Re: Extend vector ops to boolean operators?
On 03/06/2012 10:10 PM, Simen Kjærås wrote:
On Tue, 06 Mar 2012 21:35:11 +0100, Timon Gehr wrote:
On 03/06/2012 09:30 PM, H. S. Teoh wrote:
It'd be really cool if I could do this:
void func(int[] vector, int[] bounds) {
assert(vector[]>= 0&& vector[]< bounds[]);
...
}
Is there any reason why we shouldn't implement this?
T
Comparing arrays already does lexical-style comparison (which makes
sense).
Comparing two arrays makes sense, absolutely. Comparing one T[] and
one T currently does not. Also, foo[] already changes the behavior of
operators on foo,
No, it does not change their behavior, it is just required to allow
them. Array slices have value semantics, and foo[] is currently
shorthand for foo[0..$].
making it do a per-element compare would be in line
with this pattern.
Making array slicing change the behavior of some operator wouldn't have
any precedent.
This is also already in bugzilla:
http://d.puremagic.com/issues/show_bug.cgi?id=5636
Re: Extend vector ops to boolean operators?
On 03/06/2012 09:58 PM, H. S. Teoh wrote:
On Tue, Mar 06, 2012 at 09:35:11PM +0100, Timon Gehr wrote:
On 03/06/2012 09:30 PM, H. S. Teoh wrote:
It'd be really cool if I could do this:
void func(int[] vector, int[] bounds) {
assert(vector[]>= 0&& vector[]< bounds[]);
...
}
Is there any reason why we shouldn't implement this?
[...]
Comparing arrays already does lexical-style comparison (which makes sense).
What I wanted is not lexicographic comparison, but per-element
comparison:
v[]>=0 means v[0]>0&& v[1]>0&& v[2]>0&& ...
v[]
I know. You asked for reasons why it shouldn't be implemented. The main
reason is that it is already valid code with different semantics.
Re: Extend vector ops to boolean operators?
On 7 March 2012 17:03, bearophile wrote: > James Miller: > >> What? I'm assuming you mean that you expect an array of `bool`s? > > Right. Vector operations like a[] To see how this is useful you probably must think in terms of vector-style > programming. In NumPy the use of arrays of booleans is common: > from numpy import * a = array([3,6,8,9]) a == 6 > array([False, True, False, False], dtype=bool) a >= 7 > array([False, False, True, True], dtype=bool) a < 5 > array([ True, False, False, False], dtype=bool) # count all the even numbers sum( (a%2) == 0 ) > 2 b = array([2,6,7,10]) a == b > array([False, True, False, False], dtype=bool) a < b > array([False, False, False, True], dtype=bool) > > > They are sometimes used as masks, it's useful if you have a Vector type that > supports multi-index syntax: > > i = scipy.array([0,1,2,1]) # array of indices for the first axis > j = scipy.array([1,2,3,4]) # array of indices for the second axis > a[i,j] # return array([a[0,1], a[1,2], a[2,3], a[1,4]]) > b = scipy.array([True, False, True, False]) > a[b] # return array([a[0], a[2]]) since only b[0] and b[2] are True > > > Using the new CPU AVX registers you are able to perform a loop and work on > the items of an array in parallel until all the booleans of an array are > false. See this, expecially Listing 5: > > http://software.intel.com/en-us/articles/introduction-to-intel-advanced-vector-extensions/ > > http://www.cs.uaf.edu/2011/spring/cs641/lecture/04_12_AVX.html > > Vector comparisons have a natural hardware implementataion with AVX/AVX2 > instructions like _mm256_cmp_ps. > > Bye, > bearophile Hmm, I see your point, I think that with D's current operator overloading you could implement most of that. Other than the comparison syntax. If vector comparison gets added, then there should be some very clear documentation that it returns a vector, because I can imagine using it in an if statement and then wondering why it always went through...
Re: Extend vector ops to boolean operators?
