Apache
/
lucene
mirror of https://github.com/apache/lucene.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

[10.x] Ensure Panama float vector distance impls inlinable (#14031) 

This commit reduces the Panama vector distance float implementations to less than the maximum bytecode size of a hot method to be inlined (325).

E.g. Previously:  org.apache.lucene.internal.vectorization.PanamaVectorUtilSupport::dotProductBody (355 bytes)   failed to inline: callee is too large.

After: org.apache.lucene.internal.vectorization.PanamaVectorUtilSupport::dotProductBody (3xx bytes)   inline (hot)

This helps things a little.

Co-authored-by: Robert Muir <rmuir@apache.org>
This commit is contained in:
Chris Hegarty 2024-12-03 10:49:33 +00:00 committed by ChrisHegarty
parent 8762de7f11
commit 290847a80e
2 changed files with 50 additions and 36 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
lucene/CHANGES.txt
View File

@ -96,6 +96,9 @@ Optimizations
* GITHUB#14032: Speed up PostingsEnum when positions are requested. * GITHUB#14032: Speed up PostingsEnum when positions are requested.
(Adrien Grand) (Adrien Grand)
* GITHUB#14031: Ensure Panama float vector distance impls inlinable.
(Robert Muir, Chris Hegarty)
Bug Fixes Bug Fixes
--------------------- ---------------------
* GITHUB#13832: Fixed an issue where the DefaultPassageFormatter.format method did not format passages as intended * GITHUB#13832: Fixed an issue where the DefaultPassageFormatter.format method did not format passages as intended

83
lucene/core/src/java21/org/apache/lucene/internal/vectorization/PanamaVectorUtilSupport.java
View File

