// SPDX-License-Identifier: AGPL-3.0-or-later
// IoU Loss - Intersection over Union loss (f64 canonical)
// Direct optimization of IoU metric
// Used in segmentation and object detection
//
// Algorithm:
// IoU = (intersection + smooth_val) / (union + smooth_val)
// IoU Loss = 1 - IoU
// where intersection = sum(predictions * targets)
// union = sum(predictions + targets - predictions * targets)
//
// Uses workgroup shared memory for correct float reduction (not broken atomic bitcast)
struct Params {
size: u32,
smooth_val: f64,
num_partials: u32, // number of workgroup partial results to reduce in pass 2
_pad1: u32,
}
@group(0) @binding(0) var<storage, read> predictions: array<f64>;
@group(0) @binding(1) var<storage, read> targets: array<f64>;
@group(0) @binding(2) var<storage, read_write> intersection_buffer: array<f64>; // partial sums per workgroup
@group(0) @binding(3) var<storage, read_write> union_buffer: array<f64>; // partial sums per workgroup
@group(0) @binding(4) var<storage, read_write> output: array<f64>; // [1] - final loss
@group(0) @binding(5) var<uniform> params: Params;
var<workgroup> shared_intersection: array<f64, 256>;
var<workgroup> shared_union: 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 idx = global_id.x;
let local_idx = local_id.x;
var local_intersection: f64 = 0.0;
var local_union: f64 = 0.0;
if (idx < params.size) {
let pred = predictions[idx];
let targ = targets[idx];
local_intersection = pred * targ;
local_union = pred + targ - local_intersection;
}
shared_intersection[local_idx] = local_intersection;
shared_union[local_idx] = local_union;
workgroupBarrier();
// Parallel reduction in shared memory
var stride = 128u;
while (stride >= 1u) {
if (local_idx < stride && local_idx + stride < 256u) {
shared_intersection[local_idx] = shared_intersection[local_idx] + shared_intersection[local_idx + stride];
shared_union[local_idx] = shared_union[local_idx] + shared_union[local_idx + stride];
}
workgroupBarrier();
stride = stride / 2u;
}
// First thread writes workgroup partial result
if (local_idx == 0u) {
intersection_buffer[workgroup_id.x] = shared_intersection[0];
union_buffer[workgroup_id.x] = shared_union[0];
}
}
// Second pass: sum partial results and compute final loss (single workgroup)
@compute @workgroup_size(256)
fn compute_loss(
@builtin(global_invocation_id) global_id: vec3<u32>,
@builtin(local_invocation_id) local_id: vec3<u32>
) {
let idx = global_id.x;
let local_idx = local_id.x;
let num_wg = params.num_partials;
// Load partial results into shared memory
var partial_int: f64 = 0.0;
var partial_un: f64 = 0.0;
if (idx < num_wg) {
partial_int = intersection_buffer[idx];
partial_un = union_buffer[idx];
}
shared_intersection[local_idx] = partial_int;
shared_union[local_idx] = partial_un;
workgroupBarrier();
// Reduce
var stride = 128u;
while (stride >= 1u) {
if (local_idx < stride && local_idx + stride < 256u) {
shared_intersection[local_idx] = shared_intersection[local_idx] + shared_intersection[local_idx + stride];
shared_union[local_idx] = shared_union[local_idx] + shared_union[local_idx + stride];
}
workgroupBarrier();
stride = stride / 2u;
}
if (local_idx == 0u) {
let intersection = shared_intersection[0];
let union_val = shared_union[0];
let iou = (intersection + params.smooth_val) / (union_val + params.smooth_val + 1e-8);
output[0] = 1.0 - iou;
}
}