#version 450 #extension GL_GOOGLE_include_directive : require // YIN's difference function for tuner.comp, spread over many workgroups: // d(tau) = sum over the newest W samples of (x[j] - x[j - tau])^2. Each // workgroup takes LAGS_PER_GROUP lags, with THREADS / LAGS_PER_GROUP // threads summing every so-many term of each, then adding up. layout(local_size_x = 256) in; #include "TunerParams.glsl" layout(set = 0, binding = 0) readonly buffer Tape { float tape[]; }; layout(set = 1, binding = 0) buffer Difference { float d[]; }; // Match layouts.zig and tuner.comp. const uint W = 2048; const uint LAGS = 1700; const uint THREADS = 256; const uint LAGS_PER_GROUP = 16; const uint PER_LAG = THREADS / LAGS_PER_GROUP; shared float x[W + LAGS]; shared float partial[THREADS]; void main() { uint t = gl_LocalInvocationID.x; int first = int(u.end) - int(W + LAGS); for (uint i = t; i < W + LAGS; i += THREADS) { int at = first + int(i); x[i] = at >= 0 && at < int(u.samples) ? tape[u.offset + uint(at)] : 0.0; } barrier(); uint tau = gl_WorkGroupID.x * LAGS_PER_GROUP + t / PER_LAG; uint lane = t % PER_LAG; float sum = 0.0; if (tau < LAGS) { for (uint j = lane; j < W; j += PER_LAG) { float diff = x[LAGS + j] - x[LAGS + j - tau]; sum += diff * diff; } } partial[t] = sum; barrier(); for (uint stride = PER_LAG / 2u; stride > 0u; stride >>= 1u) { if (lane < stride) partial[t] += partial[t + stride]; barrier(); } if (lane == 0u && tau < LAGS) d[tau] = partial[t]; }