def ext_arr_to_matrix(arr: ext_arr(), mat: template(), as_vector: template()):
for I in ti.grouped(mat):
for p in ti.static(range(mat.n)):
for q in ti.static(range(mat.m)):
if ti.static(as_vector):
mat[I][p] = arr[I, p]
else:
mat[I][p, q] = arr[I, p, q]
def extract_arguments(self):
sig = inspect.signature(self.func)
if sig.return_annotation not in (inspect._empty, None):
self.return_type = sig.return_annotation
params = sig.parameters
arg_names = params.keys()
for i, arg_name in enumerate(arg_names):
param = params[arg_name]
if param.kind == inspect.Parameter.VAR_KEYWORD:
raise KernelDefError(
'Taichi kernels do not support variable keyword parameters (i.e., **kwargs)'
)
if param.kind == inspect.Parameter.VAR_POSITIONAL:
raise KernelDefError(
'Taichi kernels do not support variable positional parameters (i.e., *args)'
)
if param.default is not inspect.Parameter.empty:
raise KernelDefError(
'Taichi kernels do not support default values for arguments'
)
if param.kind == inspect.Parameter.KEYWORD_ONLY:
raise KernelDefError(
'Taichi kernels do not support keyword parameters')
if param.kind != inspect.Parameter.POSITIONAL_OR_KEYWORD:
raise KernelDefError(
'Taichi kernels only support "positional or keyword" parameters'
)
annotation = param.annotation
if param.annotation is inspect.Parameter.empty:
if i == 0 and self.classkernel: # The |self| parameter
annotation = template()
else:
_taichi_skip_traceback = 1
raise KernelDefError(
'Taichi kernels parameters must be type annotated')
else:
if isinstance(annotation, (template, ext_arr, any_arr)):
pass
elif id(annotation) in primitive_types.type_ids:
pass
elif isinstance(annotation, sparse_matrix_builder):
pass
else:
_taichi_skip_traceback = 1
raise KernelDefError(
f'Invalid type annotation (argument {i}) of Taichi kernel: {annotation}'
)
self.argument_annotations.append(annotation)
self.argument_names.append(param.name)
def extract_arguments(self):
sig = inspect.signature(self.func)
if sig.return_annotation not in (inspect._empty, None):
self.return_type = sig.return_annotation
params = sig.parameters
arg_names = params.keys()
for i, arg_name in enumerate(arg_names):
param = params[arg_name]
if param.kind == inspect.Parameter.VAR_KEYWORD:
raise KernelDefError(
'Taichi functions do not support variable keyword parameters (i.e., **kwargs)'
)
if param.kind == inspect.Parameter.VAR_POSITIONAL:
raise KernelDefError(
'Taichi functions do not support variable positional parameters (i.e., *args)'
)
if param.kind == inspect.Parameter.KEYWORD_ONLY:
raise KernelDefError(
'Taichi functions do not support keyword parameters')
if param.kind != inspect.Parameter.POSITIONAL_OR_KEYWORD:
raise KernelDefError(
'Taichi functions only support "positional or keyword" parameters'
)
annotation = param.annotation
if annotation is inspect.Parameter.empty:
if i == 0 and self.classfunc:
annotation = template()
else:
if id(annotation) in primitive_types.type_ids:
ti.warning(
'Data type annotations are unnecessary for Taichi'
' functions, consider removing it',
stacklevel=4)
elif not isinstance(annotation, template):
raise KernelDefError(
f'Invalid type annotation (argument {i}) of Taichi function: {annotation}'
)
self.arguments.append(annotation)
self.argument_names.append(param.name)
def vector_to_fast_image(img: template(), out: ext_arr()):
# FIXME: Why is ``for i, j in img:`` slower than:
for i, j in ti.ndrange(*img.shape):
r, g, b = 0, 0, 0
color = img[i, img.shape[1] - 1 - j]
if ti.static(img.dtype in [ti.f32, ti.f64]):
r, g, b = min(255, max(0, int(color * 255)))
else:
impl.static_assert(img.dtype == ti.u8)
r, g, b = color
idx = j * img.shape[0] + i
# We use i32 for |out| since OpenGL and Metal doesn't support u8 types
if ti.static(settings.get_os_name() != 'osx'):
out[idx] = (r << 16) + (g << 8) + b
else:
# What's -16777216?
#
# On Mac, we need to set the alpha channel to 0xff. Since Mac's GUI
# is big-endian, the color is stored in ABGR order, and we need to
# add 0xff000000, which is -16777216 in I32's legit range. (Albeit
# the clarity, adding 0xff000000 doesn't work.)
alpha = -16777216
out[idx] = (b << 16) + (g << 8) + r + alpha
def ext_arr_to_tensor(arr: ext_arr(), tensor: template()):
for I in ti.grouped(tensor):
tensor[I] = arr[I]
def tensor_to_tensor(tensor: template(), other: template()):
for I in ti.grouped(tensor):
tensor[I] = other[I]
def vector_to_image(mat: template(), arr: ext_arr()):
for I in ti.grouped(mat):
for p in ti.static(range(mat.n)):
arr[I, p] = ti.cast(mat[I][p], ti.f32)
if ti.static(mat.n <= 2):
arr[I, 2] = 0
def tensor_to_image(tensor: template(), arr: ext_arr()):
for I in ti.grouped(tensor):
t = ti.cast(tensor[I], ti.f32)
arr[I, 0] = t
arr[I, 1] = t
arr[I, 2] = t
def snode_deactivate_dynamic(b: template()):
for I in ti.grouped(b.parent()):
ti.deactivate(b, I)
def fill_tensor(tensor: template(), val: template()):
for I in ti.grouped(tensor):
tensor[I] = val
def snode_deactivate(b: template()):
for I in ti.grouped(b):
ti.deactivate(b, I)
def fill_matrix(mat: template(), vals: template()):
for I in ti.grouped(mat):
for p in ti.static(range(mat.n)):
for q in ti.static(range(mat.m)):
mat[I][p, q] = vals[p][q]
def clear_loss(l: template()):
# Using SNode writers would result in a forced sync, therefore we wrap these
# writes into a kernel.
l[None] = 0
l.grad[None] = 1
def clear_gradients(vars: template()):
for I in ti.grouped(ScalarField(Expr(vars[0]))):
for s in ti.static(vars):
ScalarField(Expr(s))[I] = 0
def to_torch_template(expr: template(), torch_tensor: ext_arr()):
for i in expr:
torch_tensor[i] = expr[i]
def from_torch_template(expr: template(), torch_tensor: ext_arr()):
for i in expr:
expr[i] = torch_tensor[i]
def tensor_to_ext_arr(tensor: template(), arr: ext_arr()):
for I in ti.grouped(tensor):
arr[I] = tensor[I]
