@woosh/meep-engine 3.43.0 → 3.45.0

package/src/shade/renderer/texture/encode/chunk_bc6h_encode_block_mode11.js

@@ -1,346 +1,349 @@
-import { CodeChunk } from "../../shader/compiler/CodeChunk.js";
-import { chunk_bc6h_pack_mode11 } from "./chunk_bc6h_pack_mode11.js";
-
-/**
- * Encode sixteen HDR RGB texels as one BC6H block in mode 11, the unsigned format.
- *
- * **The line is fitted in the half-float bit domain, because that is the domain the decoder
- * interpolates in.** A BC6H endpoint is a half-float bit pattern quantised to ten bits; the decoder
- * unquantises it to a sixteen-bit integer, interpolates the two ends linearly with BC7's four-bit
- * weights, scales the result by 31/64 and reinterprets the integer *as a half*. So the palette is
- * linear in bit patterns, not in values — which is a log-like scale, sixteen steps between two
- * ends a few octaves apart being spaced by ratio rather than by difference. Fitting in that domain
- * is what Narkowicz's real-time BC6H does with `f32tof16` before anything else, and it is why an
- * error measured there is a relative error, which is what an HDR image wants judged.
- *
- * The fit is {@link chunk_bc1_encode_block}'s line in that domain: the principal axis, the extreme
- * projections, two rounds of least squares with the sixteen decode weights, then quantisation to
- * ten bits and one round of coordinate descent over the six stored components, every candidate
- * scored against the palette the decoder builds. The anchor is honoured by writing the block the
- * other way up when texel 0 lands in the upper half.
- *
- * **Mode 11 only.** It has the finest endpoints of the one-region modes and no delta to fit; the
- * two-region modes are where a hard edge inside a block would gain, and they are a later chunk. On
- * the environment maps the format exists for, Narkowicz measured his mode-11-only fast path within
- * reach of offline encoders' fast presets.
- *
- * Half floats hold no infinity or NaN here and no sign: a negative texel clamps to zero, and
- * anything over the largest finite half to that. The conversion to half bits is written out rather
- * than taken from `pack2x16float`, so that the emulator and every device round the same way.
- *
- * @type {CodeChunk}
- */
-export const chunk_bc6h_encode_block_mode11 = CodeChunk.from(
-    //language=WGSL
-    `
-// the bit pattern of the half nearest a non-negative finite value, ties to even, as f32tof16 gives
-fn bc6h_half_bits(x: f32) -> f32 {
-    let v = clamp(select(x, 0.0, x != x), 0.0, 65504.0);
-
-    if (v == 0.0) {
-        return 0.0;
-    }
-
-    let bits = bitcast<u32>(v);
-    let exponent = i32((bits >> 23u) & 0xFFu) - 127;
-
-    // a half's subnormals count in units of 2^-24; the tie goes to even, written out because the
-    // emulator's round is half-up and a device's is half-even
-    if (exponent < -14) {
-        let scaled = v * 16777216.0;
-        let whole = floor(scaled);
-        let fraction = scaled - whole;
-
-        let up = fraction > 0.5 || (fraction == 0.5 && (u32(whole) & 1u) == 1u);
-
-        return whole + select(0.0, 1.0, up);
-    }
-
-    let mantissa = bits & 0x7FFFFFu;
-    let kept = mantissa >> 13u;
-    let dropped = mantissa & 0x1FFFu;
-
-    var pattern = (u32(exponent + 15) << 10u) | kept;
-
-    if (dropped > 0x1000u || (dropped == 0x1000u && (kept & 1u) == 1u)) {
