On 08/23/2013 02:02 PM, Paul Berry wrote:
On 23 August 2013 13:19, Matt Turner <matts...@gmail.com
<mailto:matts...@gmail.com>> wrote:
On Fri, Aug 23, 2013 at 8:55 AM, Paul Berry <stereotype...@gmail.com
<mailto:stereotype...@gmail.com>> wrote:
> On 22 August 2013 16:08, Matt Turner <matts...@gmail.com
<mailto:matts...@gmail.com>> wrote:
>>
>> ---
>> src/glsl/builtins/ir/frexp.ir <http://frexp.ir>
| 25
>> +++++++++++++++++++++++++
>> src/glsl/builtins/ir/ldexp.ir <http://ldexp.ir>
| 25
>> +++++++++++++++++++++++++
>> src/glsl/builtins/profiles/ARB_gpu_shader5.glsl | 10 ++++++++++
>> 3 files changed, 60 insertions(+)
>> create mode 100644 src/glsl/builtins/ir/frexp.ir <http://frexp.ir>
>> create mode 100644 src/glsl/builtins/ir/ldexp.ir <http://ldexp.ir>
>>
>> diff --git a/src/glsl/builtins/ir/frexp.ir <http://frexp.ir>
b/src/glsl/builtins/ir/frexp.ir <http://frexp.ir>
>> new file mode 100644
>> index 0000000..a514994
>> --- /dev/null
>> +++ b/src/glsl/builtins/ir/frexp.ir <http://frexp.ir>
>> @@ -0,0 +1,25 @@
>> +((function frexp
>> + (signature float
>> + (parameters
>> + (declare (in) float x)
>> + (declare (out) int exp))
>> + ((return (expression float frexp (var_ref x) (var_ref exp)))))
>
>
> Having an ir_expression that writes to one of its parameters is
going to
> break assumptions in a lot of our optimization passes.
I'm concerned that that may be a problem we have to solve anyway.
While our hardware doesn't support an frexp instruction (like e.g.,
AMD does) and we could probably do what you suggest, we do have
instructions that correspond directly to the uaddCarry() and
usubBorrow() built-ins in this same extension. They return a value and
also have an out parameter.
genUType uaddCarry(genUType x, genUType y, out genUType carry);
genUType usubBorrow(genUType x, genUType y, out genUType borrow);
We could probably avoid the problem you describe by lowering them, but
it's feeling increasingly distasteful.
Your code would make a good piglit test. I'll do some experiments.
Hmm, interesting.
The way LLVM solves this problem, as I understand it, is through
so-called "intrinsic functions"
(http://llvm.org/docs/LangRef.html#intrinsic-functions). I wonder if we
should start doing that in Mesa.
Briefly, here is what it would look like, using uaddCarry as an example:
1. First we do an inefficient implementation of uaddCarry in terms of
existing GLSL functions, much like you did for frexp in your
frexp_to_arith lowering pass, except that we do it in
src/glsl/builtins/glsl/uaddCarry.glsl, so it's a little easier to review
:). Optimization passes already deal with function "out" parameters
properly, and function inlining automatically splices in the proper code
during linking.
2. For back-ends that don't have an efficient native way to do
uaddCarry, we're done. The uaddCarry function works as is.
3. For back-ends that do have an efficient way to do uaddCarry, we add a
mechanism to allow the back-end to tell the linker: "don't inline the
definition of this built-in. Just leave it as an ir_call because I have
my own special implementation of it"*.
I had thought about solving this in a slightly different way, but there
are a couple potential tricky bits.
Provide an implementation of the built-in function in the GLSL library.
float frexp(float x, out int exponent)
{
return __intrinsic_frexp(x, exponent);
}
Provide a default implementation of the intrinsic elsewhere.
Allow drivers to supply an alternate library with custom versions of the
intrinsics.
Since the GLSL library's frexp is the same in either case, the problem
Paul identifies below should be avoided.
The tricky bit, and the problem we always come to when talking about
intrinsics is dealing with constant expressions. That doesn't
(shouldn't?) apply to this case because of the out parameter, but it may
apply to other cases.
Right now an application could do:
float foo[packUnorm2x16(vec2(1,0))];
If packUnorm2x16 becomes __intrinsic_packUnorm2x16, the constant
expression evaluator has to be able to handle whatever
__intrinsic_packUnorm2x16 becomes.
4. In the back-end visitor code, the ir_call visitor looks at the name
of the function being called. If it's "uaddCarry", then the back-end
visitor just emits the efficient back-end code. Any other ir_calls
should have been eliminated by the function inlining.
*We'll need to be careful to make sure that the right thing happens if
the user overrides uaddCarry with their own user-defined function, of
course :)
Now that I've actually thought through it, I'm really excited about this
idea. It seems way more straightforward than what we are currently
doing (e.g. in lower_packing_builtins.cpp), and it works nicely with the
other back-ends because if a back-end doesn't advertise an intrinsic
definition of a given function, it automtically gets the version
declared in src/glsl/builtins without having to do any extra work.
What do you think?
_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/mesa-dev
_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/mesa-dev
