Skip to content

8370691: Add new HalffloatVector type and enable intrinsification of float16 vector operations #28002

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

Open
wants to merge 11 commits into
base: master
Choose a base branch
from
+66,081 −12,074
Open

8370691: Add new HalffloatVector type and enable intrinsification of float16 vector operations #28002

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
11 commits
Select commit Hold shift + click to select a range
0a37da2
8370691: Add new HalffloatVector type and enable intrinsification of …
jbhateja Oct 27, 2025
2f43bb1
Build failure fixes
jbhateja Oct 28, 2025
adc1baa
Fix JTREG failures
jbhateja Oct 29, 2025
295b2ca
Adding IR Framework test
jbhateja Oct 30, 2025
911c816
Adding a HalffloatVectorBenchmark having benchmarking kernel parity w…
jbhateja Oct 31, 2025
86a6a29
Removing redundant warmup constraint
jbhateja Nov 3, 2025
68a1f78
Cleanups
jbhateja Nov 3, 2025
55ef3a0
Revamped JTreg test generation and bug fixes
jbhateja Nov 7, 2025
db1c469
Merge branch 'master' of http://github.com/openjdk/jdk into JDK-8370691
jbhateja Nov 7, 2025
1493ef8
Fix some JTREG failures
jbhateja Nov 7, 2025
c60d533
Some cleanups
jbhateja Nov 7, 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
33 changes: 29 additions & 4 deletions src/hotspot/share/classfile/vmIntrinsics.hpp
View file
Open in desktop

Large diffs are not rendered by default.

1 change: 1 addition & 0 deletions src/hotspot/share/classfile/vmSymbols.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -99,6 +99,7 @@ class SerializeClosure;
template(jdk_internal_vm_vector_VectorShuffle, "jdk/internal/vm/vector/VectorSupport$VectorShuffle") \
template(payload_name, "payload") \
template(ETYPE_name, "ETYPE") \
template(CTYPE_name, "CTYPE") \
template(VLENGTH_name, "VLENGTH") \
\
template(jdk_internal_vm_FillerObject, "jdk/internal/vm/FillerObject") \
Expand Down
815 changes: 496 additions & 319 deletions src/hotspot/share/opto/vectorIntrinsics.cpp
View file
Open in desktop

Large diffs are not rendered by default.

46 changes: 40 additions & 6 deletions src/hotspot/share/prims/vectorSupport.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -55,7 +55,7 @@ BasicType VectorSupport::klass2bt(InstanceKlass* ik) {
assert(ik->is_subclass_of(vmClasses::vector_VectorPayload_klass()), "%s not a VectorPayload", ik->name()->as_C_string());
fieldDescriptor fd; // find_field initializes fd if found
// static final Class<?> ETYPE;
Klass* holder = ik->find_field(vmSymbols::ETYPE_name(), vmSymbols::class_signature(), &fd);
Klass* holder = ik->find_field(vmSymbols::CTYPE_name(), vmSymbols::class_signature(), &fd);

assert(holder != nullptr, "sanity");
assert(fd.is_static(), "");
Expand Down Expand Up @@ -199,13 +199,17 @@ bool VectorSupport::is_unsigned_op(jint id) {
}
}

