InterlockedAdd in HLSL, and what to do for floats

Count with InterlockedAdd in an HLSL compute shader on Vulkan, Metal and WebGPU, and add floats atomically with a compare-exchange loop, since there is no float InterlockedAdd. Includes what OneShader 1.2.0 doesn’t compile.

InterlockedAdd, Min, Max, And, Or, Xor, Exchange and CompareExchange work on 32-bit int and uint, in RW structured and byte-address buffers and in groupshared memory. Since 1.2.0 they also work on int64_t and uint64_t on SPIR-V (see below). A histogram shows both the plain counter and the float sum, Histogram.hlsl:

Texture2D<float4> Image : register(t0);
RWStructuredBuffer<uint> Bins : register(u1);     // 256 bins
RWStructuredBuffer<uint> Total : register(u2);    // one float, stored as its bits
 
// There is no float InterlockedAdd: add with a compare-exchange loop on the bits.
void AtomicAddFloat(uint index, float value)
{
    uint expected = Total[index];
    for (;;)
    {
        uint desired = asuint(asfloat(expected) + value);
        uint original;
        InterlockedCompareExchange(Total[index], expected, desired, original);
        if (original == expected)
            break;
        expected = original;
    }
}
 
[shader("compute")]
[numthreads(16, 16, 1)]
void CSMain(uint3 id : SV_DispatchThreadID)
{
    float3 rgb = Image[id.xy].rgb;
    float luma = dot(rgb, float3(0.2126, 0.7152, 0.0722));
    uint bin = min(uint(luma * 256.0), 255u);
 
    InterlockedAdd(Bins[bin], 1u);   // 32-bit int or uint
    AtomicAddFloat(0, luma);
}

oneshader --target spirv --target msl --target wgsl Histogram.hlsl -Fo out/histogram
spirv-val --target-env vulkan1.2 out/histogram.spv
naga out/histogram.wgsl

The compile succeeds for all three targets and writes histogram.spv, histogram.metal and histogram.wgsl. The SPIR-V passes spirv-val. The Metal output uses atomic_fetch_add_explicit and atomic_compare_exchange_weak_explicit, and the WGSL output uses atomicAdd on array<atomic<u32>>. naga, however, rejects the WGSL:

error: Function [0] 'osh_atomic_cas_storage_u32' is invalid
   ┌─ histogram.wgsl:7:1
   │  
 7 │ ╭ fn osh_atomic_cas_storage_u32(p: ptr<storage, atomic<u32>, read_write>, compare: u32, value: u32) -> u32 {
...
   = Argument 'p' at index 0 is a pointer of space Storage { access: StorageAccess(LOAD | STORE) }, which can't be passed into functions.

Limits in 1.2.0

• Compare-exchange on WGSL. For InterlockedCompareExchange and InterlockedCompareStore (on buffers and on groupshared memory), and for InterlockedMin/Max on signed int buffers, the WGSL that 1.2.0 writes passes a pointer to a helper function, and naga rejects that. So the float add above works on Vulkan and Metal, but not on WebGPU yet. InterlockedAdd, And, Or, Xor and Exchange, and Min/Max on uint, produce WGSL that validates.

• No float atomics. InterlockedAdd on a float is an error that says what to do:

  floatadd.hlsl(3,62): error OSH0314: 'InterlockedAdd' works on int, uint, int64_t and uint64_t, but the first argument is 'float'. Float atomics are not supported; reinterpret the data as uint with asuint()/asfloat().

• 64-bit atomics are mostly SPIR-V. On SPIR-V, 64-bit Interlocked* work on structured buffers and groupshared memory, and the device needs Int64Atomics (in the entry’s RequiredFeatures). Metal 2.4 has only unsigned InterlockedMax/InterlockedMin on a buffer whose result isn’t used, and WGSL has no 64-bit integers: other forms are errors (OSH5102, OSH5202) when those targets are selected. See 16-bit and 64-bit types.

• No atomics on textures for Metal and WGSL (Metal needs 3.1). Use a buffer.

Good to know

• The compare-exchange loop retries until no other thread changed the value between the read and the write. Under heavy contention it’s slow, so reduce first: sum within the thread group in groupshared memory or with WaveActiveSum, and do one atomic per group or per wave. See Wave intrinsics.
• Floating-point addition isn’t associative, so a sum built from atomics differs in the last bits from run to run, in any compiler.
• The two-argument form InterlockedAdd(dest, value) discards the old value. The three-argument form returns it.

For the dispatch and layouts around a compute shader, see Write a compute shader for Vulkan and WebGPU.