stream)
        query_boxes_dev = cuda.to_device(query_boxes.reshape([-1]), stream)
        iou_dev = cuda.to_device(iou.reshape([-1]), stream)
        rotate_iou_kernel_eval[blockspergrid, threadsPerBlock, stream](
            N, K, boxes_dev, query_boxes_dev, iou_dev, criterion)
        iou_dev.copy_to_host(iou.reshape([-1]), stream=stream)
    return iou.astype(boxes.dtype) 

# 需要导入模块: from numba import cuda [as 别名]
# 或者: from numba.cuda import select_device [as 别名]
def rotate_iou_gpu_eval(boxes, query_boxes, criterion=-1, device_id=0):
    """rotated box iou running in gpu. 500x faster than cpu version
    (take 5ms in one example with numba.cuda code).
    convert from [this project](
        https://github.com/hongzhenwang/RRPN-revise/tree/master/lib/rotation).
    
    Args:
        boxes (float tensor: [N, 5]): rbboxes. format: centers, dims, 
            angles(clockwise when positive)
        query_boxes (float tensor: [K, 5]): [description]
        device_id (int, optional): Defaults to 0. [description]
    
    Returns:
        [type]: [description]
    """
    box_dtype = boxes.dtype
    boxes = boxes.astype(np.float32)
    query_boxes = query_boxes.astype(np.float32)
    N = boxes.shape[0]
    K = query_boxes.shape[0]
    iou = np.zeros((N, K), dtype=np.float32)
    if N == 0 or K == 0:
        return iou
    threadsPerBlock = 8 * 8
    cuda.select_device(device_id)
    blockspergrid = (div_up(N, threadsPerBlock), div_up(K, threadsPerBlock))

    stream = cuda.stream()
    with stream.auto_synchronize():
        boxes_dev = cuda.to_device(boxes.reshape([-1]), stream)
        query_boxes_dev = cuda.to_device(query_boxes.reshape([-1]), stream)
        iou_dev = cuda.to_device(iou.reshape([-1]), stream)
        rotate_iou_kernel_eval[blockspergrid, threadsPerBlock, stream](
            N, K, boxes_dev, query_boxes_dev, iou_dev, criterion)
        iou_dev.copy_to_host(iou.reshape([-1]), stream=stream)
    return iou.astype(boxes.dtype) 

# 需要导入模块: from numba import cuda [as 别名]
# 或者: from numba.cuda import select_device [as 别名]
def nms_gpu(dets, nms_overlap_thresh, device_id=0):
    """nms in gpu.

    Args:
        dets ([type]): [description]
        nms_overlap_thresh ([type]): [description]
        device_id ([type], optional): Defaults to 0. [description]

    Returns:
        [type]: [description]
    """

    boxes_num = dets.shape[0]
    keep_out = np.zeros([boxes_num], dtype=np.int32)
    scores = dets[:, 4]
    order = scores.argsort()[::-1].astype(np.int32)
    boxes_host = dets[order, :]

    threadsPerBlock = 8 * 8
    col_blocks = div_up(boxes_num, threadsPerBlock)
    cuda.select_device(device_id)
    mask_host = np.zeros((boxes_num * col_blocks,), dtype=np.uint64)
    blockspergrid = (
        div_up(boxes_num, threadsPerBlock),
        div_up(boxes_num, threadsPerBlock),
    )
    stream = cuda.stream()
    with stream.auto_synchronize():
        boxes_dev = cuda.to_device(boxes_host.reshape([-1]), stream)
        mask_dev = cuda.to_device(mask_host, stream)
        nms_kernel[blockspergrid, threadsPerBlock, stream](
            boxes_num, nms_overlap_thresh, boxes_dev, mask_dev
        )
        mask_dev.copy_to_host(mask_host, stream=stream)
    # stream.synchronize()
    num_out = nms_postprocess(keep_out, mask_host, boxes_num)
    keep = keep_out[:num_out]
    return list(order[keep]) 

# 需要导入模块: from numba import cuda [as 别名]
# 或者: from numba.cuda import select_device [as 别名]
def rotate_nms_gpu(dets, nms_overlap_thresh, device_id=0):
    """nms in gpu. WARNING: this function can provide right result
    but its performance isn't be tested

    Args:
        dets ([type]): [description]
        nms_overlap_thresh ([type]): [description]
        device_id ([type], optional): Defaults to 0. [description]

    Returns:
        [type]: [description]
    """
    dets = dets.astype(np.float32)
    boxes_num = dets.shape[0]
    keep_out = np.zeros([boxes_num], dtype=np.int32)
    scores = dets[:, 5]
    order = scores.argsort()[::-1].astype(np.int32)
    boxes_host = dets[order, :]

    threadsPerBlock = 8 * 8
    col_blocks = div_up(boxes_num, threadsPerBlock)
    cuda.select_device(device_id)
    # mask_host shape: boxes_num * col_blocks * sizeof(np.uint64)
    mask_host = np.zeros((boxes_num * col_blocks,), dtype=np.uint64)
    blockspergrid = (
        div_up(boxes_num, threadsPerBlock),
        div_up(boxes_num, threadsPerBlock),
    )
    stream = cuda.stream()
    with stream.auto_synchronize():
        boxes_dev = cuda.to_device(boxes_host.reshape([-1]), stream)
        mask_dev = cuda.to_device(mask_host, stream)
        rotate_nms_kernel[blockspergrid, threadsPerBlock, stream](
            boxes_num, nms_overlap_thresh, boxes_dev, mask_dev
        )
        mask_dev.copy_to_host(mask_host, stream=stream)
    num_out = nms_postprocess(keep_out, mask_host, boxes_num)
    keep = keep_out[:num_out]
    return list(order[keep]) 

# 需要导入模块: from numba import cuda [as 别名]
# 或者: from numba.cuda import select_device [as 别名]
def rotate_iou_gpu(boxes, query_boxes, device_id=0):
    """rotated box iou running in gpu. 500x faster than cpu version
    (take 5ms in one example with numba.cuda code).
    convert from [this project](
        https://github.com/hongzhenwang/RRPN-revise/tree/master/lib/rotation).