int VectorSupport::vop2ideal(jint id, BasicType bt) {
int VectorSupport::vop2ideal(jint id, BasicType bt, int operType) {
VectorOperation vop = (VectorOperation)id;
switch (vop) {
case VECTOR_OP_ADD: {
switch (bt) {
case T_BYTE: // fall-through
case T_SHORT: // fall-through
case T_SHORT:
if (operType == VECTOR_TYPE_FP16) {
return Op_AddHF;
}
// fall-through
case T_INT: return Op_AddI;
case T_LONG: return Op_AddL;
case T_FLOAT: return Op_AddF;
Expand All @@ -217,7 +221,11 @@ int VectorSupport::vop2ideal(jint id, BasicType bt) {
case VECTOR_OP_SUB: {
switch (bt) {
case T_BYTE: // fall-through
case T_SHORT: // fall-through
case T_SHORT:
if (operType == VECTOR_TYPE_FP16) {
return Op_SubHF;
}
// fall-through
case T_INT: return Op_SubI;
case T_LONG: return Op_SubL;
case T_FLOAT: return Op_SubF;
Expand All @@ -229,7 +237,11 @@ int VectorSupport::vop2ideal(jint id, BasicType bt) {
case VECTOR_OP_MUL: {
switch (bt) {
case T_BYTE: // fall-through
case T_SHORT: // fall-through
case T_SHORT:
if (operType == VECTOR_TYPE_FP16) {
return Op_MulHF;
}
// fall-through
case T_INT: return Op_MulI;
case T_LONG: return Op_MulL;
case T_FLOAT: return Op_MulF;
Expand All @@ -241,7 +253,11 @@ int VectorSupport::vop2ideal(jint id, BasicType bt) {
case VECTOR_OP_DIV: {
switch (bt) {
case T_BYTE: // fall-through
case T_SHORT: // fall-through
case T_SHORT:
if (operType == VECTOR_TYPE_FP16) {
return Op_DivHF;
}
// fall-through
case T_INT: return Op_DivI;
case T_LONG: return Op_DivL;
case T_FLOAT: return Op_DivF;
Expand All @@ -254,6 +270,10 @@ int VectorSupport::vop2ideal(jint id, BasicType bt) {
switch (bt) {
case T_BYTE:
case T_SHORT:
if (operType == VECTOR_TYPE_FP16) {
return Op_MinHF;
}
// fall-through
case T_INT: return Op_MinI;
case T_LONG: return Op_MinL;
case T_FLOAT: return Op_MinF;
Expand All @@ -266,6 +286,10 @@ int VectorSupport::vop2ideal(jint id, BasicType bt) {
switch (bt) {
case T_BYTE:
case T_SHORT:
if (operType == VECTOR_TYPE_FP16) {
return Op_MaxHF;
}
// fall-through
case T_INT: return Op_MaxI;
case T_LONG: return Op_MaxL;
case T_FLOAT: return Op_MaxF;
Expand Down Expand Up @@ -298,6 +322,10 @@ int VectorSupport::vop2ideal(jint id, BasicType bt) {
switch (bt) {
case T_BYTE: // fall-through
case T_SHORT: // fall-through
if (operType == VECTOR_TYPE_FP16) {
return 0;
}
// fall-through
case T_INT: return Op_AbsI;
case T_LONG: return Op_AbsL;
case T_FLOAT: return Op_AbsF;
Expand All @@ -310,6 +338,10 @@ int VectorSupport::vop2ideal(jint id, BasicType bt) {
switch (bt) {
case T_BYTE: // fall-through
case T_SHORT: // fall-through
if (operType == VECTOR_TYPE_FP16) {
return 0;
}
// fall-through
case T_INT: return Op_NegI;
case T_LONG: return Op_NegL;
case T_FLOAT: return Op_NegF;
Expand Down Expand Up @@ -350,6 +382,7 @@ int VectorSupport::vop2ideal(jint id, BasicType bt) {
}
case VECTOR_OP_SQRT: {
switch (bt) {
case T_SHORT: return Op_SqrtHF;
case T_FLOAT: return Op_SqrtF;
case T_DOUBLE: return Op_SqrtD;
default: fatal("SQRT: %s", type2name(bt));
Expand All @@ -358,6 +391,7 @@ int VectorSupport::vop2ideal(jint id, BasicType bt) {
}
case VECTOR_OP_FMA: {
switch (bt) {
case T_SHORT: return Op_FmaHF;
case T_FLOAT: return Op_FmaF;
case T_DOUBLE: return Op_FmaD;
default: fatal("FMA: %s", type2name(bt));
Expand Down
9 changes: 8 additions & 1 deletion src/hotspot/share/prims/vectorSupport.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -137,12 +137,19 @@ class VectorSupport : AllStatic {
NUM_VEC_SIZES = 5
};

