Backport AVX512 implementation from devel branch to 3.3 (the 3.3 version had many issues)

This commit is contained in:
Gael Guennebaud
2020-08-26 12:34:20 +02:00
parent def70b2e37
commit efd72cddcd
3 changed files with 394 additions and 406 deletions

View File

@@ -248,12 +248,13 @@ template<typename Scalar> void packetmath()
VERIFY(isApproxAbs(ref[0], internal::predux(internal::pload<Packet>(data1)), refvalue) && "internal::predux");
{
for (int i=0; i<4; ++i)
int newsize = PacketSize>4?PacketSize/2:PacketSize;
for (int i=0; i<newsize; ++i)
ref[i] = 0;
for (int i=0; i<PacketSize; ++i)
ref[i%4] += data1[i];
ref[i%newsize] += data1[i];
internal::pstore(data2, internal::predux_downto4(internal::pload<Packet>(data1)));
VERIFY(areApprox(ref, data2, PacketSize>4?PacketSize/2:PacketSize) && "internal::predux_downto4");
VERIFY(areApprox(ref, data2, newsize) && "internal::predux_downto4");
}
ref[0] = 1;
@@ -304,7 +305,7 @@ template<typename Scalar> void packetmath()
}
}
if (PacketTraits::HasBlend || g_vectorize_sse) {
if (PacketTraits::HasBlend) {
// pinsertfirst
for (int i=0; i<PacketSize; ++i)
ref[i] = data1[i];
@@ -314,7 +315,7 @@ template<typename Scalar> void packetmath()
VERIFY(areApprox(ref, data2, PacketSize) && "internal::pinsertfirst");
}
if (PacketTraits::HasBlend || g_vectorize_sse) {
if (PacketTraits::HasBlend) {
// pinsertlast
for (int i=0; i<PacketSize; ++i)
ref[i] = data1[i];