@ -75,6 +75,9 @@ final class PanamaVectorUtilSupport implements VectorUtilSupport {
} }
} }
// cached vector sizes for smaller method bodies
private static final int FLOAT_SPECIES_LENGTH = FLOAT_SPECIES.length();
// the way FMA should work! if available use it, otherwise fall back to mul/add // the way FMA should work! if available use it, otherwise fall back to mul/add
private static FloatVector fma(FloatVector a, FloatVector b, FloatVector c) { private static FloatVector fma(FloatVector a, FloatVector b, FloatVector c) {
if (Constants.HAS_FAST_VECTOR_FMA) { if (Constants.HAS_FAST_VECTOR_FMA) {
@ -99,7 +102,7 @@ final class PanamaVectorUtilSupport implements VectorUtilSupport {
float res = 0; float res = 0;
// if the array size is large (> 2x platform vector size), its worth the overhead to vectorize // if the array size is large (> 2x platform vector size), its worth the overhead to vectorize
if (a.length > 2 * FLOAT_SPECIES.length()) { if (a.length > 2 * FLOAT_SPECIES_LENGTH) {
i += FLOAT_SPECIES.loopBound(a.length); i += FLOAT_SPECIES.loopBound(a.length);
res += dotProductBody(a, b, i); res += dotProductBody(a, b, i);
} }
@ -120,30 +123,33 @@ final class PanamaVectorUtilSupport implements VectorUtilSupport {
FloatVector acc2 = FloatVector.zero(FLOAT_SPECIES); FloatVector acc2 = FloatVector.zero(FLOAT_SPECIES);
FloatVector acc3 = FloatVector.zero(FLOAT_SPECIES); FloatVector acc3 = FloatVector.zero(FLOAT_SPECIES);
FloatVector acc4 = FloatVector.zero(FLOAT_SPECIES); FloatVector acc4 = FloatVector.zero(FLOAT_SPECIES);
int unrolledLimit = limit - 3 * FLOAT_SPECIES.length(); final int unrolledLimit = limit - 3 * FLOAT_SPECIES_LENGTH;
for (; i < unrolledLimit; i += 4 * FLOAT_SPECIES.length()) { for (; i < unrolledLimit; i += 4 * FLOAT_SPECIES_LENGTH) {
// one // one
FloatVector va = FloatVector.fromArray(FLOAT_SPECIES, a, i); FloatVector va = FloatVector.fromArray(FLOAT_SPECIES, a, i);
FloatVector vb = FloatVector.fromArray(FLOAT_SPECIES, b, i); FloatVector vb = FloatVector.fromArray(FLOAT_SPECIES, b, i);
acc1 = fma(va, vb, acc1); acc1 = fma(va, vb, acc1);
// two // two
FloatVector vc = FloatVector.fromArray(FLOAT_SPECIES, a, i + FLOAT_SPECIES.length()); final int i2 = i + FLOAT_SPECIES_LENGTH;
FloatVector vd = FloatVector.fromArray(FLOAT_SPECIES, b, i + FLOAT_SPECIES.length()); FloatVector vc = FloatVector.fromArray(FLOAT_SPECIES, a, i2);
FloatVector vd = FloatVector.fromArray(FLOAT_SPECIES, b, i2);
acc2 = fma(vc, vd, acc2); acc2 = fma(vc, vd, acc2);
// three // three
FloatVector ve = FloatVector.fromArray(FLOAT_SPECIES, a, i + 2 * FLOAT_SPECIES.length()); final int i3 = i2 + FLOAT_SPECIES_LENGTH;
FloatVector vf = FloatVector.fromArray(FLOAT_SPECIES, b, i + 2 * FLOAT_SPECIES.length()); FloatVector ve = FloatVector.fromArray(FLOAT_SPECIES, a, i3);
FloatVector vf = FloatVector.fromArray(FLOAT_SPECIES, b, i3);
acc3 = fma(ve, vf, acc3); acc3 = fma(ve, vf, acc3);
// four // four
FloatVector vg = FloatVector.fromArray(FLOAT_SPECIES, a, i + 3 * FLOAT_SPECIES.length()); final int i4 = i3 + FLOAT_SPECIES_LENGTH;
FloatVector vh = FloatVector.fromArray(FLOAT_SPECIES, b, i + 3 * FLOAT_SPECIES.length()); FloatVector vg = FloatVector.fromArray(FLOAT_SPECIES, a, i4);
FloatVector vh = FloatVector.fromArray(FLOAT_SPECIES, b, i4);
acc4 = fma(vg, vh, acc4); acc4 = fma(vg, vh, acc4);
} }
// vector tail: less scalar computations for unaligned sizes, esp with big vector sizes // vector tail: less scalar computations for unaligned sizes, esp with big vector sizes
for (; i < limit; i += FLOAT_SPECIES.length()) { for (; i < limit; i += FLOAT_SPECIES_LENGTH) {
