Bohan Hou
859498dc01
## Summary This PR adds the initial TIRx support needed for low-level programming of Blackwell-class GPU architectures. As part of the ongoing TIRx refactor, it introduces TVMScript support for directly scripting advanced hardware features without relying on scheduling as the primary programming interface. The change keeps existing `s_tir` script support intact while making direct scripting a first-class path for TIRx programs. ## Main Changes - Add TIRx operator dispatch and layout infrastructure. - Add TVMScript support for new low-level TIRx operations. - Add analysis, transform, and lowering support for TIRx IR nodes. - Add CUDA/Blackwell-oriented codegen and intrinsic coverage. - Add Python and C++ integration points for TIRx scripting and runtime support. ## Validation - `pre-commit run --all-files` - `ninja -C build -j32` - `CUDA_VISIBLE_DEVICES=2 pytest tests/python/tirx/ -n 16` - `1723 passed, 47 skipped, 32 warnings` - `CUDA_VISIBLE_DEVICES=2 python -m pytest -v tests/python/all-platform-minimal-test` - `37 passed, 105 skipped` - `TVM_TEST_TARGETS=llvm python -m pytest -v tests/python/tirx-analysis tests/python/tirx-base tests/python/tirx-transform -n 16` - `664 passed, 25 skipped, 9 xfailed, 1 xpassed` ## Local CI Notes Some full CI-equivalent jobs were not locally reproducible because this machine is missing parts of the Apache TVM CI environment, including `llvm-config-15/17`, Vulkan, ROCm, Maven, Sphinx, Doxygen, Emscripten, and ARM/QEMU cross-toolchain components. Metal-specific tests were skipped locally because no Metal runtime is available.
523 lines
14 KiB
Python
523 lines
14 KiB
Python
# Licensed to the Apache Software Foundation (ASF) under one
|
|
# or more contributor license agreements. See the NOTICE file
|
|
# distributed with this work for additional information
|
|
# regarding copyright ownership. The ASF licenses this file
|
|
# to you 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.
|
|
# pylint: disable=invalid-name, unused-import, redefined-outer-name
|
|
# ruff: noqa: F401, RUF005
|
|
"""Runtime Tensor API"""
|
|
|
|
import ctypes
|
|
import warnings
|
|
from typing import Optional
|
|
|
|
import numpy as np
|
|
|
|
try:
|
|
import ml_dtypes
|
|
except ImportError:
|
|
ml_dtypes = None
|
|
|
|
import tvm_ffi
|
|
from tvm_ffi import DLDeviceType, device
|
|
|
|
import tvm
|
|
from tvm.runtime import Device
|
|
|
|
from . import _ffi_api
|
|
|
|
|
|
def from_dlpack(ext_tensor):
|
|
"""
|
|
Convert an external tensor to an Tensor.
|
|
|
|
Parameters
|
|
----------
|
|
ext_tensor : object
|
|
The external tensor to convert.
|
|
|
|
require_alignment : int
|
|
The minimum required alignment to check for the tensor.
|
|
|
|
require_contiguous : bool
|
|
Whether to check for contiguous memory.
|
|
"""
|
|
# TODO(tvm-team): change to require_alignment=0 and require_contiguous=False
|
|
# once we update the compiler generated code to guard against misaligned access.
|
|
return tvm_ffi.from_dlpack(
|
|
ext_tensor,
|
|
require_alignment=64,
|
|
require_contiguous=True,
|
|
)
|
|
|
|
|
|
@tvm_ffi.register_object("ffi.Tensor")
|
|
class Tensor(tvm_ffi.core.Tensor):
|
|
"""Lightweight Tensor class of TVM runtime.
|
|
|
|
Strictly this is only an Array Container (a buffer object)
|
|
No arthimetic operations are defined.
|
|
All operations are performed by TVM functions.
|
|
|
|
The goal is not to re-build yet another array library.