-        pattern += 1u;
-    }
-
-    return f32(min(pattern, 0x7BFFu));
-}
-
-// BC7's four-bit interpolation weights, which BC6H shares
-fn bc6h_weight4(index: u32) -> f32 {
-    var weights = array<f32, 16>(0.0, 4.0, 9.0, 13.0, 17.0, 21.0, 26.0, 30.0, 34.0, 38.0, 43.0, 47.0, 51.0, 55.0, 60.0, 64.0);
-
-    return weights[index];
-}
-
-// a ten-bit endpoint as the decoder unquantises it before interpolating
-fn bc6h_unquantize(q: vec3<f32>) -> vec3<f32> {
-    let scaled = floor((q * 65536.0 + 32768.0) / 1024.0);
-
-    return select(select(scaled, vec3<f32>(65535.0), q >= vec3<f32>(1023.0)), vec3<f32>(0.0), q <= vec3<f32>(0.0));
-}
-
-// a palette entry, as half bits, from two unquantised ends
-fn bc6h_palette_entry(u0: vec3<f32>, u1: vec3<f32>, index: u32) -> vec3<f32> {
-    let w = bc6h_weight4(index);
-
-    let interpolated = floor((u0 * (64.0 - w) + u1 * w + 32.0) / 64.0);
-
-    return floor(interpolated * 31.0 / 64.0);
-}
-
-// a value in half bits quantised to the ten bits the block stores
-fn bc6h_quantize(bits: vec3<f32>) -> vec3<f32> {
-    return clamp(round(bits * 1024.0 / 31744.0), vec3<f32>(0.0), vec3<f32>(1023.0));
-}
-
-// the index whose palette entry is nearest a texel, and the squared distance in half bits
-fn bc6h_nearest(u0: vec3<f32>, u1: vec3<f32>, texel: vec3<f32>) -> vec2<f32> {
-    let p0 = bc6h_palette_entry(u0, u1, 0u);
-    let p15 = bc6h_palette_entry(u0, u1, 15u);
-
-    let direction = p15 - p0;
-    let span = dot(direction, direction);
-
-    var estimate = 0.0;
-
-    if (span > 0.0) {
-        estimate = clamp(round(dot(texel - p0, direction) / span * 15.0), 0.0, 15.0);
-    }
-
-    var best = estimate;
-    var best_distance = 1e30;
-
-    for (var offset = -1.0; offset <= 1.0; offset += 1.0) {
-        let candidate = clamp(estimate + offset, 0.0, 15.0);
-
-        let delta = bc6h_palette_entry(u0, u1, u32(candidate)) - texel;
-        let distance = dot(delta, delta);
-
-        if (distance < best_distance) {
-            best_distance = distance;
-            best = candidate;
-        }
-    }
-
-    return vec2<f32>(best, best_distance);
-}
-
-fn bc6h_score(texels: array<vec3<f32>, 16>, q0: vec3<f32>, q1: vec3<f32>) -> f32 {
-    let u0 = bc6h_unquantize(q0);
-    let u1 = bc6h_unquantize(q1);
-
-    var error = 0.0;
-
-    for (var i = 0u; i < 16u; i++) {
-        error += bc6h_nearest(u0, u1, texels[i]).y;
-    }
-
-    return error;
-}
-
-// the ends that best reproduce the texels at the weights they are nearest to on a real line
-fn bc6h_refine(texels: array<vec3<f32>, 16>, e0: vec3<f32>, e1: vec3<f32>) -> array<vec3<f32>, 2> {
-    let direction = e1 - e0;
-    let span = dot(direction, direction);
-
-    var aa = 0.0;
-    var ab = 0.0;
-    var bb = 0.0;
-    var ap = vec3<f32>(0.0);
-    var bp = vec3<f32>(0.0);
-
-    for (var i = 0u; i < 16u; i++) {
-        var estimate = 0.0;
-
-        if (span > 0.0) {
-            estimate = clamp(round(dot(texels[i] - e0, direction) / span * 15.0), 0.0, 15.0);
-        }
-
-        var best_weight = 0.0;
-        var best_distance = 1e30;
-
-        for (var offset = -1.0; offset <= 1.0; offset += 1.0) {
-            let candidate = clamp(estimate + offset, 0.0, 15.0);
-            let weight = bc6h_weight4(u32(candidate)) / 64.0;
-
-            let delta = mix(e0, e1, weight) - texels[i];
