On Friday, 8 July 2016 at 11:05:36 UTC, ZombineDev wrote:
On Friday, 8 July 2016 at 06:17:43 UTC, Nicholas Wilson wrote:
On Friday, 8 July 2016 at 05:53:21 UTC, Nicholas Wilson wrote:
So as part of my effort to get D running on GPUs I need to
make a "second class" pointer type that I can alter in the
backend of LDC to the correct address space. to that end I
have made the following
module dcompute.types.pointer;
enum Private = 0;
enum Global = 1;
enum Shared = 2;
enum Constant = 3;
enum Generic = 4;
pure @trusted nothrow @nogc
struct Pointer(uint p, T) if(p <= Generic)
{
T* ptr;
ref T opUnary(string op)() if(op=="*")
{
return *ptr;
}
ref T opIndex(size_t i)
{
return *(ptr+i);
}
auto opBinary(string op)(ptrdiff_t i) if(op=="+" || op ==
"-")
{
return Pointer!(p, T) (ptr + i);
}
}
...
That should be mixin("ptr"~op~"i")
I need compound add and compound subtract as well.
Why not make the address (memory) space a strongly-typed enum,
e.g.:
enum AddressSpace
{
threadLocal = 0,
groupLocal = 1,
global = 2,
constant = 3,
texture = 4
generic = 5
}
because it'll change the mangling and I just want to get it
working first, also there will be aliases for all the things
anyway. (also AFAIU there is no texture address space its just
fancy caching)
struct Pointer(T, AddressSpace type = AddressSpace.Generic)
// ...
?
is there anything else that I'm missing that you can do with
a raw pointer?
Slicing:
That's creating a new variable based on the pointer as {$-offset,
ptr + offset} not a primitive, but I take your point.
// Pseudo OpenCL-D
groupLocal T* ptr1;
groupLocal T[] slice1 = ptr1[2 .. 10];
global T* ptr2;
groupLocal T[] slice2 = ptr2[0 .. 2]; // Error
I REALLY want to avoid having to make them qualifiers and that
means I have to hack DMDFE as well as LDC.
