warp-lang 1.9.1__py3-none-manylinux_2_34_aarch64.whl → 1.10.0rc2__py3-none-manylinux_2_34_aarch64.whl

warp/_src/dlpack.py

@@ -0,0 +1,462 @@
+# SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Python specification for DLpack:
+# https://dmlc.github.io/dlpack/latest/python_spec.html
+
+import ctypes
+
+import warp
+from warp._src.thirdparty.dlpack import (
+    DLDataType,
+    DLDataTypeCode,
+    DLDevice,
+    DLDeviceType,
+    DLManagedTensor,
+    _c_str_dltensor,
+)
+
+_c_str_used_dltensor = b"used_dltensor"
+
+PyMem_RawMalloc = ctypes.pythonapi.PyMem_RawMalloc
+PyMem_RawMalloc.argtypes = [ctypes.c_size_t]
+PyMem_RawMalloc.restype = ctypes.c_void_p
+
+PyMem_RawFree = ctypes.pythonapi.PyMem_RawFree
+PyMem_RawFree.argtypes = [ctypes.c_void_p]
+PyMem_RawFree.restype = None
+
+Py_IncRef = ctypes.pythonapi.Py_IncRef
+Py_IncRef.argtypes = [ctypes.py_object]
+Py_IncRef.restype = None
+
+Py_DecRef = ctypes.pythonapi.Py_DecRef
+Py_DecRef.argtypes = [ctypes.py_object]
+Py_DecRef.restype = None
+
+PyCapsule_Destructor = ctypes.CFUNCTYPE(None, ctypes.c_void_p)
+
+PyCapsule_IsValid = ctypes.pythonapi.PyCapsule_IsValid
+PyCapsule_IsValid.argtypes = [ctypes.py_object, ctypes.c_char_p]
+PyCapsule_IsValid.restype = ctypes.c_int
+
+PyCapsule_GetPointer = ctypes.pythonapi.PyCapsule_GetPointer
+PyCapsule_GetPointer.argtypes = [ctypes.py_object, ctypes.c_char_p]
+PyCapsule_GetPointer.restype = ctypes.c_void_p
+
+PyCapsule_SetName = ctypes.pythonapi.PyCapsule_SetName
+PyCapsule_SetName.argtypes = [ctypes.py_object, ctypes.c_char_p]
+PyCapsule_SetName.restype = ctypes.c_int
+
+
+class _DLPackTensorHolder:
+    """Class responsible for deleting DLManagedTensor memory after ownership is transferred from a capsule."""
+
+    def __new__(cls, *args, **kwargs):
+        instance = super().__new__(cls)
+        instance.mem_ptr = None
+        return instance
+
+    def __init__(self, mem_ptr):
+        self.mem_ptr = mem_ptr
+
+    def __del__(self):
+        if not self.mem_ptr:
+            return
+
+        managed_tensor = DLManagedTensor.from_address(self.mem_ptr)
+        if managed_tensor.deleter:
+            managed_tensor.deleter(self.mem_ptr)
+
+
+@ctypes.CFUNCTYPE(None, ctypes.c_void_p)
+def _dlpack_tensor_deleter(managed_ptr) -> None:
+    """A function to deallocate a DLManagedTensor."""
+
+    managed_tensor = DLManagedTensor.from_address(managed_ptr)
+
+    # unreference the source array
+    manager = ctypes.cast(managed_tensor.manager_ctx, ctypes.py_object)
+    ctypes.pythonapi.Py_DecRef(manager)
+
+    # free the DLManagedTensor memory, including shape and strides
+    PyMem_RawFree(ctypes.c_void_p(managed_ptr))
+
+
+@PyCapsule_Destructor
+def _dlpack_capsule_deleter(ptr) -> None:
+    """Destructor for a capsule holding a DLManagedTensor."""
