hlsl syntax 64 bit buffer and image atomics for spirv about slang

[Ipotrick]

[[vk::binding(0,0)]] RWTexture2D<uint64_t> tex[];

[shader("compute")]
[numthreads(1, 1, 1)]
void main()
{
    imageAtomicAdd(tex[0], int2(0,0), uint64_t(0));
}

This is a repro for the crash.
It does not happen when the images are not in an array.
This does not cause a crash:

[[vk::binding(0,0)]] RWTexture2D<uint64_t> tex;

[shader("compute")]
[numthreads(1, 1, 1)]
void main()
{
    imageAtomicAdd(tex, int2(0,0), uint64_t(0));
}

[csyonghe]

import glsl; is another (preferred) way to use glsl flavored intrinsics. But we should look into the crash.

[ArielG-NV]

InterlockedAdd for 64bit types with images looks to only have a hlsl implementation:

[ForceInline]
void InterlockedAdd(__ref  int64_t dest,  int64_t value)
{
    __target_switch
    {
    case hlsl: __intrinsic_asm "InterlockedAdd";
    }
}

For now the glsl syntax variant of the operation should work as a workaround if only a spir-v/glsl target is needed: imageAtomicAdd

[Ipotrick]

How would i use imageAtomicAdd in the example code i have above? I dont understand.

[ArielG-NV]

I believe the following should work: imageAtomicAdd(tex, ivec2(0,0), uint64_t(0));

[Ipotrick]

slang does not recognize imageAromicAdd nor ivec2. Do i need to import a specific module for this to work?

[ArielG-NV]

the compiler flag -allow-glsl should enable GLSL syntax to work along-side HLSL syntax

[Ipotrick]

slang crashes when i try to use glsl syntax 😨

[ArielG-NV]

That is not the expected result.

[ArielG-NV]

This probably should be broken into 2 issues:
1 for the crash
1 for merging glsl.meta<->hlsl.meta image atomic operations

[Ipotrick]

I ll make another issue for the crash