James Miller: > What? I'm assuming you mean that you expect an array of `bool`s? Right. Vector operations like a[]>> from numpy import * >>> a = array([3,6,8,9]) >>> a == 6 array([False, True, False, False], dtype=bool) >>> a >= 7 array([False, False, True, True], dtype=bool) >>> a < 5 array([ True, False, False, False], dtype=bool) >>> # count all the even numbers >>> sum( (a%2) == 0 ) 2 >>> b = array([2,6,7,10]) >>> a == b array([False, True, False, False], dtype=bool) >>> a < b array([False, False, False, True], dtype=bool) They are sometimes used as masks, it's useful if you have a Vector type that supports multi-index syntax: i = scipy.array([0,1,2,1]) # array of indices for the first axis j = scipy.array([1,2,3,4]) # array of indices for the second axis a[i,j] # return array([a[0,1], a[1,2], a[2,3], a[1,4]]) b = scipy.array([True, False, True, False]) a[b] # return array([a[0], a[2]]) since only b[0] and b[2] are True Using the new CPU AVX registers you are able to perform a loop and work on the items of an array in parallel until all the booleans of an array are false. See this, expecially Listing 5: http://software.intel.com/en-us/articles/introduction-to-intel-advanced-vector-extensions/ http://www.cs.uaf.edu/2011/spring/cs641/lecture/04_12_AVX.html Vector comparisons have a natural hardware implementataion with AVX/AVX2 instructions like _mm256_cmp_ps. Bye, bearophile
Re: Extend vector ops to boolean operators?
On 7 March 2012 15:35, Peter Alexander wrote:
> On Tuesday, 6 March 2012 at 23:57:07 UTC, James Miller wrote:
>>
>> On 7 March 2012 10:58, Kapps wrote:
>>>
>>> On Tuesday, 6 March 2012 at 20:28:40 UTC, H. S. Teoh wrote:
It'd be really cool if I could do this:
void func(int[] vector, int[] bounds) {
assert(vector[] >= 0 && vector[] < bounds[]);
...
}
Is there any reason why we shouldn't implement this?
T
>>>
>>>
>>>
>>> Would this be possible with UFCS?
>>>
>>> int opCmp(T)T([] array, T element) { ... }
>>> int opCmp(T)(T[] array1, T[] array2) { ... }
>>
>>
>> I like this idea, at least adding an opSliceCmp operator-overload
>> would do as a start, I think thats the correct name for it. I can't be
>> bothered to check.
>>
>> --
>> James Miller
>
>
> It has to be done as vector operations.
>
> a[] < b[] should equal [a[0] < b[0], a[1] < b[1], ... ]
>
> What the OP has asked for is not a vector operation, so it shouldn't use the
> vector op syntax.
What? I'm assuming you mean that you expect an array of `bool`s? While
I agree that most vector operations should return a vector, comparison
makes more sense as returning a straight bool, most of the time you
aren't going to want to just have an array of bools. The only use case
i can think of is this:
auto c = a[] < b[]
foreach (i : c) {
if (i) doSomething();
else doSomethingElse();
}
Which can be easily re-written as
for (int i; i < a.length; i++) {
if (a[i] < b[i]) doSomething();
else doSomethingElse();
}
(ignoring things like proper checking etc.)
Of course most vector ops should return vectors, but most other ops
return proper types too, comparison ops tend to be the exception to
the rule.
--
James Miller
Re: Extend vector ops to boolean operators?
On Tuesday, 6 March 2012 at 23:57:07 UTC, James Miller wrote:
On 7 March 2012 10:58, Kapps wrote:
On Tuesday, 6 March 2012 at 20:28:40 UTC, H. S. Teoh wrote:
It'd be really cool if I could do this:
void func(int[] vector, int[] bounds) {
assert(vector[] >= 0 && vector[] <
bounds[]);
...
}
Is there any reason why we shouldn't implement this?
T
Would this be possible with UFCS?
int opCmp(T)T([] array, T element) { ... }
int opCmp(T)(T[] array1, T[] array2) { ... }
I like this idea, at least adding an opSliceCmp
operator-overload
would do as a start, I think thats the correct name for it. I
can't be
bothered to check.