+
+    capsule = ctypes.cast(ptr, ctypes.py_object)
+
+    if PyCapsule_IsValid(capsule, _c_str_dltensor):
+        managed_ptr = PyCapsule_GetPointer(capsule, _c_str_dltensor)
+        managed_tensor = DLManagedTensor.from_address(managed_ptr)
+        if managed_tensor.deleter:
+            managed_tensor.deleter(managed_ptr)
+
+
+def _device_to_dlpack(wp_device: warp._src.context.Device) -> DLDevice:
+    dl_device = DLDevice()
+
+    if wp_device.is_cpu:
+        dl_device.device_type = DLDeviceType.kDLCPU
+        dl_device.device_id = 0
+    elif wp_device.is_cuda:
+        dl_device.device_type = DLDeviceType.kDLCUDA
+        dl_device.device_id = wp_device.ordinal
+    else:
+        raise RuntimeError(f"Invalid device type converting to DLPack: {wp_device}")
+
+    return dl_device
+
+
+def device_to_dlpack(wp_device) -> DLDevice:
+    return _device_to_dlpack(warp.get_device(wp_device))
+
+
+def dtype_to_dlpack(wp_dtype) -> DLDataType:
+    if wp_dtype == warp.bool:
+        return (DLDataTypeCode.kDLBool, 8, 1)
+    if wp_dtype == warp.int8:
+        return (DLDataTypeCode.kDLInt, 8, 1)
+    elif wp_dtype == warp.uint8:
+        return (DLDataTypeCode.kDLUInt, 8, 1)
+    elif wp_dtype == warp.int16:
+        return (DLDataTypeCode.kDLInt, 16, 1)
+    elif wp_dtype == warp.uint16:
+        return (DLDataTypeCode.kDLUInt, 16, 1)
+    elif wp_dtype == warp.int32:
+        return (DLDataTypeCode.kDLInt, 32, 1)
+    elif wp_dtype == warp.uint32:
+        return (DLDataTypeCode.kDLUInt, 32, 1)
+    elif wp_dtype == warp.int64:
+        return (DLDataTypeCode.kDLInt, 64, 1)
+    elif wp_dtype == warp.uint64:
+        return (DLDataTypeCode.kDLUInt, 64, 1)
+    elif wp_dtype == warp.float16:
+        return (DLDataTypeCode.kDLFloat, 16, 1)
+    elif wp_dtype == warp.float32:
+        return (DLDataTypeCode.kDLFloat, 32, 1)
+    elif wp_dtype == warp.float64:
+        return (DLDataTypeCode.kDLFloat, 64, 1)
+    else:
+        raise RuntimeError(f"No conversion from Warp type {wp_dtype} to DLPack type")
+
+
+def dtype_from_dlpack(dl_dtype):
+    # unpack to tuple for easier comparison
+    dl_dtype = (dl_dtype.type_code.value, dl_dtype.bits)
+
+    if dl_dtype == (DLDataTypeCode.kDLUInt, 1):
+        raise RuntimeError("Warp does not support bit boolean types")
+    elif dl_dtype == (DLDataTypeCode.kDLInt, 8):
+        return warp._src.types.int8
+    elif dl_dtype == (DLDataTypeCode.kDLInt, 16):
+        return warp._src.types.int16
+    elif dl_dtype == (DLDataTypeCode.kDLInt, 32):
+        return warp._src.types.int32
+    elif dl_dtype == (DLDataTypeCode.kDLInt, 64):
+        return warp._src.types.int64
+    elif dl_dtype == (DLDataTypeCode.kDLUInt, 8):
+        return warp._src.types.uint8
+    elif dl_dtype == (DLDataTypeCode.kDLUInt, 16):
+        return warp._src.types.uint16
+    elif dl_dtype == (DLDataTypeCode.kDLUInt, 32):
+        return warp._src.types.uint32
+    elif dl_dtype == (DLDataTypeCode.kDLUInt, 64):
+        return warp._src.types.uint64
+    elif dl_dtype == (DLDataTypeCode.kDLFloat, 16):
+        return warp._src.types.float16
+    elif dl_dtype == (DLDataTypeCode.kDLFloat, 32):
+        return warp._src.types.float32
+    elif dl_dtype == (DLDataTypeCode.kDLFloat, 64):
+        return warp._src.types.float64
+    elif dl_dtype == (DLDataTypeCode.kDLComplex, 64):
+        raise RuntimeError("Warp does not support complex types")
+    elif dl_dtype == (DLDataTypeCode.kDLComplex, 128):
+        raise RuntimeError("Warp does not support complex types")
+    else:
+        raise RuntimeError(f"Unknown DLPack datatype {dl_dtype}")
+
+
+def device_from_dlpack(dl_device):
+    assert warp._src.context.runtime is not None, "Warp not initialized, call wp.init() before use"
+
+    if dl_device.device_type.value == DLDeviceType.kDLCPU or dl_device.device_type.value == DLDeviceType.kDLCUDAHost:
+        return warp._src.context.runtime.cpu_device
+    elif (