FloatVector va = FloatVector.fromArray(FLOAT_SPECIES, a, i); FloatVector va = FloatVector.fromArray(FLOAT_SPECIES, a, i);
FloatVector vb = FloatVector.fromArray(FLOAT_SPECIES, b, i); FloatVector vb = FloatVector.fromArray(FLOAT_SPECIES, b, i);
acc1 = fma(va, vb, acc1); acc1 = fma(va, vb, acc1);
@ -162,7 +168,7 @@ final class PanamaVectorUtilSupport implements VectorUtilSupport {
float norm2 = 0; float norm2 = 0;
// if the array size is large (> 2x platform vector size), its worth the overhead to vectorize // if the array size is large (> 2x platform vector size), its worth the overhead to vectorize
if (a.length > 2 * FLOAT_SPECIES.length()) { if (a.length > 2 * FLOAT_SPECIES_LENGTH) {
i += FLOAT_SPECIES.loopBound(a.length); i += FLOAT_SPECIES.loopBound(a.length);
float[] ret = cosineBody(a, b, i); float[] ret = cosineBody(a, b, i);
sum += ret[0]; sum += ret[0];
@ -190,8 +196,8 @@ final class PanamaVectorUtilSupport implements VectorUtilSupport {
FloatVector norm1_2 = FloatVector.zero(FLOAT_SPECIES); FloatVector norm1_2 = FloatVector.zero(FLOAT_SPECIES);
FloatVector norm2_1 = FloatVector.zero(FLOAT_SPECIES); FloatVector norm2_1 = FloatVector.zero(FLOAT_SPECIES);
FloatVector norm2_2 = FloatVector.zero(FLOAT_SPECIES); FloatVector norm2_2 = FloatVector.zero(FLOAT_SPECIES);
int unrolledLimit = limit - FLOAT_SPECIES.length(); final int unrolledLimit = limit - FLOAT_SPECIES_LENGTH;
for (; i < unrolledLimit; i += 2 * FLOAT_SPECIES.length()) { for (; i < unrolledLimit; i += 2 * FLOAT_SPECIES_LENGTH) {
// one // one
FloatVector va = FloatVector.fromArray(FLOAT_SPECIES, a, i); FloatVector va = FloatVector.fromArray(FLOAT_SPECIES, a, i);
FloatVector vb = FloatVector.fromArray(FLOAT_SPECIES, b, i); FloatVector vb = FloatVector.fromArray(FLOAT_SPECIES, b, i);
@ -200,14 +206,15 @@ final class PanamaVectorUtilSupport implements VectorUtilSupport {
norm2_1 = fma(vb, vb, norm2_1); norm2_1 = fma(vb, vb, norm2_1);
// two // two
FloatVector vc = FloatVector.fromArray(FLOAT_SPECIES, a, i + FLOAT_SPECIES.length()); final int i2 = i + FLOAT_SPECIES_LENGTH;
FloatVector vd = FloatVector.fromArray(FLOAT_SPECIES, b, i + FLOAT_SPECIES.length()); FloatVector vc = FloatVector.fromArray(FLOAT_SPECIES, a, i2);
FloatVector vd = FloatVector.fromArray(FLOAT_SPECIES, b, i2);
sum2 = fma(vc, vd, sum2); sum2 = fma(vc, vd, sum2);
norm1_2 = fma(vc, vc, norm1_2); norm1_2 = fma(vc, vc, norm1_2);
norm2_2 = fma(vd, vd, norm2_2); norm2_2 = fma(vd, vd, norm2_2);
} }
// vector tail: less scalar computations for unaligned sizes, esp with big vector sizes // vector tail: less scalar computations for unaligned sizes, esp with big vector sizes
for (; i < limit; i += FLOAT_SPECIES.length()) { for (; i < limit; i += FLOAT_SPECIES_LENGTH) {
FloatVector va = FloatVector.fromArray(FLOAT_SPECIES, a, i); FloatVector va = FloatVector.fromArray(FLOAT_SPECIES, a, i);
FloatVector vb = FloatVector.fromArray(FLOAT_SPECIES, b, i); FloatVector vb = FloatVector.fromArray(FLOAT_SPECIES, b, i);
sum1 = fma(va, vb, sum1); sum1 = fma(va, vb, sum1);
@ -227,7 +234,7 @@ final class PanamaVectorUtilSupport implements VectorUtilSupport {
float res = 0; float res = 0;
// if the array size is large (> 2x platform vector size), its worth the overhead to vectorize // if the array size is large (> 2x platform vector size), its worth the overhead to vectorize
if (a.length > 2 * FLOAT_SPECIES.length()) { if (a.length > 2 * FLOAT_SPECIES_LENGTH) {
i += FLOAT_SPECIES.loopBound(a.length); i += FLOAT_SPECIES.loopBound(a.length);
res += squareDistanceBody(a, b, i); res += squareDistanceBody(a, b, i);