|
|
Instead, this is a minimal data structure to demonstrate
|
|
how can we use TVM in existing project which might have their own array containers.
|
|
"""
|
|
|
|
def __setitem__(self, in_slice, value):
|
|
"""Set ndarray value"""
|
|
if (
|
|
not isinstance(in_slice, slice)
|
|
or in_slice.start is not None
|
|
or in_slice.stop is not None
|
|
):
|
|
raise ValueError("Array only support set from numpy array")
|
|
if isinstance(value, Tensor):
|
|
if not value.same_as(self):
|
|
value.copyto(self)
|
|
elif isinstance(value, np.ndarray | np.generic):
|
|
self.copyfrom(value)
|
|
else:
|
|
raise TypeError(f"type {type(value)} not supported")
|
|
|
|
def copyfrom(self, source_array):
|
|
"""Perform a synchronous copy from the array.
|
|
|
|
Parameters
|
|
----------
|
|
source_array : array_like
|
|
The data source we should like to copy from.
|
|
|
|
Returns
|
|
-------
|
|
arr : Tensor
|
|
Reference to self.
|
|
"""
|
|
if isinstance(source_array, Tensor):
|
|
source_array.copyto(self)
|
|
return self
|
|
|
|
if not isinstance(source_array, np.ndarray):
|
|
try:
|
|
source_array = np.array(source_array, dtype=self.dtype)
|
|
except Exception:
|
|
raise TypeError(
|
|
f"array must be an array_like data, type {type(source_array)} is not supported"
|
|
)
|
|
|
|
t = tvm_ffi.dtype(self.dtype)
|
|
shape, dtype = self.shape, self.dtype
|
|
if t.lanes > 1:
|
|
shape = shape + (t.lanes,)
|
|
t = t.with_lanes(1)
|
|
dtype = str(t)
|
|
|
|
if source_array.shape != shape:
|
|
raise ValueError(
|
|
f"array shape do not match the shape of Tensor {source_array.shape} vs {shape}"
|
|
)
|
|
numpy_str_map = tvm_ffi.dtype._NUMPY_DTYPE_TO_STR
|
|
np_dtype_str = (
|
|
numpy_str_map[source_array.dtype]
|
|
if source_array.dtype in numpy_str_map
|
|
else str(source_array.dtype)
|
|
)
|
|
if (not source_array.flags["C_CONTIGUOUS"]) or (
|
|
dtype == "bfloat16" or dtype != np_dtype_str
|
|
):
|
|
if dtype == "bfloat16":
|
|
source_array = np.frombuffer(source_array.tobytes(), "uint16")
|
|
source_array = np.ascontiguousarray(
|
|
source_array, dtype="uint16" if dtype == "bfloat16" else dtype
|
|
)
|
|
if self.dtype.startswith("float4_e2m1fn"):
|
|
# we need to pack the input data when converting to float4_e2m1fn type,
|
|
data_bits = source_array.view(dtype="uint8").flatten()
|
|
if data_bits.size % 2:
|
|
data_bits = np.pad(data_bits, (0, 1), mode="constant", constant_values=0)
|
|
data_bits = data_bits.reshape(-1, 2)
|
|
packed = ((data_bits[:, 0] & 0x0F) << 4) | (data_bits[:, 1] & 0x0F)
|
|
source_array = packed.astype(np.int8)
|
|
assert source_array.flags["C_CONTIGUOUS"]
|
|
data = source_array.ctypes.data_as(ctypes.c_void_p)
|
|
nbytes = source_array.size * source_array.dtype.itemsize
|
|
_ffi_api.TVMTensorCopyFromBytes(self, data, nbytes)
|
|
return self
|
|
|
|
def __repr__(self):
|
|
# exception safety handling for chandle=None
|
|
if self.__chandle__() == 0:
|
|
return type(self).__name__ + "(chandle=None)"
|
|
res = f"<tvm.runtime.Tensor shape={self.shape}, {self.device}>\n"
|
|
res += self.numpy().__repr__()
|
|
return res
|
|
|
|
def __str__(self):
|
|
return str(self.numpy())
|
|
|
|
def numpy(self):
|
|
"""Convert this array to numpy array
|
|
|
|
Returns
|
|
-------
|
|
np_arr : numpy.ndarray
|
|
The corresponding numpy array.
