6.2 KiB
Memory aliasing
Memory aliasing is a way to tell Burst how your code uses data. This can improve and optimize the performance of your application.
Memory aliasing happens when locations in the memory overlap each other. The following documentation outlines the difference between memory aliasing, and no memory aliasing.
The following example shows a job that copies data from an input array to an output array:
[BurstCompile]
private struct CopyJob : IJob
{
[ReadOnly]
public NativeArray<float> Input;
[WriteOnly]
public NativeArray<float> Output;
public void Execute()
{
for (int i = 0; i < Input.Length; i++)
{
Output[i] = Input[i];
}
}
}
No memory aliasing
If the arrays Input
and Output
don't overlap, which means that their respective memory location doesn't overlap, the code returns the following result after running this job on a sample input/output:
If Burst is noalias aware, it can work at the scalar level to optimize the previous scalar loop. It does this through a process called vectorizing, where it rewrites the loop to process elements in a small batch. For example, Burst could work at vector level in 4 by 4 elements:
Memory with no aliasing vectorized
Memory aliasing
If the Output
array overlaps the Input
array by one element (for example Output[0]
points to Input[1]
), then this means that the memory is aliasing. This gives the following result when you run CopyJob
without the auto vectorizer:
If Burst isn't aware of the memory aliasing, it tries to auto vectorize the loop, which results in the following:
Memory with aliasing and invalid vectorized code
The result of this code is invalid and might lead to bugs if Burst can't identify them.
Generated code
In the CopyJob
example, there is an x64
assembly targeted at AVX2
in its loop. The instruction vmovups
moves 8 floats, so a single auto vectorized loop moves 4 × 8 floats, which equals 32 floats copied per loop iteration, instead of just one:
.LBB0_4:
vmovups ymm0, ymmword ptr [rcx - 96]
vmovups ymm1, ymmword ptr [rcx - 64]
vmovups ymm2, ymmword ptr [rcx - 32]
vmovups ymm3, ymmword ptr [rcx]
vmovups ymmword ptr [rdx - 96], ymm0
vmovups ymmword ptr [rdx - 64], ymm1
vmovups ymmword ptr [rdx - 32], ymm2
vmovups ymmword ptr [rdx], ymm3
sub rdx, -128
sub rcx, -128
add rsi, -32
jne .LBB0_4
test r10d, r10d
je .LBB0_8
The following example shows the same Burst compiled loop, but Burst's aliasing is artificially disabled:
.LBB0_2:
mov r8, qword ptr [rcx]
mov rdx, qword ptr [rcx + 16]
cdqe
mov edx, dword ptr [rdx + 4*rax]
mov dword ptr [r8 + 4*rax], edx
inc eax
cmp eax, dword ptr [rcx + 8]
jl .LBB0_2
The result is entirely scalar and runs approximately 32 times slower than the highly optimized, vectorized variant that the original alias analysis produces.
Function cloning
For function calls where Burst knows about the aliasing between parameters to the function, Burst can infer the aliasing. It can then propagate this onto the called function to improve optimization:
[MethodImpl(MethodImplOptions.NoInlining)]
int Bar(ref int a, ref int b)
{
a = 42;
b = 13;
return a;
}
int Foo()
{
var a = 53;
var b = -2;
return Bar(ref a, ref b);
}
The assembly for Bar
would be:
mov dword ptr [rcx], 42
mov dword ptr [rdx], 13
mov eax, dword ptr [rcx]
ret
This is because Burst doesn't know the aliasing of a
and b
within the Bar
function. This is in line with what other compiler technologies do with this code snippet.
Burst is smarter than this though. Through a process of function cloning, Burst creates a copy of Bar
where it knows that the aliasing properties of a
and b
don't alias. It then replaces the original call to Bar
with a call to the copy. This results in the following assembly:
mov dword ptr [rcx], 42
mov dword ptr [rdx], 13
mov eax, 42
ret
In this scenario, Burst doesn't perform the second load from a
.
Aliasing checks
Because aliasing is key to Burst's ability to optimize for performance, there are some aliasing intrinsics:
Unity.Burst.CompilerServices.Aliasing.ExpectAliased
expects that the two pointers do alias, and generates a compiler error if not.Unity.Burst.CompilerServices.Aliasing.ExpectNotAliased
expects that the two pointers don't alias, and generates a compiler error if not.
An example:
using static Unity.Burst.CompilerServices.Aliasing;
[BurstCompile]
private struct CopyJob : IJob
{
[ReadOnly]
public NativeArray<float> Input;
[WriteOnly]
public NativeArray<float> Output;
public unsafe void Execute()
{
// NativeContainer attributed structs (like NativeArray) cannot alias with each other in a job struct!
ExpectNotAliased(Input.getUnsafePtr(), Output.getUnsafePtr());
// NativeContainer structs cannot appear in other NativeContainer structs.
ExpectNotAliased(in Input, in Output);
ExpectNotAliased(in Input, Input.getUnsafePtr());
ExpectNotAliased(in Input, Output.getUnsafePtr());
ExpectNotAliased(in Output, Input.getUnsafePtr());
ExpectNotAliased(in Output, Output.getUnsafePtr());
// But things definitely alias with themselves!
ExpectAliased(in Input, in Input);
ExpectAliased(Input.getUnsafePtr(), Input.getUnsafePtr());
ExpectAliased(in Output, in Output);
ExpectAliased(Output.getUnsafePtr(), Output.getUnsafePtr());
}
}
These checks only run when optimizations are enabled, because proper aliasing deduction is intrinsically linked to the optimizer's ability to see through functions via inlining.