enum {
VECTOR_TYPE_PRIM = 1,
VECTOR_TYPE_FP16 = 2,
VECTOR_TYPE_FP8 = 3,
VECTOR_TYPE_INT8 = 4
};

enum {
MODE_BROADCAST = 0,
MODE_BITS_COERCED_LONG_TO_MASK = 1
};

static int vop2ideal(jint vop, BasicType bt);
static int vop2ideal(jint vop, BasicType bt, int operType = VECTOR_TYPE_PRIM);
static bool has_scalar_op(jint id);
static bool is_unsigned_op(jint id);

Expand Down
34 changes: 17 additions & 17 deletions src/java.base/share/classes/jdk/internal/misc/X-ScopedMemoryAccess.java.template
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -326,15 +326,15 @@ public final class ScopedMemoryAccess {
@ForceInline
public static
<V extends VectorSupport.Vector<E>, E, S extends VectorSupport.VectorSpecies<E>>
V loadFromMemorySegment(Class<? extends V> vmClass, Class<E> e, int length,
V loadFromMemorySegment(Class<? extends V> vmClass, Class<?> c, int operType, int length,
AbstractMemorySegmentImpl msp, long offset,
S s,
VectorSupport.LoadOperation<AbstractMemorySegmentImpl, V, S> defaultImpl) {

try {
return loadFromMemorySegmentScopedInternal(
msp.sessionImpl(),
vmClass, e, length,
vmClass, c, operType, length,
msp, offset,
s,
defaultImpl);
Expand All @@ -348,14 +348,14 @@ public final class ScopedMemoryAccess {
private static
<V extends VectorSupport.Vector<E>, E, S extends VectorSupport.VectorSpecies<E>>
V loadFromMemorySegmentScopedInternal(MemorySessionImpl session,
Class<? extends V> vmClass, Class<E> e, int length,
Class<? extends V> vmClass, Class<?> c, int operType, int length,
AbstractMemorySegmentImpl msp, long offset,
S s,
VectorSupport.LoadOperation<AbstractMemorySegmentImpl, V, S> defaultImpl) {
try {
session.checkValidStateRaw();

return VectorSupport.load(vmClass, e, length,
return VectorSupport.load(vmClass, c, operType, length,
msp.unsafeGetBase(), msp.unsafeGetOffset() + offset, true,
msp, offset, s,
defaultImpl);
Expand All @@ -368,14 +368,14 @@ public final class ScopedMemoryAccess {
public static
<V extends VectorSupport.Vector<E>, E, S extends VectorSupport.VectorSpecies<E>,
M extends VectorSupport.VectorMask<E>>
V loadFromMemorySegmentMasked(Class<? extends V> vmClass, Class<M> maskClass, Class<E> e,
V loadFromMemorySegmentMasked(Class<? extends V> vmClass, Class<M> maskClass, Class<?> c, int operType,
int length, AbstractMemorySegmentImpl msp, long offset, M m, S s, int offsetInRange,
VectorSupport.LoadVectorMaskedOperation<AbstractMemorySegmentImpl, V, S, M> defaultImpl) {

try {
return loadFromMemorySegmentMaskedScopedInternal(
msp.sessionImpl(),
vmClass, maskClass, e, length,
vmClass, maskClass, c, operType, length,
msp, offset, m,
s, offsetInRange,
defaultImpl);
Expand All @@ -390,14 +390,14 @@ public final class ScopedMemoryAccess {
<V extends VectorSupport.Vector<E>, E, S extends VectorSupport.VectorSpecies<E>,
M extends VectorSupport.VectorMask<E>>
V loadFromMemorySegmentMaskedScopedInternal(MemorySessionImpl session, Class<? extends V> vmClass,
Class<M> maskClass, Class<E> e, int length,
Class<M> maskClass, Class<?> c, int operType, int length,
AbstractMemorySegmentImpl msp, long offset, M m,
S s, int offsetInRange,
VectorSupport.LoadVectorMaskedOperation<AbstractMemorySegmentImpl, V, S, M> defaultImpl) {
try {
session.checkValidStateRaw();