|
|
"""
|
|
t = tvm_ffi.dtype(self.dtype)
|
|
shape, dtype = self.shape, self.dtype
|
|
old_dtype = dtype
|
|
if t.lanes > 1:
|
|
shape = shape + (t.lanes,)
|
|
t = t.with_lanes(1)
|
|
dtype = str(t)
|
|
if dtype == "int4":
|
|
dtype = "int8"
|
|
if dtype in [
|
|
"bfloat16",
|
|
"float8_e3m4",
|
|
"float8_e4m3",
|
|
"float8_e4m3b11fnuz",
|
|
"float8_e4m3fn",
|
|
"float8_e4m3fnuz",
|
|
"float8_e5m2",
|
|
"float8_e5m2fnuz",
|
|
"float8_e8m0fnu",
|
|
"float6_e2m3fn",
|
|
"float6_e3m2fn",
|
|
"float4_e2m1fn",
|
|
]:
|
|
if ml_dtypes is None:
|
|
raise RuntimeError(
|
|
f"ml_dtypes is not installed, cannot convert {dtype} array to numpy."
|
|
)
|
|
try:
|
|
dtype = getattr(ml_dtypes, dtype)
|
|
except AttributeError:
|
|
raise RuntimeError(f"ml_dtypes has no attribute '{dtype}', cannot convert array.")
|
|
np_arr = np.empty(shape, dtype=dtype)
|
|
assert np_arr.flags["C_CONTIGUOUS"]
|
|
data = np_arr.ctypes.data_as(ctypes.c_void_p)
|
|
# TODO(kathy): revisit and get a mirrored function of ffi::GetDataSize
|
|
# in Python to replace line below
|
|
nbytes = np_arr.size if dtype == "bool" else (np_arr.size * old_dtype.bits + 7) // 8
|
|
_ffi_api.TVMTensorCopyToBytes(self, data, nbytes)
|
|
|
|
if old_dtype == "int4" or old_dtype.startswith("float4_e2m1fn"):
|
|
length = np_arr.size
|
|
np_arr = np_arr.view("int8")
|
|
np_arr_ret = np.empty((length,), dtype="int8")
|
|
np_arr = np_arr.reshape((length,))
|
|
odd_index = np.bitwise_and(np_arr, 0x0F)
|
|
even_index = np.bitwise_and(np_arr >> 4, 0x0F)
|
|
np_arr_ret[1::2] = odd_index[0 : length // 2]
|
|
np_arr_ret[0::2] = even_index[0 : (length + 1) // 2]
|
|
return np_arr_ret.reshape(shape).view(dtype)
|
|
|
|
return np_arr
|
|
|
|
def copyto(self, target, mem_scope=None):
|
|
"""Copy array to target
|
|
|
|
Parameters
|
|
----------
|
|
target : Tensor
|
|
The target array to be copied, must have same shape as this array.
|
|
|
|
mem_scope : Optional[str]
|
|
The memory scope of the array.
|
|
"""
|
|
if isinstance(target, Tensor):
|
|
return self._copyto(target)
|
|
if isinstance(target, tvm_ffi.core.Device):
|
|
res = empty(self.shape, self.dtype, target, mem_scope)
|
|
return self._copyto(res)
|
|
raise ValueError(f"Unsupported target type {type(target)}")
|
|
|
|
def _copyto(self, target_nd):
|
|
"""Internal function that implements copy to target ndarray."""
|
|
_ffi_api.TVMTensorCopyFromTo(self, target_nd)
|
|
return target_nd
|
|
|
|
def _create_view(self, shape, dtype: str | None = None, relative_byte_offset: int = 0):
|
|
"""Create a view into an existing array.
