You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@lucene.apache.org by rm...@apache.org on 2024/12/04 20:53:07 UTC
(lucene) branch branch_10x updated: Revert "Ensure Panama float vector distance impls inlinable (#14031)" (#14041)
This is an automated email from the ASF dual-hosted git repository.
rmuir pushed a commit to branch branch_10x
in repository https://gitbox.apache.org/repos/asf/lucene.git
The following commit(s) were added to refs/heads/branch_10x by this push:
new 895177861f3 Revert "Ensure Panama float vector distance impls inlinable (#14031)" (#14041)
895177861f3 is described below
commit 895177861f3492d971de6b4df48fa90e0cb942fb
Author: Robert Muir <rm...@apache.org>
AuthorDate: Wed Dec 4 15:50:18 2024 -0500
Revert "Ensure Panama float vector distance impls inlinable (#14031)" (#14041)
This reverts commit 4f08f3dc6f6b0bc47872e5997168ea50c01bff8d.
---
lucene/CHANGES.txt | 3 -
.../vectorization/PanamaVectorUtilSupport.java | 83 ++++++++++------------
2 files changed, 36 insertions(+), 50 deletions(-)
diff --git a/lucene/CHANGES.txt b/lucene/CHANGES.txt
index 0c8eb903113..bb888113969 100644
--- a/lucene/CHANGES.txt
+++ b/lucene/CHANGES.txt
@@ -103,9 +103,6 @@ Optimizations
* GITHUB#14032: Speed up PostingsEnum when positions are requested.
(Adrien Grand)
-* GITHUB#14031: Ensure Panama float vector distance impls inlinable.
- (Robert Muir, Chris Hegarty)
-
* GITHUB#14011: Reduce allocation rate in HNSW concurrent merge. (Viliam Durina)
Bug Fixes
diff --git a/lucene/core/src/java21/org/apache/lucene/internal/vectorization/PanamaVectorUtilSupport.java b/lucene/core/src/java21/org/apache/lucene/internal/vectorization/PanamaVectorUtilSupport.java
index 18ef76914bb..9273f7c5a81 100644
--- a/lucene/core/src/java21/org/apache/lucene/internal/vectorization/PanamaVectorUtilSupport.java
+++ b/lucene/core/src/java21/org/apache/lucene/internal/vectorization/PanamaVectorUtilSupport.java
@@ -75,9 +75,6 @@ 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
private static FloatVector fma(FloatVector a, FloatVector b, FloatVector c) {
if (Constants.HAS_FAST_VECTOR_FMA) {
@@ -102,7 +99,7 @@ final class PanamaVectorUtilSupport implements VectorUtilSupport {
float res = 0;
// 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);
res += dotProductBody(a, b, i);
}
@@ -123,33 +120,30 @@ final class PanamaVectorUtilSupport implements VectorUtilSupport {
FloatVector acc2 = FloatVector.zero(FLOAT_SPECIES);
FloatVector acc3 = FloatVector.zero(FLOAT_SPECIES);
FloatVector acc4 = FloatVector.zero(FLOAT_SPECIES);
- final int unrolledLimit = limit - 3 * FLOAT_SPECIES_LENGTH;
- for (; i < unrolledLimit; i += 4 * FLOAT_SPECIES_LENGTH) {
+ int unrolledLimit = limit - 3 * FLOAT_SPECIES.length();
+ for (; i < unrolledLimit; i += 4 * FLOAT_SPECIES.length()) {
// one
FloatVector va = FloatVector.fromArray(FLOAT_SPECIES, a, i);