    Args:
        boxes (float tensor: [N, 5]): rbboxes. format: centers, dims,
            angles(clockwise when positive)
        query_boxes (float tensor: [K, 5]): [description]
        device_id (int, optional): Defaults to 0. [description]

    Returns:
        [type]: [description]
    """
    box_dtype = boxes.dtype
    boxes = boxes.astype(np.float32)
    query_boxes = query_boxes.astype(np.float32)
    N = boxes.shape[0]
    K = query_boxes.shape[0]
    iou = np.zeros((N, K), dtype=np.float32)
    if N == 0 or K == 0:
        return iou
    threadsPerBlock = 8 * 8
    cuda.select_device(device_id)
    blockspergrid = (div_up(N, threadsPerBlock), div_up(K, threadsPerBlock))

    stream = cuda.stream()
    with stream.auto_synchronize():
        boxes_dev = cuda.to_device(boxes.reshape([-1]), stream)
        query_boxes_dev = cuda.to_device(query_boxes.reshape([-1]), stream)
        iou_dev = cuda.to_device(iou.reshape([-1]), stream)
        rotate_iou_kernel[blockspergrid, threadsPerBlock, stream](
            N, K, boxes_dev, query_boxes_dev, iou_dev
        )
        iou_dev.copy_to_host(iou.reshape([-1]), stream=stream)
    return iou.astype(boxes.dtype) 

# 需要导入模块: from numba import cuda [as 别名]
# 或者: from numba.cuda import select_device [as 别名]
def rotate_iou_gpu_eval(boxes, query_boxes, criterion=-1, device_id=0):
    """rotated box iou running in gpu. 500x faster than cpu version
    (take 5ms in one example with numba.cuda code).
    convert from [this project](
        https://github.com/hongzhenwang/RRPN-revise/tree/master/lib/rotation).

    Args:
        boxes (float tensor: [N, 5]): rbboxes. format: centers, dims,
            angles(clockwise when positive)
        query_boxes (float tensor: [K, 5]): [description]
        device_id (int, optional): Defaults to 0. [description]

    Returns:
        [type]: [description]
    """
    box_dtype = boxes.dtype
    boxes = boxes.astype(np.float32)
    query_boxes = query_boxes.astype(np.float32)
    N = boxes.shape[0]
    K = query_boxes.shape[0]
    iou = np.zeros((N, K), dtype=np.float32)
    if N == 0 or K == 0:
        return iou
    threadsPerBlock = 8 * 8
    cuda.select_device(device_id)
    blockspergrid = (div_up(N, threadsPerBlock), div_up(K, threadsPerBlock))

    stream = cuda.stream()
    with stream.auto_synchronize():
        boxes_dev = cuda.to_device(boxes.reshape([-1]), stream)
        query_boxes_dev = cuda.to_device(query_boxes.reshape([-1]), stream)
        iou_dev = cuda.to_device(iou.reshape([-1]), stream)
        rotate_iou_kernel_eval[blockspergrid, threadsPerBlock, stream](
            N, K, boxes_dev, query_boxes_dev, iou_dev, criterion
        )
        iou_dev.copy_to_host(iou.reshape([-1]), stream=stream)
    return iou.astype(boxes.dtype) 

# 需要导入模块: from numba import cuda [as 别名]
# 或者: from numba.cuda import select_device [as 别名]
def rotate_iou_gpu_eval(boxes, query_boxes, criterion=-1, device_id=0):
    """rotated box iou running in gpu. 500x faster than cpu version
    (take 5ms in one example with numba.cuda code).
    convert from [this project](
        https://github.com/hongzhenwang/RRPN-revise/tree/master/lib/rotation).
    
    Args:
        boxes (float tensor: [N, 5]): rbboxes. format: centers, dims, 
            angles(clockwise when positive)
        query_boxes (float tensor: [K, 5]): [description]
        device_id (int, optional): Defaults to 0. [description]
    
    Returns:
        [type]: [description]
    """
    box_dtype = boxes.dtype
    boxes = boxes.astype(np.float32)
    query_boxes = query_boxes.astype(np.float32)
    N = boxes.shape[0]
    K = query_boxes.shape[0]
    iou = np.zeros((N, K), dtype=np.float32)
    if N == 0 or K == 0:
        return iou
    threadsPerBlock = 8 * 8
    cuda.select_device(device_id)
    blockspergrid = (div_up(N, threadsPerBlock), div_up(K, threadsPerBlock))

    stream = cuda.stream()
    with stream.auto_synchronize():
        boxes_dev = cuda.to_device(boxes.reshape([-1]), stream)
        query_boxes_dev = cuda.to_device(query_boxes.reshape([-1]), stream)
        iou_dev = cuda.to_device(iou.reshape([-1]), stream)
        rotate_iou_kernel_eval[blockspergrid, threadsPerBlock, stream](
            N, K, boxes_dev, query_boxes_dev, iou_dev, criterion
        )
        iou_dev.copy_to_host(iou.reshape([-1]), stream=stream)
    return iou.astype(boxes.dtype)