Skip to content

Add panama implementations of byte-bit and float-bit script operations #124722

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 14 commits into from
Mar 25, 2025
Merged

Add panama implementations of byte-bit and float-bit script operations #124722

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
14 commits
Select commit Hold shift + click to select a range
1493794
Add bit vectors to DistanceFunctionBenchmark
thecoop Mar 12, 2025
987dd5e
Add float-bit panama implementation
thecoop Mar 12, 2025
8c1ac09
Add byte-bit implementation
thecoop Mar 13, 2025
346e294
Unroll loops for more oomph
thecoop Mar 13, 2025
c983834
Check when the unrolled loops should be run
thecoop Mar 13, 2025
646f1b8
Get length calculations correct
thecoop Mar 14, 2025
45b46be
Merge branch 'main' into vector-bit-dot-product
thecoop Mar 24, 2025
d66ee15
Tweak test code
thecoop Mar 24, 2025
34fa081
Merge branch 'main' into vector-bit-dot-product
thecoop Mar 24, 2025
9279e05
More test tweaks
thecoop Mar 24, 2025
6c7ed1a
Don't unroll the filler operations
thecoop Mar 24, 2025
17b29cd
No need for a scalar tail at all
thecoop Mar 24, 2025
f5da55d
Still need a tail for 256-bit impls
thecoop Mar 24, 2025
8762ac1
Update docs/changelog/124722.yaml
thecoop Mar 24, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
Prev Previous commit
Next Next commit
Get length calculations correct
  • Loading branch information
@thecoop
thecoop committed Mar 14, 2025
commit 646f1b8fe92acfbe6fef8e7e153d11fdcb2f5b7c
52 changes: 36 additions & 16 deletions ...in21/java/org/elasticsearch/simdvec/internal/vectorization/PanamaESVectorUtilSupport.java
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -194,17 +194,23 @@ public static long ipByteBin128(byte[] q, byte[] d) {
VectorShape.forBitSize(INT_SPECIES_256.vectorBitSize() / Integer.BYTES)
);

private static int limit(int length, int sectionSize) {
return length - (length % sectionSize);
}

static int ipByteBit512(byte[] q, byte[] d) {
assert q.length == d.length * Byte.SIZE;
int i = 0;
int sum = 0;

if (q.length >= INT_SPECIES_512.length() * 4) {
int sectionLength = INT_SPECIES_512.length() * 4;
if (q.length >= sectionLength) {
IntVector acc0 = IntVector.zero(INT_SPECIES_512);
IntVector acc1 = IntVector.zero(INT_SPECIES_512);
IntVector acc2 = IntVector.zero(INT_SPECIES_512);
IntVector acc3 = IntVector.zero(INT_SPECIES_512);
for (; i < INT_SPECIES_512.loopBound(q.length); i += INT_SPECIES_512.length() * 4) {
int limit = limit(q.length, sectionLength);
for (; i < limit; i += sectionLength) {
var vals0 = ByteVector.fromArray(BYTE_SPECIES_FOR_INT_512, q, i).castShape(INT_SPECIES_512, 0);
var vals1 = ByteVector.fromArray(BYTE_SPECIES_FOR_INT_512, q, i + INT_SPECIES_512.length()).castShape(INT_SPECIES_512, 0);
var vals2 = ByteVector.fromArray(BYTE_SPECIES_FOR_INT_512, q, i + INT_SPECIES_512.length() * 2)
Expand All @@ -227,12 +233,14 @@ static int ipByteBit512(byte[] q, byte[] d) {
+ acc3.reduceLanes(VectorOperators.ADD);
}

if (q.length - i >= INT_SPECIES_256.length() * 4) {
sectionLength = INT_SPECIES_256.length() * 4;
if (q.length - i >= sectionLength) {
IntVector acc0 = IntVector.zero(INT_SPECIES_256);
IntVector acc1 = IntVector.zero(INT_SPECIES_256);
IntVector acc2 = IntVector.zero(INT_SPECIES_256);
IntVector acc3 = IntVector.zero(INT_SPECIES_256);
for (; i < INT_SPECIES_256.loopBound(q.length); i += INT_SPECIES_256.length() * 4) {
int limit = limit(q.length, sectionLength);
for (; i < limit; i += sectionLength) {
var vals0 = ByteVector.fromArray(BYTE_SPECIES_FOR_INT_256, q, i).castShape(INT_SPECIES_256, 0);
var vals1 = ByteVector.fromArray(BYTE_SPECIES_FOR_INT_256, q, i + INT_SPECIES_256.length()).castShape(INT_SPECIES_256, 0);
var vals2 = ByteVector.fromArray(BYTE_SPECIES_FOR_INT_256, q, i + INT_SPECIES_256.length() * 2)
Expand All @@ -257,7 +265,8 @@ static int ipByteBit512(byte[] q, byte[] d) {

