Why doesn't my IJobChunk Burst vectorize?

Branches inside the loop, function calls Burst can't inline, and AoS layout (Vector3 from UnityEngine) all defeat SIMD. Use float3 (Unity.Mathematics), structure-of-arrays, and branchless math (math.select).

How do I check vectorization?

Burst Inspector. Find the loop. Look for SIMD register names (xmm0, ymm0). Scalar code uses xmm0[0]. If you see scalar everywhere, the loop didn't vectorize. Add the math.select rewrite and re-check.

Should I split jobs by component type?

For ECS, yes — one IJobChunk per archetype keeps memory access linear. Mixing many components per chunk reads scattered memory and prevents Burst from emitting wide loads.

Fix: Unity IJobChunk Burst Not Vectorizing

Quick answer: Use Unity.Mathematics types (float3, float4). Replace branches with math.select. Inspect Burst output for SIMD register names. Split by archetype so memory access stays linear.

IJobChunk runs in Burst but only 1.5x faster than C#. Burst Inspector shows scalar code where you expected SIMD. Branches and AoS layout sabotaged vectorization.

The Symptom

Burst-compiled job runs but performance is mediocre. Burst Inspector shows xmm0[0] (scalar) instead of xmm0 (vector). Hot loop has if/else patterns.

The Fix

using Unity.Mathematics;

[BurstCompile]
struct MoveJob : IJobChunk
{
    public ComponentTypeHandle<LocalTransform> transformH;
    [ReadOnly] public ComponentTypeHandle<Velocity> velocityH;
    public float dt;

    public void Execute(in ArchetypeChunk chunk, int idx, bool useEnabled, in v128 mask)
    {
        var transforms = chunk.GetNativeArray(ref transformH);
        var velocities = chunk.GetNativeArray(ref velocityH);

        for (int i = 0; i < chunk.Count; i++)
        {
            var t = transforms[i];
            var v = velocities[i].value;

            // Branchless: clamp speed
            var speed = math.length(v);
            var clamped = math.select(v, v * (10f / speed), speed > 10f);

            t.Position += clamped * dt;
            transforms[i] = t;
        }
    }
}

math.select replaces if/else. Burst can vectorize the entire loop body when there are no control-flow divergences.

Inspect

Window → Burst → Burst Inspector. Find your job. Show LLVM IR or assembly. Look for SIMD register names. Compare before/after the rewrite.

Verifying

Profile cycles per element. SIMD on float3 should give roughly 4x scalar. Use Burst’s timing column or Profiler markers.

“Mathematics types. Branchless. Linear memory. Burst vectorizes.”

Related Issues

For Burst BC1006, see BC1006. For Job data race, see data race.

math.select. float3. Wide registers.