Skip to content

Commit

Permalink
improves performance of DXT cluster fit
Browse files Browse the repository at this point in the history 
by about 50%. Disables previous test as implementation is not identical,
any longer.
  • Loading branch information
@memo33
memo33 committed Apr 24, 2016
1 parent 5d6f325 commit 6a95573
Show file tree
Hide file tree
Showing 2 changed files with 60 additions and 18 deletions.
68 changes: 55 additions & 13 deletions src/main/java/com/github/memo33/jsquish/CompressorCluster.java
View file
Original file line number Diff line number Diff line change
Expand Up @@ -29,9 +29,19 @@ a copy of this software and associated documentation files (the
import com.github.memo33.jsquish.Squish.CompressionType;
import com.github.memo33.jsquish.Squish.CompressionMetric;

/* Compared to the original code, we improve the performance by iterating
* only over clusters that are similar to a canonical cluster choice,
* which saves about 50% of total iterations (number of least squares
* problems). This makes subiterations shorter, but on average a bit deeper,
* so we increase MAX_ITERATIONS a bit (this bound is rarely reached anyway).
*
* As the number of iterations is reduced, the error might increase
* slightly (usually less than 0.1%) by running into a different local
* minimum.
*/
final class CompressorCluster extends CompressorColourFit {

private static final int MAX_ITERATIONS = 8;
private static final int MAX_ITERATIONS = 10;

private final int[] indices = new int[16];
private final int[] bestIndices = new int[16];
Expand Down Expand Up @@ -80,9 +90,11 @@ void compress3(final byte[] block, final int offset) {
final Vec b = new Vec();

// prepare an ordering using the principle axis
constructOrdering(principle, 0);
int[] canonical = constructOrderingAndCanonicalCluster(principle, 0, false);

// check all possible clusters and iterate on the total order
// (instead of checking all clusters, we only check those that
// are similar to the canonical one - saves about >50% of iterations)
int bestIteration = 0;
for ( int iteration = 0; ; ) {
// first cluster [0,i) is at the start
Expand All @@ -91,13 +103,15 @@ void compress3(final byte[] block, final int offset) {
alpha[m] = weights[m];
beta[m] = 0.0f;
}
for ( int i = count; i >= 0; --i ) {
for (int x = -canonical[1]; x <= canonical[0]; x++) {
int i = canonical[0] - x;
// second cluster [i,j) is half along
for ( int m = i; m < count; ++m ) {
indices[m] = 2;
alpha[m] = beta[m] = 0.5f * weights[m];
}
for ( int j = count; j >= i; --j ) {
for (int y = -canonical[2]; y <= canonical[1] && y <= canonical[1]+x; y++) {
int j = i + canonical[1] + x - y;
// last cluster [j,k) is at the end
if ( j < count ) {
indices[j] = 1;
Expand Down Expand Up @@ -128,7 +142,8 @@ void compress3(final byte[] block, final int offset) {
break;

// stop if a new iteration is an ordering that has already been tried
if ( !constructOrdering(a.set(bestEnd).sub(bestStart), iteration) )
canonical = constructOrderingAndCanonicalCluster(a.set(bestEnd).sub(bestStart), iteration, false);
if (canonical == null)
break;
}

Expand All @@ -146,6 +161,7 @@ void compress3(final byte[] block, final int offset) {

// save the error
this.totalBestError = bestError;

}
}

Expand All @@ -160,9 +176,11 @@ void compress4(final byte[] block, final int offset) {
final Vec end = new Vec();

// prepare an ordering using the principle axis
constructOrdering(principle, 0);
int[] canonical = constructOrderingAndCanonicalCluster(principle, 0, true);

