Compiler Explorer

Source code

struct float2_SoA
{
    float* x;
    float* y;
};

typedef float<3> float3;

static uint seed = 0x12345678;
inline uint RandomUInt()
{
	seed ^= seed << 13;
	seed ^= seed >> 17;
	seed ^= seed << 5;
	return seed;
}

#pragma ignore warning //uint to float is slow
inline float RandomFloat() { return RandomUInt() * 2.3283064365387e-10f; }

unmasked void ParticleSystem_Tick( const uniform int32 count, uniform float2_SoA &pos, uniform float2_SoA &dir, uniform uint32 frame[], 
                                 const uniform int32 frameChange[], const uniform int32 mapWidth, const uniform int32 mapHeight,
                                 const uniform float3 peaks[], const uniform int32 peaksSize )
{
    float mapWidthMinusOnef = (float) (mapWidth-1);
    float mapHeightf = (float) mapHeight;

for (uniform int32 p = 0; p < peaksSize; p++)
    {
        varying float peakX = peaks[p].x;
        varying float peakY = peaks[p].y;
        varying float peakZ = peaks[p].z; // multiply by 0.02f could be hoisted here

foreach(i = 0 ... count)
        {
            // these operations don't need to be masked.
            pos.x[i] += dir.x[i];
            pos.y[i] += dir.y[i];
            
            dir.y[i] *= 0.95f;

// inside this scope should be masked
            if ( pos.x[i] < 0.0f )
            {
                pos.x[i] = (varying float) mapWidthMinusOnef;

#pragma ignore warning // modulus/integer division is slow
                pos.y[i] = RandomFloat() * mapHeight;

dir.x[i] = -1.0f - RandomFloat() * 2.0f;
                dir.y[i] = 0.0f;
            }

// and this bit here does not need be masked.
            float toPeak_x = peakX - pos.x[i];
            float toPeak_y = peakY - pos.y[i];

float toPeak_dot = toPeak_x * toPeak_x + toPeak_y * toPeak_y;
            float toPeak_len = sqrt(toPeak_dot);
            float g = peakZ * 0.02f / toPeak_len;  // this scalar should be able to be applied to line 32

float invLen = 1.0 / toPeak_len;  // this would be beneficical to reorder above the previous statement and save the divide
            toPeak_y *= invLen;

dir.y[i] -= toPeak_y * g;

dir.y[i] += RandomFloat() * 0.05f - 0.025f;
            frame[i] = (frame[i] + frameChange[i] + 256) & 256;
        }
    }
}