Abstracting over things that can be unpacked
Hi all, These ideas are still in very early stages. I present them here in hope of starting a discussion. (We discussed this quite a bit at last year's ICFP, I hope this slightly different take on the problem might lead to new ideas.) I think the next big step in Haskell performance is going to come from using better data representation in common types such as list, sets, and maps. Today these polymorphic data structures use both more memory and have more indirections than necessary, due to boxing of values. This boxing is due to the values being stored in fields of polymorphic type. First idea: instead of rejecting unpack pragmas on polymorphic fields, have them require a class constraint on the field types. Example: data UnboxPair a b = (Unbox a, Unbox b) => UP {-# UNPACK #-} !a {-# UNPACK #-} !b The Unbox type class would be similar in spirit to the class with the same name in the vector package, but be implemented internally by GHC. To a first approximation instances would only exist for fields that unpack to non-pointer types (e.g. Int.) Second idea: Introduce a new pragma, that has similar effect on representations as DPH's [::] vector type. This new pragma does deep unpacking, allowing for more types to be instances of the Unbox type e.g. pairs. Example: data T = C {-# UNWRAP #-} (a, b) If you squint a bit this pragma does the same as [: (a, b) :], except no vectors are involved. The final representation would be the unpacked representation of a and b, concatenated together (e.g. (Int, Int) would result in the field above being 128-bit wide on a 64-bit machine. The meta-idea tying these two ideas together is to allow for some abstraction over representation transforming pragmas, such as UNPACK. P.S. Before someone suggest using type families. Please read my email titled "Avoiding O(n^2) instances when using associated data types to unpack values into constructors." Cheers, Johan
+1 !
On Fri, Mar 2, 2012 at 7:40 PM, Johan Tibell
Hi all,
These ideas are still in very early stages. I present them here in hope of starting a discussion. (We discussed this quite a bit at last year's ICFP, I hope this slightly different take on the problem might lead to new ideas.)
I think the next big step in Haskell performance is going to come from using better data representation in common types such as list, sets, and maps. Today these polymorphic data structures use both more memory and have more indirections than necessary, due to boxing of values. This boxing is due to the values being stored in fields of polymorphic type.
First idea: instead of rejecting unpack pragmas on polymorphic fields, have them require a class constraint on the field types. Example:
data UnboxPair a b = (Unbox a, Unbox b) => UP {-# UNPACK #-} !a {-# UNPACK #-} !b
The Unbox type class would be similar in spirit to the class with the same name in the vector package, but be implemented internally by GHC. To a first approximation instances would only exist for fields that unpack to non-pointer types (e.g. Int.)
Second idea: Introduce a new pragma, that has similar effect on representations as DPH's [::] vector type. This new pragma does deep unpacking, allowing for more types to be instances of the Unbox type e.g. pairs. Example:
data T = C {-# UNWRAP #-} (a, b)
If you squint a bit this pragma does the same as [: (a, b) :], except no vectors are involved. The final representation would be the unpacked representation of a and b, concatenated together (e.g. (Int, Int) would result in the field above being 128-bit wide on a 64-bit machine.
The meta-idea tying these two ideas together is to allow for some abstraction over representation transforming pragmas, such as UNPACK.
P.S. Before someone suggest using type families. Please read my email titled "Avoiding O(n^2) instances when using associated data types to unpack values into constructors."
Cheers, Johan
_______________________________________________ Glasgow-haskell-users mailing list Glasgow-haskell-users@haskell.org http://www.haskell.org/mailman/listinfo/glasgow-haskell-users
On 03/03/12 01:40, Johan Tibell wrote:
Hi all,
These ideas are still in very early stages. I present them here in hope of starting a discussion. (We discussed this quite a bit at last year's ICFP, I hope this slightly different take on the problem might lead to new ideas.)
I think the next big step in Haskell performance is going to come from using better data representation in common types such as list, sets, and maps. Today these polymorphic data structures use both more memory and have more indirections than necessary, due to boxing of values. This boxing is due to the values being stored in fields of polymorphic type.