return VectorSupport.loadMasked(vmClass, maskClass, e, length,
return VectorSupport.loadMasked(vmClass, maskClass, c, operType, length,
msp.unsafeGetBase(), msp.unsafeGetOffset() + offset, true, m, offsetInRange,
msp, offset, s,
defaultImpl);
Expand All @@ -409,15 +409,15 @@ public final class ScopedMemoryAccess {
@ForceInline
public static
<V extends VectorSupport.Vector<E>, E>
void storeIntoMemorySegment(Class<? extends V> vmClass, Class<E> e, int length,
void storeIntoMemorySegment(Class<? extends V> vmClass, Class<?> c, int operType, int length,
V v,
AbstractMemorySegmentImpl msp, long offset,
VectorSupport.StoreVectorOperation<AbstractMemorySegmentImpl, V> defaultImpl) {

try {
storeIntoMemorySegmentScopedInternal(
msp.sessionImpl(),
vmClass, e, length,
vmClass, c, operType, length,
v,
msp, offset,
defaultImpl);
Expand All @@ -431,14 +431,14 @@ public final class ScopedMemoryAccess {
private static
<V extends VectorSupport.Vector<E>, E>
void storeIntoMemorySegmentScopedInternal(MemorySessionImpl session,
Class<? extends V> vmClass, Class<E> e, int length,
Class<? extends V> vmClass, Class<?> c, int operType, int length,
V v,
AbstractMemorySegmentImpl msp, long offset,
VectorSupport.StoreVectorOperation<AbstractMemorySegmentImpl, V> defaultImpl) {
try {
session.checkValidStateRaw();

VectorSupport.store(vmClass, e, length,
VectorSupport.store(vmClass, c, operType, length,
msp.unsafeGetBase(), msp.unsafeGetOffset() + offset, true,
v,
msp, offset,
Expand All @@ -451,15 +451,15 @@ public final class ScopedMemoryAccess {
@ForceInline
public static
<V extends VectorSupport.Vector<E>, E, M extends VectorSupport.VectorMask<E>>
void storeIntoMemorySegmentMasked(Class<? extends V> vmClass, Class<M> maskClass, Class<E> e,
int length, V v, M m,
void storeIntoMemorySegmentMasked(Class<? extends V> vmClass, Class<M> maskClass, Class<?> c,
int operType, int length, V v, M m,
AbstractMemorySegmentImpl msp, long offset,
VectorSupport.StoreVectorMaskedOperation<AbstractMemorySegmentImpl, V, M> defaultImpl) {

try {
storeIntoMemorySegmentMaskedScopedInternal(
msp.sessionImpl(),
vmClass, maskClass, e, length,
vmClass, maskClass, c, operType, length,
v, m,
msp, offset,
defaultImpl);
Expand All @@ -474,13 +474,13 @@ public final class ScopedMemoryAccess {
<V extends VectorSupport.Vector<E>, E, M extends VectorSupport.VectorMask<E>>
void storeIntoMemorySegmentMaskedScopedInternal(MemorySessionImpl session,
Class<? extends V> vmClass, Class<M> maskClass,
Class<E> e, int length, V v, M m,
Class<?> c, int operType, int length, V v, M m,
AbstractMemorySegmentImpl msp, long offset,
VectorSupport.StoreVectorMaskedOperation<AbstractMemorySegmentImpl, V, M> defaultImpl) {
try {
session.checkValidStateRaw();

VectorSupport.storeMasked(vmClass, maskClass, e, length,
VectorSupport.storeMasked(vmClass, maskClass, c, operType, length,
msp.unsafeGetBase(), msp.unsafeGetOffset() + offset, true,
v, m,
msp, offset,
Expand Down
Loading