-            let distance = dot(delta, delta);
-
-            if (distance < best_distance) {
-                best_distance = distance;
-                best_weight = weight;
-            }
-        }
-
-        let b = best_weight;
-        let a = 1.0 - b;
-
-        aa += a * a;
-        ab += a * b;
-        bb += b * b;
-        ap += a * texels[i];
-        bp += b * texels[i];
-    }
-
-    let determinant = aa * bb - ab * ab;
-
-    var out: array<vec3<f32>, 2>;
-
-    if (abs(determinant) < 1e-6) {
-        out[0] = e0;
-        out[1] = e1;
-
-        return out;
-    }
-
-    out[0] = (bb * ap - ab * bp) / determinant;
-    out[1] = (aa * bp - ab * ap) / determinant;
-
-    return out;
-}
-
-fn bc6h_with_component(v: vec3<f32>, channel: u32, value: f32) -> vec3<f32> {
-    return vec3<f32>(
-        select(v.x, value, channel == 0u),
-        select(v.y, value, channel == 1u),
-        select(v.z, value, channel == 2u)
-    );
-}
-
-fn bc6h_encode_block_mode11(texels: array<vec3<f32>, 16>) -> vec4<u32> {
-    var bits: array<vec3<f32>, 16>;
-
-    var low = vec3<f32>(65535.0);
-    var high = vec3<f32>(0.0);
-    var mean = vec3<f32>(0.0);
-
-    for (var i = 0u; i < 16u; i++) {
-        let v = vec3<f32>(bc6h_half_bits(texels[i].x), bc6h_half_bits(texels[i].y), bc6h_half_bits(texels[i].z));
-
-        bits[i] = v;
-        low = min(low, v);
-        high = max(high, v);
-        mean += v;
-    }
-
-    mean = mean / 16.0;
-
-    let extent = high - low;
-
-    var e0 = mean;
-    var e1 = mean;
-
-    if (max(extent.x, max(extent.y, extent.z)) >= 0.5) {
-        var cov = array<f32, 6>(0.0, 0.0, 0.0, 0.0, 0.0, 0.0);
-
-        for (var i = 0u; i < 16u; i++) {
-            let d = bits[i] - mean;
-
-            cov[0] += d.x * d.x;
-            cov[1] += d.x * d.y;
-            cov[2] += d.x * d.z;
-            cov[3] += d.y * d.y;
-            cov[4] += d.y * d.z;
-            cov[5] += d.z * d.z;
-        }
-
-        var axis = extent;
-
-        for (var iteration = 0u; iteration < 4u; iteration++) {
-            let next = vec3<f32>(
-                cov[0] * axis.x + cov[1] * axis.y + cov[2] * axis.z,
-                cov[1] * axis.x + cov[3] * axis.y + cov[4] * axis.z,
-                cov[2] * axis.x + cov[4] * axis.y + cov[5] * axis.z
-            );
-
-            let magnitude = length(next);
-
-            if (magnitude < 1e-3) {
-                break;
-            }
-
-            axis = next / magnitude;
-        }
-
-        axis = normalize(axis);
-
-        var t_min = 1e30;
-        var t_max = -1e30;
-
-        for (var i = 0u; i < 16u; i++) {
-            let t = dot(bits[i] - mean, axis);
-
-            t_min = min(t_min, t);
-            t_max = max(t_max, t);
-        }
-
-        e0 = mean + axis * t_min;
-        e1 = mean + axis * t_max;
-
-        for (var round = 0u; round < 2u; round++) {
-            let refined = bc6h_refine(bits, e0, e1);
-
-            e0 = clamp(refined[0], vec3<f32>(0.0), vec3<f32>(31743.0));
-            e1 = clamp(refined[1], vec3<f32>(0.0), vec3<f32>(31743.0));
-        }
-    }
-
-    var q0 = bc6h_quantize(e0);
-    var q1 = bc6h_quantize(e1);
-
-    var best = bc6h_score(bits, q0, q1);
-
-    // coordinate descent over the six stored components, one round
-    for (var component = 0u; component < 6u; component++) {