First idea: instead of rejecting unpack pragmas on polymorphic fields, have them require a class constraint on the field types. Example:
data UnboxPair a b = (Unbox a, Unbox b) => UP {-# UNPACK #-} !a {-# UNPACK #-} !b
I expect that this will not be easy to implement, because it requires interaction with things like the garbage collector. For example, UnboxPair will need a different info table for different a and b. It might be possible to essentially specialize UnboxPair for each different a and b for which it is used, but that gets tricky with generic code.
The Unbox type class would be similar in spirit to the class with the same name in the vector package, but be implemented internally by GHC. To a first approximation instances would only exist for fields that unpack to non-pointer types (e.g. Int.)
Second idea: Introduce a new pragma, that has similar effect on representations as DPH's [::] vector type. This new pragma does deep unpacking, allowing for more types to be instances of the Unbox type.
Could this be handled by just having/deriving an Unbox instance for (a,b)? I imagine the Unbox type class would have to contain essentially the same things as Storable, maybe something like type UnboxedRepr :: Int -> # class Unbox a where type Repr a :: # -- gives size and alignment unbox :: a -> Repr a box :: Repr a -> a A problem with an instance (Unboxed a, Unboxed b) => Unboxed (a,b) is that it allows arbitrarily large unboxed values to be created at runtime. That doesn't work when you use specialization to create the needed info tables. Twan
On Sat, Mar 3, 2012 at 8:06 AM, Twan van Laarhoven
I expect that this will not be easy to implement, because it requires interaction with things like the garbage collector. For example, UnboxPair will need a different info table for different a and b.
It might be possible to essentially specialize UnboxPair for each different a and b for which it is used, but that gets tricky with generic code.
I believe use-site code generation is the way to go. This is what C++ does (at compile time) and C# (at runtime, using the JIT) does.
The Unbox type class would be similar in spirit to the class with the
same name in the vector package, but be implemented internally by GHC. To a first approximation instances would only exist for fields that unpack to non-pointer types (e.g. Int.)
Second idea: Introduce a new pragma, that has similar effect on representations as DPH's [::] vector type. This new pragma does deep unpacking, allowing for more types to be instances of the Unbox type.
Could this be handled by just having/deriving an Unbox instance for (a,b)? I imagine the Unbox type class would have to contain essentially the same things as Storable, maybe something like
type UnboxedRepr :: Int -> # class Unbox a where type Repr a :: # -- gives size and alignment unbox :: a -> Repr a box :: Repr a -> a
A problem with an instance (Unboxed a, Unboxed b) => Unboxed (a,b) is that it allows arbitrarily large unboxed values to be created at runtime. That doesn't work when you use specialization to create the needed info tables.
As I mentioned further up in the email, I think this needs to be done at compile time. However, I'm not sure type classes are the right mechanism, as they don't guarantee that the polymorphism is resolved at compile time. Perhaps type families, in some form, is the right solution. -- Johan
As I mentioned further up in the email, I think this needs to be done at compile time. However, I'm not sure type classes are the right mechanism, as they don't guarantee that the polymorphism is resolved at compile time. Perhaps type families, in some form, is the right solution.
There is problem with type families. Currently GHC is unable to unpack them[1]. [1] http://hackage.haskell.org/trac/ghc/ticket/3990
First idea: instead of rejecting unpack pragmas on polymorphic fields, have them require a class constraint on the field types. Example: data UnboxPair a b = (Unbox a, Unbox b) => UP {-# UNPACK #-} !a {-# UNPACK #-} !b The Unbox type class would be similar in spirit to the class with the same name in the vector package, but be implemented internally by GHC. To a first approximation instances would only exist for fields that unpack to non-pointer types (e.g. Int.) Johan I don't understand how this would work. What methods does the class Unbox have? What is the final representation of a UP constructor? How many words? Pointers or non-pointers? At the moment each constructor has a fixed layout. I'm really not sure what you have in mind here. More detail needed! Simon
participants (5)
-
Aleksey Khudyakov -
Johan Tibell -
Ryan Newton -
Simon Peyton-Jones -
Twan van Laarhoven