Arrays of user types #7
Conversation
inducer
left a comment
inducer
left a comment
There was a problem hiding this comment.
Thanks! I think this one will require some thought...
dagrt/builtins_python.py
Outdated
| return np.vdot(a, b) | ||
|
|
||
|
|
||
| def builtin_cumulative_product(a, axis): |
There was a problem hiding this comment.
This was only being used in one place in BDF, and I managed to remove it (562b2b5)
dagrt/builtins_python.py
Outdated
| def builtin_array_utype(n, x): | ||
| import numpy as np | ||
| if n != np.floor(n): | ||
| raise ValueError("array() argument n is not an integer") | ||
| n = int(n) | ||
|
|
||
| return np.empty((n, x.size), dtype=x.dtype) | ||
| #return np.zeros((n, x.size), dtype=x.dtype) |
There was a problem hiding this comment.
This builtin is needed for when we need to index into a set of user types at runtime (i.e. we cannot create a set of distinct usertype variables at generation time). In the context of BDF/this PR, the use case is Nordsieck history arrays (see this paper, eq. 2.6), which in the context of variable order must (at least I think) be a single array of user types.
dagrt/function_registry.py
Outdated
| identifier = "<builtin>norm_2" | ||
|
|
||
|
|
||
| class NormWRMS(_NormBase): |
There was a problem hiding this comment.
Can you add these to the docs? (i.e. make it so that they show up in the documentation) There's a list in the module docstring.
dagrt/builtins_python.py
Outdated
| return np.vdot(a, b) | ||
|
|
||
|
|
||
| def builtin_cumulative_product(a, axis): |
There was a problem hiding this comment.
This was only being used in one place in BDF, and I managed to remove it (562b2b5)
dagrt/function_registry.py
Outdated
| if check and not isinstance(x_kind, (NoneType, Array, UserType)): | ||
| raise TypeError("argument 'x' of 'array_utype' is not a user type") | ||
|
|
||
| return (Array(is_real_valued=True),) |
There was a problem hiding this comment.
This isn't true, really. Particularly the is_real_valued part.
dagrt/function_registry.py
Outdated
|
|
||
| is_real_valued = a_kind.is_real_valued and b_kind.is_real_valued | ||
|
|
||
| return (Array(is_real_valued),) |
dagrt/builtins_python.py
Outdated
| def builtin_reshape(a, a_cols): | ||
| return a.reshape(-1, a_cols, order="F") |
There was a problem hiding this comment.
Since dagrt arrays don't have shapes other than "1D vector", "reshape" is an odd addition IMO.
There was a problem hiding this comment.
This was only being used in one place in BDF, and I managed to remove it (562b2b5)
product builtins
for multiple UserTypes
builtin (avoids multi-UserType issue)
function end for temps
reusing user_type_move
moves, to avoid cluttering emit_assign_expr_inner
|
Could you remind me of the status of this? Is it ready to review? (If not, please mark as draft.) |
|
It will be ready soon, but isn't quite ready (need to put in place the outside-the-loop alloc/dealloc we talked about yesterday) |
|
@inducer this is ready for review - I'm removing the draft status after fixing what I could of the memory movement within loops. |
|
@inducer I've just looked this over again and brought it up to speed with the main If a refresher on motivation is needed, I would direct the reader to this paper, and in particular the use of Nordsieck history arrays for BDF integration as per this PR. |
2 participants
This PR aims to add functionality for arrays of user types to
dagrt's builtins, for the purposes of generating code for the instructions provided by Leap in this PR.