FloatVector vb = FloatVector.fromArray(FLOAT_SPECIES, b, i);
acc1 = fma(va, vb, acc1);
// two
- final int i2 = i + FLOAT_SPECIES_LENGTH;
- FloatVector vc = FloatVector.fromArray(FLOAT_SPECIES, a, i2);
- FloatVector vd = FloatVector.fromArray(FLOAT_SPECIES, b, i2);
+ FloatVector vc = FloatVector.fromArray(FLOAT_SPECIES, a, i + FLOAT_SPECIES.length());
+ FloatVector vd = FloatVector.fromArray(FLOAT_SPECIES, b, i + FLOAT_SPECIES.length());
acc2 = fma(vc, vd, acc2);
// three
- final int i3 = i2 + FLOAT_SPECIES_LENGTH;
- FloatVector ve = FloatVector.fromArray(FLOAT_SPECIES, a, i3);
- FloatVector vf = FloatVector.fromArray(FLOAT_SPECIES, b, i3);
+ FloatVector ve = FloatVector.fromArray(FLOAT_SPECIES, a, i + 2 * FLOAT_SPECIES.length());
+ FloatVector vf = FloatVector.fromArray(FLOAT_SPECIES, b, i + 2 * FLOAT_SPECIES.length());
acc3 = fma(ve, vf, acc3);
// four
- final int i4 = i3 + FLOAT_SPECIES_LENGTH;
- FloatVector vg = FloatVector.fromArray(FLOAT_SPECIES, a, i4);
- FloatVector vh = FloatVector.fromArray(FLOAT_SPECIES, b, i4);
+ FloatVector vg = FloatVector.fromArray(FLOAT_SPECIES, a, i + 3 * FLOAT_SPECIES.length());
+ FloatVector vh = FloatVector.fromArray(FLOAT_SPECIES, b, i + 3 * FLOAT_SPECIES.length());
acc4 = fma(vg, vh, acc4);
}
// 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 vb = FloatVector.fromArray(FLOAT_SPECIES, b, i);
acc1 = fma(va, vb, acc1);
@@ -168,7 +162,7 @@ final class PanamaVectorUtilSupport implements VectorUtilSupport {
float norm2 = 0;
// 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);
float[] ret = cosineBody(a, b, i);
sum += ret[0];
@@ -196,8 +190,8 @@ final class PanamaVectorUtilSupport implements VectorUtilSupport {
FloatVector norm1_2 = FloatVector.zero(FLOAT_SPECIES);
FloatVector norm2_1 = FloatVector.zero(FLOAT_SPECIES);
FloatVector norm2_2 = FloatVector.zero(FLOAT_SPECIES);
- final int unrolledLimit = limit - FLOAT_SPECIES_LENGTH;
- for (; i < unrolledLimit; i += 2 * FLOAT_SPECIES_LENGTH) {
+ int unrolledLimit = limit - FLOAT_SPECIES.length();
+ for (; i < unrolledLimit; i += 2 * FLOAT_SPECIES.length()) {
// one
FloatVector va = FloatVector.fromArray(FLOAT_SPECIES, a, i);
FloatVector vb = FloatVector.fromArray(FLOAT_SPECIES, b, i);
@@ -206,15 +200,14 @@ final class PanamaVectorUtilSupport implements VectorUtilSupport {
norm2_1 = fma(vb, vb, norm2_1);
// two
- final int i2 = i + FLOAT_SPECIES_LENGTH;
- FloatVector vc = FloatVector.fromArray(FLOAT_SPECIES, a, i2);
- FloatVector vd = FloatVector.fromArray(FLOAT_SPECIES, b, i2);
+ FloatVector vc = FloatVector.fromArray(FLOAT_SPECIES, a, i + FLOAT_SPECIES.length());
+ FloatVector vd = FloatVector.fromArray(FLOAT_SPECIES, b, i + FLOAT_SPECIES.length());
sum2 = fma(vc, vd, sum2);
norm1_2 = fma(vc, vc, norm1_2);
norm2_2 = fma(vd, vd, norm2_2);
}
// 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 vb = FloatVector.fromArray(FLOAT_SPECIES, b, i);