-        let channel = component % 3u;
-
-        for (var direction = -1.0; direction <= 1.0; direction += 2.0) {
-            var c0 = q0;
-            var c1 = q1;
-
-            if (component < 3u) {
-                c0 = bc6h_with_component(q0, channel, clamp(q0[channel] + direction, 0.0, 1023.0));
-            } else {
-                c1 = bc6h_with_component(q1, channel, clamp(q1[channel] + direction, 0.0, 1023.0));
-            }
-
-            let error = bc6h_score(bits, c0, c1);
-
-            if (error < best) {
-                best = error;
-                q0 = c0;
-                q1 = c1;
-            }
-        }
-    }
-
-    var indices: array<u32, 16>;
-
-    let u0 = bc6h_unquantize(q0);
-    let u1 = bc6h_unquantize(q1);
-
-    for (var i = 0u; i < 16u; i++) {
-        indices[i] = u32(bc6h_nearest(u0, u1, bits[i]).x);
-    }
-
-    // the anchor: texel 0's top bit is implied zero
-    if (indices[0] >= 8u) {
-        let swap = q0;
-
-        q0 = q1;
-        q1 = swap;
-
-        for (var i = 0u; i < 16u; i++) {
-            indices[i] = 15u - indices[i];
-        }
-    }
-
-    return bc6h_pack_mode11(vec3<u32>(q0), vec3<u32>(q1), indices);
-}
-`, [
-        chunk_bc6h_pack_mode11,
-    ]);
+import { CodeChunk } from "../../shader/compiler/CodeChunk.js";
+import { chunk_bc6h_pack_mode11 } from "./chunk_bc6h_pack_mode11.js";
+
+/**
+ * Encode sixteen HDR RGB texels as one BC6H block in mode 11, the unsigned format.
+ *
+ * **The line is fitted in the half-float bit domain, because that is the domain the decoder
+ * interpolates in.** A BC6H endpoint is a half-float bit pattern quantised to ten bits; the decoder
+ * unquantises it to a sixteen-bit integer, interpolates the two ends linearly with BC7's four-bit
+ * weights, scales the result by 31/64 and reinterprets the integer *as a half*. So the palette is
+ * linear in bit patterns, not in values — which is a log-like scale, sixteen steps between two
+ * ends a few octaves apart being spaced by ratio rather than by difference. Fitting in that domain
+ * is what Narkowicz's real-time BC6H does with `f32tof16` before anything else, and it is why an
+ * error measured there is a relative error, which is what an HDR image wants judged.
+ *
+ * The fit is {@link chunk_bc1_encode_block}'s line in that domain: the principal axis, the extreme
+ * projections, two rounds of least squares with the sixteen decode weights, then quantisation to
+ * ten bits and one round of coordinate descent over the six stored components, every candidate
+ * scored against the palette the decoder builds. The anchor is honoured by writing the block the
+ * other way up when texel 0 lands in the upper half.
+ *
+ * **Mode 11 only.** It has the finest endpoints of the one-region modes and no delta to fit; the
+ * two-region modes are where a hard edge inside a block would gain, and they are a later chunk. On
+ * the environment maps the format exists for, Narkowicz measured his mode-11-only fast path within
+ * reach of offline encoders' fast presets.
+ *
+ * Half floats hold no infinity or NaN here and no sign: a negative texel clamps to zero, and
+ * anything over the largest finite half to that. The conversion to half bits is written out rather
+ * than taken from `pack2x16float`, so that the emulator and every device round the same way.
