Re-optimized unswizzle

[shinjich]

modified code that optimized for Xenon architecture, and vectorized unity table generators.

[walbourn]

Make sure this still runs with the associated xdsp test here.

The CMakeLists.txt in that project should also allow you to verify it builds without warnings with clang/LLVM.

[shinjich]

The updated XDSP adopted XMVectorSinCos for vectorization. I noticed that compiling with clang LLVM would cause an accuracy error of about 1 bit and it would fail compareson on XDSPTest.

define TESTEPSILON 0.00001f -> 0.0000124f can absorb accuracy errors, but if we need to resolve the accuracy error itself, we will need to use standard library's sinf/cosf instead of XMVectorSinCos.

Note that this issue does not occurs with VC compiler.

[walbourn]

The updated XDSP adopted XMVectorSinCos for vectorization. I noticed that compiling with clang LLVM would cause an accuracy error of about 1 bit and it would fail compareson on XDSPTest.

define TESTEPSILON 0.00001f -> 0.0000124f can absorb accuracy errors, but if we need to resolve the accuracy error itself, we will need to use standard library's sinf/cosf instead of XMVectorSinCos.

Note that this issue does not occurs with VC compiler.

Feel free to submit a PR for the test as well. It's pretty common to need to adjust epsilon between implementations. You should see what I hit originally moving from Intel to ARM :)

[shinjich]

Thanks Chuck, I'll submit a PR for the test later. As you pointed out, precision will be different between NEON and SSE.