Shaders frequently need fine-grained control over the way the compiler optimizes
(or doesn’t) math optimizations. To enhance developer control in a more robust
way than the existing precise qualifier, HLSL should gain a higher level
concept of math modes.
The precise qualifier is particularly gnarly to implement and is wrought with
bugs that are inscrutable to developers. The source of the implementation
challenges come from the fact that the precise attribute is applied to a
declaration, but does not impact the uses of the declaration. Instead it impacts
reverse-propagates and applies to all of the math expressions that are used to
compute the declared variable.
The goal of this proposal is to create a new language feature which allows shader authors the control they need to replace precise, without the tooling complexity.
First and foremost, since the goal is to replace the precise qualifier, this
proposal should deprecate precise in the language. Allowing the new feature
and precise to exist together for a single language version will ease
adoption, however precise will be removed in the future.
The core of this proposal is to introduce new math modes strict and fast.
The strict math mode prohibits optimizations which may impact the precision of
results. This prevents optimizations like fusing multiply and add instructions.
The fast math mode prioritizes speed over precision guarantees. This does not
mean that fast math is less precise. In some cases optimal fused operations
are faster, but the results are less consistent across hardware architectures.
As a simplified view, strict can be viewed as “execute the math as written in
source”, while fast is “let the optimizer do its thing”.
The new math modes are exposed through a variety of mechanisms described below.
The new hlsl::math_mode() attribute can be set to either strict, fast, or
default. The default math mode is fast, however this can be changed with the
compiler flag:math-mode=precise|fast.
The math attribute can be applied to functions or expressions, but it does not
propagate in the way that the precise keyword did.
For example:
[hlsl::math_mode(strict)]
float fma(float x, float y, float z) {
return x * y + z;
}
float fma2(float x, float y, float z) {
return [hlsl::math_mode(strict)] x * y + z;
}
In both of the above functions optimization to an FMA instruction is prevented.
New math namespaces are also added which provide simplified access to math operations with the expected performance characteristics.
For example:
float f = hlsl::strict::fma(1.0, 2.0, 3.0); // strict mode math
float f = hlsl::fast::fma(1.0, 2.0, 3.0); // fast mode math
float f = hlsl::fma(1.0, 2.0, 3.0); // default mode math