if (i < q.length) {
// do the tail
sum += DefaultESVectorUtilSupport.ipByteBitImpl(q, d, i);
// default implementation uses length of data vector, not query vector
sum += DefaultESVectorUtilSupport.ipByteBitImpl(q, d, i / 8);
}
return sum;
}
Expand All @@ -267,12 +276,14 @@ static int ipByteBit256(byte[] q, byte[] d) {
int i = 0;
int sum = 0;

if (q.length >= INT_SPECIES_256.length() * 4) {
int sectionLength = INT_SPECIES_256.length() * 4;
if (q.length >= sectionLength) {
IntVector acc0 = IntVector.zero(INT_SPECIES_256);
IntVector acc1 = IntVector.zero(INT_SPECIES_256);
IntVector acc2 = IntVector.zero(INT_SPECIES_256);
IntVector acc3 = IntVector.zero(INT_SPECIES_256);
for (; i < INT_SPECIES_256.loopBound(q.length); i += INT_SPECIES_256.length() * 4) {
int limit = limit(q.length, sectionLength);
for (; i < limit; i += sectionLength) {
var vals0 = ByteVector.fromArray(BYTE_SPECIES_FOR_INT_256, q, i).castShape(INT_SPECIES_256, 0);
var vals1 = ByteVector.fromArray(BYTE_SPECIES_FOR_INT_256, q, i + INT_SPECIES_256.length()).castShape(INT_SPECIES_256, 0);
var vals2 = ByteVector.fromArray(BYTE_SPECIES_FOR_INT_256, q, i + INT_SPECIES_256.length() * 2)
Expand All @@ -297,7 +308,8 @@ static int ipByteBit256(byte[] q, byte[] d) {

if (i < q.length) {
// do the tail
sum += DefaultESVectorUtilSupport.ipByteBitImpl(q, d, i);
// default implementation uses length of data vector, not query vector
sum += DefaultESVectorUtilSupport.ipByteBitImpl(q, d, i / 8);
}
return sum;
}
Expand All @@ -310,12 +322,14 @@ static float ipFloatBit512(float[] q, byte[] d) {
int i = 0;
float sum = 0;

if (q.length >= FLOAT_SPECIES_512.length() * 4) {
int sectionLength = FLOAT_SPECIES_512.length() * 4;
if (q.length >= sectionLength) {
FloatVector acc0 = FloatVector.zero(FLOAT_SPECIES_512);
FloatVector acc1 = FloatVector.zero(FLOAT_SPECIES_512);
FloatVector acc2 = FloatVector.zero(FLOAT_SPECIES_512);
FloatVector acc3 = FloatVector.zero(FLOAT_SPECIES_512);
for (; i < FLOAT_SPECIES_512.loopBound(q.length); i += FLOAT_SPECIES_512.length() * 4) {
int limit = limit(q.length, sectionLength);
for (; i < limit; i += sectionLength) {
var floats0 = FloatVector.fromArray(FLOAT_SPECIES_512, q, i);
var floats1 = FloatVector.fromArray(FLOAT_SPECIES_512, q, i + FLOAT_SPECIES_512.length());
var floats2 = FloatVector.fromArray(FLOAT_SPECIES_512, q, i + FLOAT_SPECIES_512.length() * 2);
Expand All @@ -336,12 +350,14 @@ static float ipFloatBit512(float[] q, byte[] d) {
+ acc3.reduceLanes(VectorOperators.ADD);
}

if (q.length - i >= FLOAT_SPECIES_256.length() * 4) {
sectionLength = FLOAT_SPECIES_256.length() * 4;
if (q.length - i >= sectionLength) {
FloatVector acc0 = FloatVector.zero(FLOAT_SPECIES_256);
FloatVector acc1 = FloatVector.zero(FLOAT_SPECIES_256);
FloatVector acc2 = FloatVector.zero(FLOAT_SPECIES_256);
FloatVector acc3 = FloatVector.zero(FLOAT_SPECIES_256);
for (; i < FLOAT_SPECIES_256.loopBound(q.length); i += FLOAT_SPECIES_256.length() * 4) {
int limit = limit(q.length, sectionLength);
for (; i < limit; i += sectionLength) {
var floats0 = FloatVector.fromArray(FLOAT_SPECIES_256, q, i);
var floats1 = FloatVector.fromArray(FLOAT_SPECIES_256, q, i + FLOAT_SPECIES_256.length());
var floats2 = FloatVector.fromArray(FLOAT_SPECIES_256, q, i + FLOAT_SPECIES_256.length() * 2);
Expand All @@ -364,7 +380,8 @@ static float ipFloatBit512(float[] q, byte[] d) {

