aboutsummaryrefslogtreecommitdiff
path: root/arm_compute/core/FixedPoint.inl
diff options
context:
space:
mode:
authorGeorgios Pinitas <georgios.pinitas@arm.com>2017-06-28 18:29:47 +0100
committerAnthony Barbier <anthony.barbier@arm.com>2018-09-17 14:15:39 +0100
commit9247c92bd8c53be4d0c4ae931f51ca8f88e4150b (patch)
tree3d457a263c0aa6ddcf3d05a4a2323640c486aa36 /arm_compute/core/FixedPoint.inl
parent097967568f9363d06df3ac21403edcab57de39d7 (diff)
downloadComputeLibrary-9247c92bd8c53be4d0c4ae931f51ca8f88e4150b.tar.gz
COMPMID-428: Port NESoftmaxLayer to 16-bit fixed point.
Change-Id: I65122950bab9124b9758c27096c0f458b77aeabb Reviewed-on: http://mpd-gerrit.cambridge.arm.com/79365 Reviewed-by: Moritz Pflanzer <moritz.pflanzer@arm.com> Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com> Reviewed-by: Steven Niu <steven.niu@arm.com>
Diffstat (limited to 'arm_compute/core/FixedPoint.inl')
-rw-r--r--arm_compute/core/FixedPoint.inl17
1 files changed, 16 insertions, 1 deletions
diff --git a/arm_compute/core/FixedPoint.inl b/arm_compute/core/FixedPoint.inl
index fdbc3f0c06..b921b32ed9 100644
--- a/arm_compute/core/FixedPoint.inl
+++ b/arm_compute/core/FixedPoint.inl
@@ -90,13 +90,22 @@ inline qint8_t sqadd_qs8(qint8_t a, qint8_t b)
inline qint16_t sqadd_qs16(qint16_t a, qint16_t b)
{
- // We need to store the temporary result in qint16_t otherwise we cannot evaluate the overflow
+ // We need to store the temporary result in qint32_t otherwise we cannot evaluate the overflow
qint32_t tmp = (static_cast<qint32_t>(a) + static_cast<qint32_t>(b));
// Saturate the result in case of overflow and cast to qint16_t
return saturate_convert<qint32_t, qint16_t>(tmp);
}
+inline qint32_t sqadd_qs32(qint32_t a, qint32_t b)
+{
+ // We need to store the temporary result in qint64_t otherwise we cannot evaluate the overflow
+ qint64_t tmp = (static_cast<qint64_t>(a) + static_cast<qint64_t>(b));
+
+ // Saturate the result in case of overflow and cast to qint32_t
+ return saturate_convert<qint64_t, qint32_t>(tmp);
+}
+
inline qint8_t ssub_qs8(qint8_t a, qint8_t b)
{
return a - b;
@@ -388,4 +397,10 @@ inline qint8_t sqmovn_qs16(qint16_t a)
// Saturate the result in case of overflow and cast to qint8_t
return saturate_convert<qint16_t, qint8_t>(a);
}
+
+inline qint16_t sqmovn_qs32(qint32_t a)
+{
+ // Saturate the result in case of overflow and cast to qint16_t
+ return saturate_convert<qint32_t, qint16_t>(a);
+}
}