} }
@ -240,6 +247,12 @@ final class PanamaVectorUtilSupport implements VectorUtilSupport {
return res; return res;
} }
/** helper: returns fma(a.sub(b), a.sub(b), c) */
private static FloatVector square(FloatVector a, FloatVector b, FloatVector c) {
FloatVector diff = a.sub(b);
return fma(diff, diff, c);
}
/** vectorized square distance body */ /** vectorized square distance body */
private float squareDistanceBody(float[] a, float[] b, int limit) { private float squareDistanceBody(float[] a, float[] b, int limit) {
int i = 0; int i = 0;
@ -249,38 +262,36 @@ final class PanamaVectorUtilSupport implements VectorUtilSupport {
FloatVector acc2 = FloatVector.zero(FLOAT_SPECIES); FloatVector acc2 = FloatVector.zero(FLOAT_SPECIES);
FloatVector acc3 = FloatVector.zero(FLOAT_SPECIES); FloatVector acc3 = FloatVector.zero(FLOAT_SPECIES);
FloatVector acc4 = FloatVector.zero(FLOAT_SPECIES); FloatVector acc4 = FloatVector.zero(FLOAT_SPECIES);
int unrolledLimit = limit - 3 * FLOAT_SPECIES.length(); final int unrolledLimit = limit - 3 * FLOAT_SPECIES_LENGTH;
for (; i < unrolledLimit; i += 4 * FLOAT_SPECIES.length()) { for (; i < unrolledLimit; i += 4 * FLOAT_SPECIES_LENGTH) {
// one // one
FloatVector va = FloatVector.fromArray(FLOAT_SPECIES, a, i); FloatVector va = FloatVector.fromArray(FLOAT_SPECIES, a, i);
FloatVector vb = FloatVector.fromArray(FLOAT_SPECIES, b, i); FloatVector vb = FloatVector.fromArray(FLOAT_SPECIES, b, i);
FloatVector diff1 = va.sub(vb); acc1 = square(va, vb, acc1);
acc1 = fma(diff1, diff1, acc1);
// two // two
FloatVector vc = FloatVector.fromArray(FLOAT_SPECIES, a, i + FLOAT_SPECIES.length()); final int i2 = i + FLOAT_SPECIES_LENGTH;
FloatVector vd = FloatVector.fromArray(FLOAT_SPECIES, b, i + FLOAT_SPECIES.length()); FloatVector vc = FloatVector.fromArray(FLOAT_SPECIES, a, i2);
FloatVector diff2 = vc.sub(vd); FloatVector vd = FloatVector.fromArray(FLOAT_SPECIES, b, i2);
acc2 = fma(diff2, diff2, acc2); acc2 = square(vc, vd, acc2);
// three // three
FloatVector ve = FloatVector.fromArray(FLOAT_SPECIES, a, i + 2 * FLOAT_SPECIES.length()); final int i3 = i2 + FLOAT_SPECIES_LENGTH;
FloatVector vf = FloatVector.fromArray(FLOAT_SPECIES, b, i + 2 * FLOAT_SPECIES.length()); FloatVector ve = FloatVector.fromArray(FLOAT_SPECIES, a, i3);
FloatVector diff3 = ve.sub(vf); FloatVector vf = FloatVector.fromArray(FLOAT_SPECIES, b, i3);
acc3 = fma(diff3, diff3, acc3); acc3 = square(ve, vf, acc3);
// four // four
FloatVector vg = FloatVector.fromArray(FLOAT_SPECIES, a, i + 3 * FLOAT_SPECIES.length()); final int i4 = i3 + FLOAT_SPECIES_LENGTH;
FloatVector vh = FloatVector.fromArray(FLOAT_SPECIES, b, i + 3 * FLOAT_SPECIES.length()); FloatVector vg = FloatVector.fromArray(FLOAT_SPECIES, a, i4);
FloatVector diff4 = vg.sub(vh); FloatVector vh = FloatVector.fromArray(FLOAT_SPECIES, b, i4);
acc4 = fma(diff4, diff4, acc4); acc4 = square(vg, vh, acc4);
} }
// vector tail: less scalar computations for unaligned sizes, esp with big vector sizes // vector tail: less scalar computations for unaligned sizes, esp with big vector sizes
for (; i < limit; i += FLOAT_SPECIES.length()) { for (; i < limit; i += FLOAT_SPECIES_LENGTH) {
FloatVector va = FloatVector.fromArray(FLOAT_SPECIES, a, i); FloatVector va = FloatVector.fromArray(FLOAT_SPECIES, a, i);
FloatVector vb = FloatVector.fromArray(FLOAT_SPECIES, b, i); FloatVector vb = FloatVector.fromArray(FLOAT_SPECIES, b, i);
FloatVector diff = va.sub(vb); acc1 = square(va, vb, acc1);
acc1 = fma(diff, diff, acc1);
} }
// reduce // reduce
FloatVector res1 = acc1.add(acc2); FloatVector res1 = acc1.add(acc2);