+        dl_device.device_type.value == DLDeviceType.kDLCUDA
+        or dl_device.device_type.value == DLDeviceType.kDLCUDAManaged
+    ):
+        return warp._src.context.runtime.cuda_devices[dl_device.device_id]
+    else:
+        raise RuntimeError(f"Unknown device type from DLPack: {dl_device.device_type.value}")
+
+
+def shape_to_dlpack(shape):
+    a = (ctypes.c_int64 * len(shape))(*shape)
+    return a
+
+
+def strides_to_dlpack(strides, dtype):
+    # convert from byte count to element count
+    ndim = len(strides)
+    a = (ctypes.c_int64 * ndim)()
+    dtype_size = warp._src.types.type_size_in_bytes(dtype)
+    for i in range(ndim):
+        a[i] = strides[i] // dtype_size
+    return a
+
+
+def to_dlpack(wp_array: warp.array):
+    """Convert a Warp array to another type of DLPack-compatible array.
+
+    Args:
+        wp_array: The source Warp array that will be converted.
+
+    Returns:
+        A capsule containing a DLManagedTensor that can be converted
+        to another array type without copying the underlying memory.
+    """
+
+    # DLPack does not support structured arrays
+    if isinstance(wp_array.dtype, warp._src.codegen.Struct):
+        raise RuntimeError("Cannot convert structured Warp arrays to DLPack.")
+
+    # handle vector types
+    if hasattr(wp_array.dtype, "_wp_scalar_type_"):
+        # vector type, flatten the dimensions into one tuple
+        target_dtype = wp_array.dtype._wp_scalar_type_
+        target_ndim = wp_array.ndim + len(wp_array.dtype._shape_)
+        target_shape = (*wp_array.shape, *wp_array.dtype._shape_)
+        dtype_strides = warp._src.types.strides_from_shape(wp_array.dtype._shape_, wp_array.dtype._wp_scalar_type_)
+        target_strides = (*wp_array.strides, *dtype_strides)
+    else:
+        # scalar type
+        target_dtype = wp_array.dtype
+        target_ndim = wp_array.ndim
+        target_shape = wp_array.shape
+        target_strides = wp_array.strides
+
+    if wp_array.pinned:
+        dl_device = DLDevice()
+        dl_device.device_type = DLDeviceType.kDLCUDAHost
+        dl_device.device_id = 0
+    else:
+        dl_device = _device_to_dlpack(wp_array.device)
+
+    # allocate DLManagedTensor, shape, and strides together
+    managed_tensor_size = ctypes.sizeof(DLManagedTensor)
+    padding = managed_tensor_size & 7
+    shape_size = target_ndim * 8
+    mem_size = managed_tensor_size + padding + 2 * shape_size
+    mem_ptr = PyMem_RawMalloc(mem_size)
+    assert mem_ptr, "Failed to allocate memory for DLManagedTensor"
+
+    # set managed tensor attributes
+    managed_tensor = DLManagedTensor.from_address(mem_ptr)
+    managed_tensor.dl_tensor.data = wp_array.ptr
+    managed_tensor.dl_tensor.device = dl_device
+    managed_tensor.dl_tensor.ndim = target_ndim
+    managed_tensor.dl_tensor.dtype = dtype_to_dlpack(target_dtype)
+    managed_tensor.dl_tensor.byte_offset = 0
+
+    # shape
+    shape_offset = managed_tensor_size + padding
+    shape_ptr = ctypes.cast(mem_ptr + shape_offset, ctypes.POINTER(ctypes.c_int64))
+    for i in range(target_ndim):
+        shape_ptr[i] = target_shape[i]
+    managed_tensor.dl_tensor.shape = shape_ptr
+
+    # strides, if not contiguous
+    if wp_array.is_contiguous:
+        managed_tensor.dl_tensor.strides = None
+    else:
+        stride_offset = shape_offset + shape_size
+        stride_ptr = ctypes.cast(mem_ptr + stride_offset, ctypes.POINTER(ctypes.c_int64))
+        dtype_size = warp._src.types.type_size_in_bytes(target_dtype)
+        for i in range(target_ndim):
+            stride_ptr[i] = target_strides[i] // dtype_size
+        managed_tensor.dl_tensor.strides = stride_ptr
+
+    # DLManagedTensor holds a reference to the source array
+    managed_tensor.manager_ctx = id(wp_array)
+    Py_IncRef(wp_array)
+
+    managed_tensor.deleter = _dlpack_tensor_deleter
+
+    # NOTE: jax.ffi.pycapsule() defines the PyCapsule_New() argtypes incorrectly, which causes problems.