if (i < q.length) {
// do the tail
sum += DefaultESVectorUtilSupport.ipFloatBitImpl(q, d, i);
// default implementation uses length of data vector, not query vector
sum += DefaultESVectorUtilSupport.ipFloatBitImpl(q, d, i / 8);
}

return sum;
Expand All @@ -375,12 +392,14 @@ static float ipFloatBit256(float[] q, byte[] d) {
int i = 0;
float sum = 0;

if (q.length >= FLOAT_SPECIES_256.length() * 4) {
int sectionLength = FLOAT_SPECIES_256.length() * 4;
if (q.length >= sectionLength) {
FloatVector acc0 = FloatVector.zero(FLOAT_SPECIES_256);
FloatVector acc1 = FloatVector.zero(FLOAT_SPECIES_256);
FloatVector acc2 = FloatVector.zero(FLOAT_SPECIES_256);
FloatVector acc3 = FloatVector.zero(FLOAT_SPECIES_256);
for (; i < FLOAT_SPECIES_256.loopBound(q.length); i += FLOAT_SPECIES_256.length() * 4) {
int limit = limit(q.length, sectionLength);
for (; i < limit; i += sectionLength) {
var floats0 = FloatVector.fromArray(FLOAT_SPECIES_256, q, i);
var floats1 = FloatVector.fromArray(FLOAT_SPECIES_256, q, i + FLOAT_SPECIES_256.length());
var floats2 = FloatVector.fromArray(FLOAT_SPECIES_256, q, i + FLOAT_SPECIES_256.length() * 2);
Expand All @@ -403,7 +422,8 @@ static float ipFloatBit256(float[] q, byte[] d) {

if (i < q.length) {
// do the tail
sum += DefaultESVectorUtilSupport.ipFloatBitImpl(q, d, i);
// default implementation uses length of data vector, not query vector
sum += DefaultESVectorUtilSupport.ipFloatBitImpl(q, d, i / 8);
}

return sum;
Expand Down
40 changes: 28 additions & 12 deletions libs/simdvec/src/test/java/org/elasticsearch/simdvec/ESVectorUtilTests.java
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -22,25 +22,41 @@ public class ESVectorUtilTests extends BaseVectorizationTests {
static final ESVectorizationProvider defOrPanamaProvider = BaseVectorizationTests.maybePanamaProvider();

public void testIpByteBit() {
byte[] q = new byte[16];
byte[] d = new byte[] { (byte) Integer.parseInt("01100010", 2), (byte) Integer.parseInt("10100111", 2) };
byte[] d = new byte[random().nextInt(128)];
byte[] q = new byte[d.length * 8];
random().nextBytes(d);
random().nextBytes(q);
int expected = q[1] + q[2] + q[6] + q[8] + q[10] + q[13] + q[14] + q[15];
assertEquals(expected, ESVectorUtil.ipByteBit(q, d));
assertEquals(expected, defaultedProvider.getVectorUtilSupport().ipByteBit(q, d));
assertEquals(expected, defOrPanamaProvider.getVectorUtilSupport().ipByteBit(q, d));

int sum = 0;
for (int i = 0; i < q.length; i++) {
if (((d[i / 8] << (i % 8)) & 0x80) == 0x80) {
sum += q[i];
}
}

assertEquals(sum, ESVectorUtil.ipByteBit(q, d));
assertEquals(sum, defaultedProvider.getVectorUtilSupport().ipByteBit(q, d));
assertEquals(sum, defOrPanamaProvider.getVectorUtilSupport().ipByteBit(q, d));
}

public void testIpFloatBit() {
float[] q = new float[16];
byte[] d = new byte[] { (byte) Integer.parseInt("01100010", 2), (byte) Integer.parseInt("10100111", 2) };
byte[] d = new byte[random().nextInt(128)];
float[] q = new float[d.length * 8];
random().nextBytes(d);

float sum = 0;
for (int i = 0; i < q.length; i++) {
q[i] = random().nextFloat();
if (((d[i / 8] << (i % 8)) & 0x80) == 0x80) {
sum += q[i];
}
}
float expected = q[1] + q[2] + q[6] + q[8] + q[10] + q[13] + q[14] + q[15];
assertEquals(expected, ESVectorUtil.ipFloatBit(q, d), 1e-6);
assertEquals(expected, defaultedProvider.getVectorUtilSupport().ipFloatBit(q, d), 1e-6);
assertEquals(expected, defOrPanamaProvider.getVectorUtilSupport().ipFloatBit(q, d), 1e-6);

double delta = 1e-5 * q.length;

assertEquals(sum, ESVectorUtil.ipFloatBit(q, d), delta);
assertEquals(sum, defaultedProvider.getVectorUtilSupport().ipFloatBit(q, d), delta);
assertEquals(sum, defOrPanamaProvider.getVectorUtilSupport().ipFloatBit(q, d), delta);
}

public void testIpFloatByte() {
Expand Down