diff --git a/crates/gpu/src/shaders/fx_lut.wgsl b/crates/gpu/src/shaders/fx_lut.wgsl index 96f325642..f9081f834 100644 --- a/crates/gpu/src/shaders/fx_lut.wgsl +++ b/crates/gpu/src/shaders/fx_lut.wgsl @@ -130,12 +130,15 @@ fn fxl_shape(at: u32, n: u32, x: f32, flip: bool) -> f32 { // partition_point(|p| p.x < x). var lo = 0u; var hi = n; - while (lo < hi) { - let mid = (lo + hi) / 2u; - if (data[at + mid * 2u + kx] < x) { - lo = mid + 1u; - } else { - hi = mid; + // Bound the partition search as for inverse 1D LUTs below. + for (var step = 0u; step < 32u; step++) { + if (lo < hi) { + let mid = lo + (hi - lo) / 2u; + if (data[at + mid * 2u + kx] < x) { + lo = mid + 1u; + } else { + hi = mid; + } } } let i = clamp(lo, 1u, n - 1u); @@ -156,13 +159,22 @@ fn fxl_shaper_at(s: u32, k: u32) -> vec2 { return vec2(at, u32(data[s + k])); } +// Evaluate independent shaper channels without dynamically writing vector lanes. +fn fxl_shaper_apply(sh: u32, c: vec3, flip: bool) -> vec3 { + let r = fxl_shaper_at(sh, 0u); + let g = fxl_shaper_at(sh, 1u); + let b = fxl_shaper_at(sh, 2u); + return vec3( + fxl_shape(r.x, r.y, c.x, flip), + fxl_shape(g.x, g.y, c.y, flip), + fxl_shape(b.x, b.y, c.z, flip) + ); +} + fn fxl_file_fwd(o: u32, sh: i32, c0: vec3, interp: u32) -> vec3 { var c = c0; if (sh >= 0) { - for (var k = 0u; k < 3u; k++) { - let a = fxl_shaper_at(u32(sh), k); - c[k] = fxl_shape(a.x, a.y, c[k], false); - } + c = fxl_shaper_apply(u32(sh), c, false); } return fxl_lut_apply(o, c, interp); } @@ -209,10 +221,7 @@ fn fxl_file(o: u32, sh: i32, c: vec3, interp: u32, inverse: bool) -> vec3(fxl_inv_1d(o, 0u, c.x), fxl_inv_1d(o, 1u, c.y), fxl_inv_1d(o, 2u, c.z)); } if (sh >= 0) { - for (var k = 0u; k < 3u; k++) { - let a = fxl_shaper_at(u32(sh), k); - x[k] = fxl_shape(a.x, a.y, x[k], true); - } + x = fxl_shaper_apply(u32(sh), x, true); } return x; } diff --git a/crates/gpu/src/tests/fx_lut.rs b/crates/gpu/src/tests/fx_lut.rs index f4bac9f7d..acac88835 100644 --- a/crates/gpu/src/tests/fx_lut.rs +++ b/crates/gpu/src/tests/fx_lut.rs @@ -112,6 +112,20 @@ fn profile_conversion_on_a_fresh_device_matches_cpu() { effect_direct("ec.utility.colorprofileconverter", &[("inputProfile", e(1)), ("outputProfile", e(3)), ("intent", e(0))], false); } +// Unequal CSP channel point counts exercise shaper offsets on a fresh device. +#[test] +fn csp_shaper_on_a_fresh_device_matches_cpu() { + effect_direct("ec.color.ociofile", &[("file", s(&csp())), ("interpolation", e(1))], false); + effect_direct("ec.color.ociofile", &[("file", s(&csp())), ("direction", e(1))], false); +} + +// A 1D CSP takes the direct inverse-shaper path (flip=true), unlike a 3D CSP. +#[test] +fn inverse_1d_csp_shaper_on_a_fresh_device_matches_cpu() { + let lut = "CSPLUTV100\n1D\n3\n0 0.5 1\n0 0.3 1\n2\n0 1\n0 1\n4\n0 0.2 0.6 1\n0 0.4 0.7 1\n2\n0 0 0\n1 1 1\n"; + effect_direct("ec.color.ociofile", &[("file", s(lut)), ("direction", e(1))], false); +} + #[test] fn ocio_file_transform() { for interp in 1..4 {