The isometric transform, like the affine transform, has an implicit last
row of [0, 0, 0, 1]. This was not being properly initialized, as verified
by a new test function.
This makes both the small and huge argument cases faster because:
- for small inputs this removes the last pselect
- for large inputs only the reduction part follows a scalar path,
the rest use the same SIMD path as the small-argument case.
The reinterpret_casts used in ptranspose(PacketBlock<Packet8cf,4>&)
ptranspose(PacketBlock<Packet8cf,8>&) don't appear to be working
correctly. They're used to convert the kernel parameters to
PacketBlock<Packet8d,T>& so that the complex number versions of
ptranspose can be written using the existing double implementations.
Unfortunately, they don't seem to work and are responsible for 9 unit
test failures in the AVX512 build of tensorflow master. This commit
fixes the issue by manually initialising PacketBlock<Packet8d,T>
variables with the contents of the kernel parameter before calling
the double version of ptranspose, and then copying the resulting
values back into the kernel parameter before returning.
- no FMA: 1ULP up to 3pi, 2ULP up to sin(25966) and cos(18838), fallback to std::sin/cos for larger inputs
- FMA: 1ULP up to sin(117435.992) and cos(71476.0625), fallback to std::sin/cos for larger inputs
Commit c53eececb0
introduced AVX512 support for complex numbers but required
avx512dq to build. Commit 1d683ae2f5
fixed some but not, it would seem all,
of the hard avx512dq dependencies. Build failures are still evident on
Eigen and TensorFlow when compiling with just avx512f and no avx512dq
using gcc 7.3. Looking at the code there does indeed seem to be a problem.
Commit c53eececb0
calls avx512dq intrinsics directly, e.g, _mm512_extractf32x8_ps
and _mm512_and_ps. This commit fixes the issue by replacing the direct
intrinsic calls with the various wrapper functions that are safe to use on
avx512f only builds.
