// SPDX-License-Identifier: AGPL-3.0-or-later
// Reduce: Compute sum/max/min/mean over input (f64 canonical)
// CUDA equivalent: thrust::reduce, cub::DeviceReduce
// Algorithm: Tree reduction (work-efficient)
// Use cases: Loss computation, gradient accumulation
@group(0) @binding(0) var<storage, read> input: array<f64>;
@group(0) @binding(1) var<storage, read_write> output: array<f64>; // Partial results
struct Params {
size: u32,
operation: u32, // 0=Sum, 1=Max, 2=Min, 3=Mean
_pad0: u32,
_pad1: u32,
}
@group(0) @binding(2) var<uniform> params: Params;
var<workgroup> shared_data: array<f64, 256>;
@compute @workgroup_size(256)
fn main(
@builtin(global_invocation_id) global_id: vec3<u32>,
@builtin(local_invocation_id) local_id: vec3<u32>,
@builtin(workgroup_id) workgroup_id: vec3<u32>,
) {
let tid = local_id.x;
let gid = global_id.x;
// Load data into shared memory
var value: f64;
if (gid < params.size) {
value = input[gid];
} else {
// Initialize with identity based on operation
if (params.operation == 0u || params.operation == 3u) { // Sum, Mean
value = 0.0;
} else if (params.operation == 1u) { // Max
value = -1.7976931348623157e+308; // -DBL_MAX
} else if (params.operation == 2u) { // Min
value = 1.7976931348623157e+308; // DBL_MAX
} else {
value = 0.0;
}
}
shared_data[tid] = value;
workgroupBarrier();
// Tree reduction in shared memory
// Use exact same pattern as sum_reduce.wgsl which is known to work
for (var stride = 128u; stride > 0u; stride = stride / 2u) {
if (tid < stride) {
let a = shared_data[tid];
let b = shared_data[tid + stride];
var result: f64;
if (params.operation == 0u || params.operation == 3u) { // Sum, Mean
result = a + b;
} else if (params.operation == 1u) { // Max
result = max(a, b);
} else if (params.operation == 2u) { // Min
result = min(a, b);
} else {
result = a;
}
shared_data[tid] = result;
}
workgroupBarrier();
}
// Thread 0 writes workgroup result
if (tid == 0u) {
var result = shared_data[0];
// Deep Debt Fix: Implement Mean operation (divide by size)
if (params.operation == 3u) {
result = result / f64(params.size);
}
output[workgroup_id.x] = result;
}
}