+    # Here we make sure that the PyCapsule_Destructor callback is correctly defined.
+    PyCapsule_New = ctypes.pythonapi.PyCapsule_New
+    PyCapsule_New.argtypes = [ctypes.c_void_p, ctypes.c_char_p, PyCapsule_Destructor]
+    PyCapsule_New.restype = ctypes.py_object
+
+    capsule = PyCapsule_New(
+        ctypes.byref(managed_tensor),
+        _c_str_dltensor,
+        _dlpack_capsule_deleter,
+    )
+
+    return capsule
+
+
+def dtype_is_compatible(dl_dtype, wp_dtype):
+    if dl_dtype.bits % 8 != 0:
+        raise RuntimeError("Data types with less than 8 bits are not supported")
+
+    if dl_dtype.type_code.value == DLDataTypeCode.kDLFloat:
+        if dl_dtype.bits == 16:
+            return wp_dtype == warp.float16
+        elif dl_dtype.bits == 32:
+            return wp_dtype == warp.float32
+        elif dl_dtype.bits == 64:
+            return wp_dtype == warp.float64
+    elif dl_dtype.type_code.value == DLDataTypeCode.kDLInt or dl_dtype.type_code.value == DLDataTypeCode.kDLUInt:
+        if dl_dtype.bits == 8:
+            return wp_dtype == warp.int8 or wp_dtype == warp.uint8
+        elif dl_dtype.bits == 16:
+            return wp_dtype == warp.int16 or wp_dtype == warp.uint16
+        elif dl_dtype.bits == 32:
+            return wp_dtype == warp.int32 or wp_dtype == warp.uint32
+        elif dl_dtype.bits == 64:
+            return wp_dtype == warp.int64 or wp_dtype == warp.uint64
+    elif dl_dtype.type_code.value == DLDataTypeCode.kDLBfloat:
+        raise RuntimeError("Bfloat data type is not supported")
+    elif dl_dtype.type_code.value == DLDataTypeCode.kDLComplex:
+        raise RuntimeError("Complex data types are not supported")
+    else:
+        raise RuntimeError(f"Unsupported DLPack dtype {(str(dl_dtype.type_code), dl_dtype.bits)}")
+
+
+def _from_dlpack(capsule, dtype=None) -> warp.array:
+    """Convert a DLPack capsule into a Warp array without copying.
+
+    Args:
+        capsule: A DLPack capsule wrapping an external array or tensor.
+        dtype: An optional Warp data type to interpret the source data.
+
+    Returns:
+        A new Warp array that uses the same underlying memory as the input capsule.