|
|
|
|
The view shares the same allocation and datatype as the
|
|
existing array, but can have a different array shape. This is
|
|
useful for runtimes that support non-flat memory, where both
|
|
the physical shape of an allocation and the logical shape of
|
|
the tensor it represents may need to be independently
|
|
specified.
|
|
|
|
Warning: This function should not be used outside of low-level
|
|
manipulations, as it breaks non-aliasing assumptions made by
|
|
TVM. This function may also be removed/replaced in the
|
|
future.
|
|
|
|
Parameters
|
|
----------
|
|
shape: Union[tvm_ffi.Shape, Sequence[typing.SupportsInt]]
|
|
|
|
The shape of the view.
|
|
|
|
dtype: Optional[str]
|
|
|
|
The datatype of the view. If None (default), the view
|
|
will be the same data type as the current array.
|
|
|
|
relative_byte_offset: int
|
|
|
|
The location of the view, relative to the location of the current
|
|
array.
|
|
|
|
Note: While the `DLTensor.byte_offset` field of the returned view
|
|
is usually the same as `relative_byte_offset`, this is not
|
|
guaranteed. The `DLTensor.byte_offset` field is relative to the
|
|
start of the backing allocation, while the `relative_byte_offset`
|
|
is relative to the start of `self`.
|
|
|
|
"""
|
|
|
|
if not isinstance(shape, tvm_ffi.Shape):
|
|
shape = tvm_ffi.Shape([int(dim) for dim in shape])
|
|
|
|
if dtype is None:
|
|
dtype = self.dtype
|
|
|
|
return _ffi_api.TVMTensorCreateView(self, shape, dtype, relative_byte_offset)
|
|
|
|
|
|
def empty(shape, dtype="float32", device=None, mem_scope=None):
|
|
"""Create an empty array given shape and device
|
|
|
|
Parameters
|
|
----------
|
|
shape : Union[tvm_ffi.Shape, Sequence[typing.SupportsInt]]
|
|
The shape of the array.
|
|
|
|
dtype : type or str
|
|
The data type of the array.
|
|
|
|
device : Device
|
|
The device of the array.
|
|
|
|
mem_scope : Optional[str]
|
|
The memory scope of the array.
|
|
|
|
Returns
|
|
-------
|
|
arr : tvm.runtime.Tensor
|
|
The array tvm supported.
|
|
"""
|
|
device = device or cpu()
|
|
if not isinstance(shape, tvm_ffi.Shape):
|
|
shape = tvm_ffi.Shape([int(dim) for dim in shape])
|
|
dtype = tvm_ffi.dtype(dtype)
|
|
arr = _ffi_api.TVMTensorAllocWithScope(shape, dtype, device, mem_scope)
|
|
return arr
|
|
|
|
|
|
def tensor(arr, device=None, mem_scope=None):
|
|
"""Create an tensor from source arr.