sum1 = fma(va, vb, sum1);
@@ -234,7 +227,7 @@ final class PanamaVectorUtilSupport implements VectorUtilSupport {
float res = 0;
// 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);
res += squareDistanceBody(a, b, i);
}
@@ -247,12 +240,6 @@ final class PanamaVectorUtilSupport implements VectorUtilSupport {
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 */
private float squareDistanceBody(float[] a, float[] b, int limit) {
int i = 0;
@@ -262,36 +249,38 @@ final class PanamaVectorUtilSupport implements VectorUtilSupport {
FloatVector acc2 = FloatVector.zero(FLOAT_SPECIES);
FloatVector acc3 = FloatVector.zero(FLOAT_SPECIES);
FloatVector acc4 = FloatVector.zero(FLOAT_SPECIES);
- final int unrolledLimit = limit - 3 * FLOAT_SPECIES_LENGTH;
- for (; i < unrolledLimit; i += 4 * FLOAT_SPECIES_LENGTH) {
+ int unrolledLimit = limit - 3 * FLOAT_SPECIES.length();
+ for (; i < unrolledLimit; i += 4 * FLOAT_SPECIES.length()) {
// one
FloatVector va = FloatVector.fromArray(FLOAT_SPECIES, a, i);
FloatVector vb = FloatVector.fromArray(FLOAT_SPECIES, b, i);
- acc1 = square(va, vb, acc1);
+ FloatVector diff1 = va.sub(vb);
+ acc1 = fma(diff1, diff1, acc1);
// two
- final int i2 = i + FLOAT_SPECIES_LENGTH;
- FloatVector vc = FloatVector.fromArray(FLOAT_SPECIES, a, i2);
- FloatVector vd = FloatVector.fromArray(FLOAT_SPECIES, b, i2);
- acc2 = square(vc, vd, acc2);
+ FloatVector vc = FloatVector.fromArray(FLOAT_SPECIES, a, i + FLOAT_SPECIES.length());
+ FloatVector vd = FloatVector.fromArray(FLOAT_SPECIES, b, i + FLOAT_SPECIES.length());
+ FloatVector diff2 = vc.sub(vd);
+ acc2 = fma(diff2, diff2, acc2);
// three
- final int i3 = i2 + FLOAT_SPECIES_LENGTH;
- FloatVector ve = FloatVector.fromArray(FLOAT_SPECIES, a, i3);
- FloatVector vf = FloatVector.fromArray(FLOAT_SPECIES, b, i3);
- acc3 = square(ve, vf, acc3);
+ FloatVector ve = FloatVector.fromArray(FLOAT_SPECIES, a, i + 2 * FLOAT_SPECIES.length());
+ FloatVector vf = FloatVector.fromArray(FLOAT_SPECIES, b, i + 2 * FLOAT_SPECIES.length());
+ FloatVector diff3 = ve.sub(vf);
+ acc3 = fma(diff3, diff3, acc3);
// four
- final int i4 = i3 + FLOAT_SPECIES_LENGTH;
- FloatVector vg = FloatVector.fromArray(FLOAT_SPECIES, a, i4);
- FloatVector vh = FloatVector.fromArray(FLOAT_SPECIES, b, i4);
- acc4 = square(vg, vh, acc4);
+ FloatVector vg = FloatVector.fromArray(FLOAT_SPECIES, a, i + 3 * FLOAT_SPECIES.length());
+ FloatVector vh = FloatVector.fromArray(FLOAT_SPECIES, b, i + 3 * FLOAT_SPECIES.length());
+ FloatVector diff4 = vg.sub(vh);
+ acc4 = fma(diff4, diff4, acc4);
}
// 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 vb = FloatVector.fromArray(FLOAT_SPECIES, b, i);
- acc1 = square(va, vb, acc1);
+ FloatVector diff = va.sub(vb);
+ acc1 = fma(diff, diff, acc1);
}
// reduce
FloatVector res1 = acc1.add(acc2);