+ *
+ * @type {CodeChunk}
+ */
+export const chunk_bc6h_encode_block_mode11 = CodeChunk.from(
+    //language=WGSL
+    `
+// the bit pattern of the half nearest a non-negative finite value, ties to even, as f32tof16 gives
+fn bc6h_half_bits(x: f32) -> f32 {
+    let v = clamp(select(x, 0.0, x != x), 0.0, 65504.0);
+
+    if (v == 0.0) {
+        return 0.0;
+    }
+
+    let bits = bitcast<u32>(v);
+    let exponent = i32((bits >> 23u) & 0xFFu) - 127;
+
+    // a half's subnormals count in units of 2^-24; the tie goes to even, written out because the
+    // emulator's round is half-up and a device's is half-even
+    if (exponent < -14) {
+        let scaled = v * 16777216.0;
+        let whole = floor(scaled);
+        let fraction = scaled - whole;
+
+        let up = fraction > 0.5 || (fraction == 0.5 && (u32(whole) & 1u) == 1u);
+
+        return whole + select(0.0, 1.0, up);
+    }
+
+    let mantissa = bits & 0x7FFFFFu;
+    let kept = mantissa >> 13u;
+    let dropped = mantissa & 0x1FFFu;
+
+    var pattern = (u32(exponent + 15) << 10u) | kept;
+
+    if (dropped > 0x1000u || (dropped == 0x1000u && (kept & 1u) == 1u)) {
+        pattern += 1u;
+    }
+
+    return f32(min(pattern, 0x7BFFu));
+}
+
+// BC7's four-bit interpolation weights, which BC6H shares: 0, 4, 9, 13, 17, 21, 26, 30, 34, 38, 43,
+// 47, 51, 55, 60, 64, every one of which this computes. Computed rather than looked up, as
+// bc7_weight4 is: a function-local table indexed at run time lives in local memory, and it was an
+// eighth of this encoder's time on a 4090
+fn bc6h_weight4(index: u32) -> f32 {
+    return f32((index * 64u + 7u) / 15u);
+}
+
+// a ten-bit endpoint as the decoder unquantises it before interpolating
+fn bc6h_unquantize(q: vec3<f32>) -> vec3<f32> {
+    let scaled = floor((q * 65536.0 + 32768.0) / 1024.0);
+
+    return select(select(scaled, vec3<f32>(65535.0), q >= vec3<f32>(1023.0)), vec3<f32>(0.0), q <= vec3<f32>(0.0));
+}
+
+// a palette entry, as half bits, from two unquantised ends
+fn bc6h_palette_entry(u0: vec3<f32>, u1: vec3<f32>, index: u32) -> vec3<f32> {
+    let w = bc6h_weight4(index);
+
+    let interpolated = floor((u0 * (64.0 - w) + u1 * w + 32.0) / 64.0);
+
+    return floor(interpolated * 31.0 / 64.0);
+}
+
+// a value in half bits quantised to the ten bits the block stores
+fn bc6h_quantize(bits: vec3<f32>) -> vec3<f32> {
+    return clamp(round(bits * 1024.0 / 31744.0), vec3<f32>(0.0), vec3<f32>(1023.0));
+}
+
+// the index whose palette entry is nearest a texel, and the squared distance in half bits
+fn bc6h_nearest(u0: vec3<f32>, u1: vec3<f32>, texel: vec3<f32>) -> vec2<f32> {
+    let p0 = bc6h_palette_entry(u0, u1, 0u);
+    let p15 = bc6h_palette_entry(u0, u1, 15u);
+
+    let direction = p15 - p0;
+    let span = dot(direction, direction);
+
+    var estimate = 0.0;
+
+    if (span > 0.0) {
+        estimate = clamp(round(dot(texel - p0, direction) / span * 15.0), 0.0, 15.0);
+    }
+
+    var best = estimate;
+    var best_distance = 1e30;
+
+    // counted in integers: counted in floats, the same loops made the whole encoder a fifth slower on