+    """
+
+    assert PyCapsule_IsValid(capsule, _c_str_dltensor), "Invalid capsule"
+    mem_ptr = PyCapsule_GetPointer(capsule, _c_str_dltensor)
+    managed_tensor = DLManagedTensor.from_address(mem_ptr)
+
+    dlt = managed_tensor.dl_tensor
+
+    device = device_from_dlpack(dlt.device)
+    pinned = dlt.device.device_type.value == DLDeviceType.kDLCUDAHost
+    shape = tuple(dlt.shape[dim] for dim in range(dlt.ndim))
+
+    # strides, if not contiguous
+    itemsize = dlt.dtype.bits // 8
+    if dlt.strides:
+        strides = tuple(dlt.strides[dim] * itemsize for dim in range(dlt.ndim))
+    else:
+        strides = None
+
+    # handle multi-lane dtypes as another dimension
+    if dlt.dtype.lanes > 1:
+        shape = (*shape, dlt.dtype.lanes)
+        if strides is not None:
+            strides = (*strides, itemsize)
+
+    if dtype is None:
+        # automatically detect dtype
+        dtype = dtype_from_dlpack(dlt.dtype)
+
+    elif hasattr(dtype, "_wp_scalar_type_"):
+        # handle vector/matrix types
+
+        if not dtype_is_compatible(dlt.dtype, dtype._wp_scalar_type_):
+            raise RuntimeError(f"Incompatible data types: {dlt.dtype} and {dtype}")
+
+        dtype_shape = dtype._shape_
+        dtype_dims = len(dtype._shape_)
+        if dtype_dims > len(shape) or dtype_shape != shape[-dtype_dims:]:
+            raise RuntimeError(
+                f"Could not convert DLPack tensor with shape {shape} to Warp array with dtype={dtype}, ensure that source inner shape is {dtype_shape}"
+            )
+
+        if strides is not None:
+            # ensure the inner strides are contiguous
+            stride = itemsize
+            for i in range(dtype_dims):
+                if strides[-i - 1] != stride:
+                    raise RuntimeError(
+                        f"Could not convert DLPack tensor with shape {shape} to Warp array with dtype={dtype}, because the source inner strides are not contiguous"
+                    )
+                stride *= dtype_shape[-i - 1]
+            strides = tuple(strides[:-dtype_dims]) or (itemsize,)
+
+        shape = tuple(shape[:-dtype_dims]) or (1,)
+
+    elif not dtype_is_compatible(dlt.dtype, dtype):
+        # incompatible dtype requested
+        raise RuntimeError(f"Incompatible data types: {dlt.dtype} and {dtype}")
+
+    a = warp._src.types.array(
+        ptr=dlt.data, dtype=dtype, shape=shape, strides=strides, copy=False, device=device, pinned=pinned
+    )
+
+    # take ownership of the DLManagedTensor
+    a._dlpack_tensor_holder = _DLPackTensorHolder(mem_ptr)
+
+    # rename the capsule so that it no longer owns the DLManagedTensor
+    PyCapsule_SetName(capsule, _c_str_used_dltensor)
+
+    return a
+
+
+def from_dlpack(source, dtype=None) -> warp.array:
+    """Convert a source array or DLPack capsule into a Warp array without copying.
+
+    Args:
+        source: A DLPack-compatible array or PyCapsule
+        dtype: An optional Warp data type to interpret the source data.
+
+    Returns:
+        A new Warp array that uses the same underlying memory as the input
+        pycapsule.
+    """
+
+    # See https://data-apis.org/array-api/2022.12/API_specification/generated/array_api.array.__dlpack__.html
+
+    if hasattr(source, "__dlpack__"):
+        device_type, device_id = source.__dlpack_device__()
+        # Check if the source lives on a CUDA device
+        if device_type in (DLDeviceType.kDLCUDA, DLDeviceType.kDLCUDAManaged):
+            # Assume that the caller will use the array on its device's current stream.
+            # Note that we pass 1 for the null stream, per DLPack spec.
+            cuda_stream = warp.get_cuda_device(device_id).stream.cuda_stream or 1
+        elif device_type == DLDeviceType.kDLCPU:
+            # No stream sync for CPU arrays.
+            cuda_stream = None
+        elif device_type == DLDeviceType.kDLCUDAHost:
+            # For pinned memory, we sync with the current CUDA device's stream.
+            # Note that we pass 1 for the null stream, per DLPack spec.
+            cuda_stream = warp.get_cuda_device().stream.cuda_stream or 1
+        else:
+            raise TypeError("Unsupported source device")
+
+        capsule = source.__dlpack__(stream=cuda_stream)
+
+    else:
+        # legacy behaviour, assume source is a capsule
+        capsule = source
+
+    return _from_dlpack(capsule, dtype=dtype)