--
James Miller
It has to be done as vector operations.
a[] < b[] should equal [a[0] < b[0], a[1] < b[1], ... ]
What the OP has asked for is not a vector operation, so it
shouldn't use the vector op syntax.
Re: Extend vector ops to boolean operators?
On 7 March 2012 10:58, Kapps wrote:
> On Tuesday, 6 March 2012 at 20:28:40 UTC, H. S. Teoh wrote:
>>
>> It'd be really cool if I could do this:
>>
>> void func(int[] vector, int[] bounds) {
>> assert(vector[] >= 0 && vector[] < bounds[]);
>>
>> ...
>> }
>>
>> Is there any reason why we shouldn't implement this?
>>
>>
>> T
>
>
> Would this be possible with UFCS?
>
> int opCmp(T)T([] array, T element) { ... }
> int opCmp(T)(T[] array1, T[] array2) { ... }
I like this idea, at least adding an opSliceCmp operator-overload
would do as a start, I think thats the correct name for it. I can't be
bothered to check.
--
James Miller
Re: Extend vector ops to boolean operators?
On Tuesday, 6 March 2012 at 20:28:40 UTC, H. S. Teoh wrote:
It'd be really cool if I could do this:
void func(int[] vector, int[] bounds) {
assert(vector[] >= 0 && vector[] < bounds[]);
...
}
Is there any reason why we shouldn't implement this?
T
Would this be possible with UFCS?
int opCmp(T)T([] array, T element) { ... }
int opCmp(T)(T[] array1, T[] array2) { ... }
Re: Extend vector ops to boolean operators?
On Tue, 06 Mar 2012 21:35:11 +0100, Timon Gehr wrote:
On 03/06/2012 09:30 PM, H. S. Teoh wrote:
It'd be really cool if I could do this:
void func(int[] vector, int[] bounds) {
assert(vector[]>= 0&& vector[]< bounds[]);
...
}
Is there any reason why we shouldn't implement this?
T
Comparing arrays already does lexical-style comparison (which makes
sense).
Comparing two arrays makes sense, absolutely. Comparing one T[] and
one T currently does not. Also, foo[] already changes the behavior of
operators on foo, making it do a per-element compare would be in line
with this pattern.
This is also already in bugzilla:
http://d.puremagic.com/issues/show_bug.cgi?id=5636
Re: Extend vector ops to boolean operators?
On 3/6/2012 2:30 PM, H. S. Teoh wrote:
It'd be really cool if I could do this:
void func(int[] vector, int[] bounds) {
assert(vector[]>= 0&& vector[]< bounds[]);
...
}
Is there any reason why we shouldn't implement this?
T
This same problem exists for making proper syntactical sugar for simd
comparison functions.
!= == (opEquals is required to return a bool)
and
<= >= < > (opCmp is required to return an int)
Granted its possible to live without the sugar but the code looks more
like asm, and reading the code takes longer without the operators in it.
Re: Extend vector ops to boolean operators?
On Tue, Mar 06, 2012 at 09:35:11PM +0100, Timon Gehr wrote:
> On 03/06/2012 09:30 PM, H. S. Teoh wrote:
> >It'd be really cool if I could do this:
> >
> > void func(int[] vector, int[] bounds) {
> > assert(vector[]>= 0&& vector[]< bounds[]);
> > ...
> > }
> >
> >Is there any reason why we shouldn't implement this?
[...]
>
> Comparing arrays already does lexical-style comparison (which makes sense).
What I wanted is not lexicographic comparison, but per-element
comparison:
v[]>=0 means v[0]>0 && v[1]>0 && v[2]>0 && ...
v[]
Re: Extend vector ops to boolean operators?
On 03/06/2012 09:30 PM, H. S. Teoh wrote:
It'd be really cool if I could do this:
void func(int[] vector, int[] bounds) {
assert(vector[]>= 0&& vector[]< bounds[]);
...
}
Is there any reason why we shouldn't implement this?
T
Comparing arrays already does lexical-style comparison (which makes sense).