+    // an RTX 4090, the bytes identical
+    for (var offset = -1; offset <= 1; offset++) {
+        let candidate = clamp(estimate + f32(offset), 0.0, 15.0);
+
+        let delta = bc6h_palette_entry(u0, u1, u32(candidate)) - texel;
+        let distance = dot(delta, delta);
+
+        if (distance < best_distance) {
+            best_distance = distance;
+            best = candidate;
+        }
+    }
+
+    return vec2<f32>(best, best_distance);
+}
+
+fn bc6h_score(texels: array<vec3<f32>, 16>, q0: vec3<f32>, q1: vec3<f32>) -> f32 {
+    let u0 = bc6h_unquantize(q0);
+    let u1 = bc6h_unquantize(q1);
+
+    var error = 0.0;
+
+    for (var i = 0u; i < 16u; i++) {
+        error += bc6h_nearest(u0, u1, texels[i]).y;
+    }
+
+    return error;
+}
+
+// the ends that best reproduce the texels at the weights they are nearest to on a real line
+fn bc6h_refine(texels: array<vec3<f32>, 16>, e0: vec3<f32>, e1: vec3<f32>) -> array<vec3<f32>, 2> {
+    let direction = e1 - e0;
+    let span = dot(direction, direction);
+
+    var aa = 0.0;
+    var ab = 0.0;
+    var bb = 0.0;
+    var ap = vec3<f32>(0.0);
+    var bp = vec3<f32>(0.0);
+
+    for (var i = 0u; i < 16u; i++) {
+        var estimate = 0.0;
+
+        if (span > 0.0) {
+            estimate = clamp(round(dot(texels[i] - e0, direction) / span * 15.0), 0.0, 15.0);
+        }
+
+        var best_weight = 0.0;
+        var best_distance = 1e30;
+
+        for (var offset = -1; offset <= 1; offset++) {
+            let candidate = clamp(estimate + f32(offset), 0.0, 15.0);
+            let weight = bc6h_weight4(u32(candidate)) / 64.0;
+
+            let delta = mix(e0, e1, weight) - texels[i];
+            let distance = dot(delta, delta);
+
+            if (distance < best_distance) {
+                best_distance = distance;
+                best_weight = weight;
+            }
+        }
+
+        let b = best_weight;
+        let a = 1.0 - b;
+
+        aa += a * a;
+        ab += a * b;
+        bb += b * b;
+        ap += a * texels[i];
+        bp += b * texels[i];
+    }
+
+    let determinant = aa * bb - ab * ab;
+
+    var out: array<vec3<f32>, 2>;
+
+    if (abs(determinant) < 1e-6) {
+        out[0] = e0;
+        out[1] = e1;
+
+        return out;
+    }
+
+    out[0] = (bb * ap - ab * bp) / determinant;
+    out[1] = (aa * bp - ab * ap) / determinant;
+
+    return out;
+}
+
+fn bc6h_with_component(v: vec3<f32>, channel: u32, value: f32) -> vec3<f32> {
+    return vec3<f32>(
+        select(v.x, value, channel == 0u),
+        select(v.y, value, channel == 1u),
+        select(v.z, value, channel == 2u)
+    );
+}
+
+fn bc6h_encode_block_mode11(texels: array<vec3<f32>, 16>) -> vec4<u32> {
+    var bits: array<vec3<f32>, 16>;
+
+    var low = vec3<f32>(65535.0);
+    var high = vec3<f32>(0.0);
+    var mean = vec3<f32>(0.0);
+
+    for (var i = 0u; i < 16u; i++) {
+        let v = vec3<f32>(bc6h_half_bits(texels[i].x), bc6h_half_bits(texels[i].y), bc6h_half_bits(texels[i].z));
+
+        bits[i] = v;
+        low = min(low, v);
+        high = max(high, v);
+        mean += v;
+    }
+
+    mean = mean / 16.0;