// check all possible clusters and iterate on the total order
// (instead of checking all clusters, we only check those that
// are similar to the canonical one - saves about >50% of iterations)
int bestIteration = 0;
for ( int iteration = 0; ; ) {
// first cluster [0,i) is at the start
Expand All @@ -171,21 +189,24 @@ void compress4(final byte[] block, final int offset) {
alpha[m] = weights[m];
beta[m] = 0.0f;
}
for ( int i = count; i >= 0; --i ) {
for (int x = -canonical[1]; x <= canonical[0]; x++) {
int i = canonical[0] - x;
// second cluster [i,j) is one third along
for ( int m = i; m < count; ++m ) {
indices[m] = 2;
alpha[m] = (2.0f / 3.0f) * weights[m];
beta[m] = (1.0f / 3.0f) * weights[m];
}
for ( int j = count; j >= i; --j ) {
for (int y = -canonical[2]; y <= canonical[1] && y <= canonical[1]+x; y++) {
int j = i + canonical[1] + x - y;
// third cluster [j,k) is two thirds along
for ( int m = j; m < count; ++m ) {
indices[m] = 3;
alpha[m] = (1.0f / 3.0f) * weights[m];
beta[m] = (2.0f / 3.0f) * weights[m];
}
for ( int k = count; k >= j; --k ) {
for (int z = -canonical[3]; z <= canonical[2] && z <= canonical[2]+y; z++) {
int k = j + canonical[2] + y - z;
// last cluster [k,n) is at the end
if ( k < count ) {
indices[k] = 1;
Expand Down Expand Up @@ -218,7 +239,8 @@ void compress4(final byte[] block, final int offset) {
break;

// stop if a new iteration is an ordering that has already been tried
if ( !constructOrdering(start.set(bestEnd).sub(bestStart), iteration) )
canonical = constructOrderingAndCanonicalCluster(start.set(bestEnd).sub(bestStart), iteration, true);
if (canonical == null)
break;
}

Expand All @@ -235,10 +257,11 @@ void compress4(final byte[] block, final int offset) {

// save the error
this.totalBestError = bestError;

}
}

private boolean constructOrdering(final Vec axis, final int iteration) {
private int[] constructOrderingAndCanonicalCluster(final Vec axis, final int iteration, boolean isComp4) {
// cache some values
final int count = colours.getCount();
final Vec[] values = colours.getPoints();
Expand Down Expand Up @@ -275,7 +298,7 @@ private boolean constructOrdering(final Vec axis, final int iteration) {
}
}
if ( same )
return false;
return null;
}

// copy the ordering and weight all the points
Expand All @@ -301,7 +324,26 @@ private boolean constructOrdering(final Vec axis, final int iteration) {
weighted[j + 1] = wY;
weighted[j + 2] = wZ;
}
return true;
return canonicalCluster(dps, count, isComp4);
}

private static int[] canonicalCluster(float[] dps, int count, boolean isComp4) {
final int[] cluster = new int[isComp4 ? 4 : 3];
if (count == 0) return cluster;
// comp3: comp4:
// |...o...|...*...|...o...| |...o...|...*...|...*...|...o...|
// a 0 1/2 1 b a 0 1/3 2/3 1 b
final float a = dps[0];
final float b = dps[count - 1];
final float[] c = isComp4
? new float[] { (3 * a + b) / 4, (a + b) / 2, (a + 3 * b) / 4, b }
: new float[] { (2 * a + b) / 3, (a + 2 * b) / 3, b };
for (int i = 0, j = 0; i < count; i++) {
while (dps[i] > c[j])
j++;
cluster[j] = cluster[j] + 1;
}
return cluster;
}

private float solveLeastSquares(final Vec start, final Vec end) {
Expand Down
10 changes: 5 additions & 5 deletions src/test/scala/com/github/memo33/jsquish/SquishSpec.scala
View file
Original file line number Diff line number Diff line change
Expand Up @@ -41,11 +41,11 @@ class SquishSpec extends WordSpec with Matchers with PrivateMethodTester {
).result

"Squish reimplementation" should {
"agree with old implementation" in {
for ((comp1, comp2) <- compressors1 zip compressors2; a <- arrs) {
comp1(a) shouldBe comp2(a)
}
}
// "agree with old implementation" in {
// for ((comp1, comp2) <- compressors1 zip compressors2; a <- arrs) {
// comp1(a) shouldBe comp2(a)
// }
// }
"be invariant on already compressed images" in {
val comp = { arr: Array[Byte] =>
Squish.compressImage(arr, w, h, null, Squish.CompressionType.DXT3, Squish.CompressionMethod.CLUSTER_FIT)
Expand Down

0 comments on commit 6a95573

Please sign in to comment.