cellAdec implementation part 2: LPCM decoder

[capriots]

• Implements the LPCM decoder part of cellAdec
• The decoder takes 16/20/24-bit big-endian integer samples using either the Blu-ray (also used in PAMF files) or DVD-Video packing format (without the extra header you would usually find in those streams) and converts them to float samples
• Fully vectorized
• Produces identical output to LLE for valid parameters (like sample_num % channel_num == 0 , access_unit_size % sample_size == 0)
  • For invalid parameters, only the end of the output, where the samples are cut off, is different, depending on the memory contents
  • There is basically no checking of user inputs on LLE
• Set cellAdec to HLE by default since it is fully implemented with this

[elad335]

You can use a template here, and do std::bit_cast<v128> for all masked_array_t

[capriots]

Bit casting to union types isn't constexpr. But changing the existing constructor to this would work:

template <Vector128 T>
constexpr v128(const T& rhs) noexcept
	: _u(std::bit_cast<u128>(rhs))
{
}

However, it would technically be UB to access the vector through any member other than _u afterwards.
Or do you mean something different?

[elad335]

I rather it not be constexpr than, for constexpr stuff you can do the trick with masked_array_t as you have done earlier.

[elad335]

There is no significant performance difference between them in modern CPUs, hence why MSVC doed not emit it. Nor should we as the risks outweigh the advantage.

Suggested change

	return _mm_load_ps(static_cast<const f32*>(ptr));
	return _mm_loadu_ps(static_cast<const f32*>(ptr));

[elad335]

Suggested change

	return _mm_load_si128(static_cast<const __m128i*>(ptr));
	return _mm_loadu_si128(static_cast<const __m128i*>(ptr));