+
+    let extent = high - low;
+
+    var e0 = mean;
+    var e1 = mean;
+
+    if (max(extent.x, max(extent.y, extent.z)) >= 0.5) {
+        var cov = array<f32, 6>(0.0, 0.0, 0.0, 0.0, 0.0, 0.0);
+
+        for (var i = 0u; i < 16u; i++) {
+            let d = bits[i] - mean;
+
+            cov[0] += d.x * d.x;
+            cov[1] += d.x * d.y;
+            cov[2] += d.x * d.z;
+            cov[3] += d.y * d.y;
+            cov[4] += d.y * d.z;
+            cov[5] += d.z * d.z;
+        }
+
+        var axis = extent;
+
+        for (var iteration = 0u; iteration < 4u; iteration++) {
+            let next = vec3<f32>(
+                cov[0] * axis.x + cov[1] * axis.y + cov[2] * axis.z,
+                cov[1] * axis.x + cov[3] * axis.y + cov[4] * axis.z,
+                cov[2] * axis.x + cov[4] * axis.y + cov[5] * axis.z
+            );
+
+            let magnitude = length(next);
+
+            if (magnitude < 1e-3) {
+                break;
+            }
+
+            axis = next / magnitude;
+        }
+
+        axis = normalize(axis);
+
+        var t_min = 1e30;
+        var t_max = -1e30;
+
+        for (var i = 0u; i < 16u; i++) {
+            let t = dot(bits[i] - mean, axis);
+
+            t_min = min(t_min, t);
+            t_max = max(t_max, t);
+        }
+
+        e0 = mean + axis * t_min;
+        e1 = mean + axis * t_max;
+
+        for (var round = 0u; round < 2u; round++) {
+            let refined = bc6h_refine(bits, e0, e1);
+
+            e0 = clamp(refined[0], vec3<f32>(0.0), vec3<f32>(31743.0));
+            e1 = clamp(refined[1], vec3<f32>(0.0), vec3<f32>(31743.0));
+        }
+    }
+
+    var q0 = bc6h_quantize(e0);
+    var q1 = bc6h_quantize(e1);
+
+    var best = bc6h_score(bits, q0, q1);
+
+    // coordinate descent over the six stored components, one round
+    for (var component = 0u; component < 6u; component++) {
+        let channel = component % 3u;
+
+        for (var direction = -1; direction <= 1; direction += 2) {
+            var c0 = q0;
+            var c1 = q1;
+
+            if (component < 3u) {
+                c0 = bc6h_with_component(q0, channel, clamp(q0[channel] + f32(direction), 0.0, 1023.0));
+            } else {
+                c1 = bc6h_with_component(q1, channel, clamp(q1[channel] + f32(direction), 0.0, 1023.0));
+            }
+
+            let error = bc6h_score(bits, c0, c1);
+
+            if (error < best) {
+                best = error;
+                q0 = c0;
+                q1 = c1;
+            }
+        }
+    }
+
+    var indices: array<u32, 16>;
+
+    let u0 = bc6h_unquantize(q0);
+    let u1 = bc6h_unquantize(q1);
+
+    for (var i = 0u; i < 16u; i++) {
+        indices[i] = u32(bc6h_nearest(u0, u1, bits[i]).x);
+    }
+
+    // the anchor: texel 0's top bit is implied zero
+    if (indices[0] >= 8u) {
+        let swap = q0;
+
+        q0 = q1;
+        q1 = swap;
+
+        for (var i = 0u; i < 16u; i++) {
+            indices[i] = 15u - indices[i];
+        }
+    }
+
+    return bc6h_pack_mode11(vec3<u32>(q0), vec3<u32>(q1), indices);
+}
+`, [
+        chunk_bc6h_pack_mode11,
+    ]);