|
|
|
|
Parameters
|
|
----------
|
|
arr : numpy.ndarray
|
|
The array to be copied from
|
|
|
|
device : Device, optional
|
|
The device to create the array
|
|
|
|
mem_scope : Optional[str]
|
|
The memory scope of the array
|
|
|
|
Returns
|
|
-------
|
|
ret : Tensor
|
|
The created array
|
|
"""
|
|
device = device or cpu()
|
|
|
|
if not isinstance(arr, np.ndarray | Tensor):
|
|
arr = np.asarray(arr)
|
|
return empty(arr.shape, arr.dtype, device, mem_scope).copyfrom(arr)
|
|
|
|
|
|
def cpu(dev_id=0):
|
|
"""Construct a CPU device
|
|
|
|
Parameters
|
|
----------
|
|
dev_id : int, optional
|
|
The integer device id
|
|
|
|
Returns
|
|
-------
|
|
dev : Device
|
|
The created device
|
|
"""
|
|
return device(DLDeviceType.kDLCPU, dev_id)
|
|
|
|
|
|
def cuda(dev_id=0):
|
|
"""Construct a CUDA GPU device
|
|
|
|
Parameters
|
|
----------
|
|
dev_id : int, optional
|
|
The integer device id
|
|
|
|
Returns
|
|
-------
|
|
dev : Device
|
|
The created device
|
|
"""
|
|
return device(DLDeviceType.kDLCUDA, dev_id)
|
|
|
|
|
|
def rocm(dev_id=0):
|
|
"""Construct a ROCM device
|
|
|
|
Parameters
|
|
----------
|
|
dev_id : int, optional
|
|
The integer device id
|
|
|
|
Returns
|
|
-------
|
|
dev : Device
|
|
The created device
|
|
"""
|
|
return device(DLDeviceType.kDLROCM, dev_id)
|
|
|
|
|
|
def opencl(dev_id=0):
|
|
"""Construct a OpenCL device
|
|
|
|
Parameters
|
|
----------
|
|
dev_id : int, optional
|
|
The integer device id
|
|
|
|
Returns
|
|
-------
|
|
dev : Device
|
|
The created device
|
|
"""
|
|
return device(DLDeviceType.kDLOpenCL, dev_id)
|
|
|
|
|
|
def metal(dev_id=0):
|
|
"""Construct a metal device
|
|
|
|
Parameters
|
|
----------
|
|
dev_id : int, optional
|
|
The integer device id
|
|
|
|
Returns
|
|
-------
|
|
dev : Device
|
|
The created device
|
|
"""
|
|
return device(DLDeviceType.kDLMetal, dev_id)
|
|
|
|
|
|
def vpi(dev_id=0):
|
|
"""Construct a VPI simulated device
|
|
|
|
Parameters
|
|
----------
|
|
dev_id : int, optional
|
|
The integer device id
|
|
|
|
Returns
|
|
-------
|
|
dev : Device
|
|
The created device
|
|
"""
|
|
return device(DLDeviceType.kDLVPI, dev_id)
|
|
|
|
|
|
def vulkan(dev_id=0):
|
|
"""Construct a Vulkan device
|
|
|
|
Parameters
|
|
----------
|
|
dev_id : int, optional
|
|
The integer device id
|
|
|
|
Returns
|
|
-------
|
|
dev : Device
|
|
The created device
|
|
"""
|
|
return device(DLDeviceType.kDLVulkan, dev_id)
|
|
|
|
|
|
def ext_dev(dev_id=0):
|
|
"""Construct a extension device
|
|
|
|
Parameters
|
|
----------
|
|
dev_id : int, optional
|
|
The integer device id
|
|
|
|
Returns
|
|
-------
|
|
dev : Device
|
|
The created device
|
|
|
|
Note
|
|
----
|
|
This API is reserved for quick testing of new
|
|
device by plugin device API as ext_dev.
|
|
"""
|
|
return device(DLDeviceType.kDLExtDev, dev_id)
|
|
|
|
|
|
def hexagon(dev_id=0):
|
|
"""Construct a Hexagon device
|
|
|
|
Parameters
|
|
----------
|
|
dev_id : int, optional
|
|
The integer device id
|
|
|
|
Returns
|
|
-------
|
|
dev : Device
|
|
The created device
|
|
"""
|
|
return device(DLDeviceType.kDLHexagon, dev_id)
|
|
|
|
|
|
def webgpu(dev_id=0):
|
|
"""Construct a webgpu device.
|
|
|
|
Parameters
|
|
----------
|
|
dev_id : int, optional
|
|
The integer device id
|
|
|
|
Returns
|
|
-------
|
|
dev : Device
|
|
The created device
|
|
"""
|
|
return device(DLDeviceType.kDLWebGPU, dev_id)
|
|
|
|
|
|
# Register back to FFI
|
|
tvm_ffi.core._set_class_tensor(Tensor)
|