fix(gpu): bound inverse LUT search and write transfer components explicitly

This commit is contained in:
Dylan M. Taylor
2026-11-15 23:25:31 -05:00
parent 5a7b4e903f
commit 3678cc9bec
+15 -12
View File
@@ -176,12 +176,16 @@ fn fxl_inv_1d(o: u32, i: u32, y: f32) -> f32 {
// partition_point over the table's values.
var lo = 0u;
var hi = n;
while (lo < hi) {
let mid = (lo + hi) / 2u;
if (data[b1 + mid * 3u + i] < y) {
lo = mid + 1u;
} else {
hi = mid;
// A u32 interval needs at most 32 halvings. A fixed trip count avoids
// a divergent data-dependent loop on software Vulkan without changing lookup.
for (var step = 0u; step < 32u; step++) {
if (lo < hi) {
let mid = lo + (hi - lo) / 2u;
if (data[b1 + mid * 3u + i] < y) {
lo = mid + 1u;
} else {
hi = mid;
}
}
}
let k = clamp(lo, 1u, n - 1u);
@@ -444,12 +448,11 @@ fn fxl_run(c0: vec3<f32>) -> vec3<f32> {
case 1u, 2u: {
let tf = u32(data[pc + 1u]);
let g = data[pc + 2u];
for (var i = 0u; i < 3u; i++) {
if (op == 1u) {
c[i] = fxl_decode(tf, g, c[i]);
} else {
c[i] = fxl_encode(tf, g, c[i]);
}
// Explicit components avoid llvmpipe's dynamically indexed vector writes.
if (op == 1u) {
c = vec3<f32>(fxl_decode(tf, g, c.x), fxl_decode(tf, g, c.y), fxl_decode(tf, g, c.z));
} else {
c = vec3<f32>(fxl_encode(tf, g, c.x), fxl_encode(tf, g, c.y), fxl_encode(tf, g, c.z));
}
pc += 3u;
}