[d3d9+dxso] Rely on robustness for unset SWVP constants

[K0bin]

This implements Joshuas idea to rely on robustness2 for constants that the application hasn't set yet.

SWVP:

• Split the constant buffer into 3 based on the data type.
• Track the highest used constant for each type and create a buffer with the required size. Copy the data.

HWVP/PS:

• Reorder the constant buffer to put the ints in front. There's only 16 Int constants (16 x 4 ints) so we just copy those every time.
• Track the highest used float constant and copy only the required data. Instead of creating differently sized buffers we just bind a smaller range.
• I think it still makes sense for this case because we theoretically need to copy up to 3840 bytes less data (on Nvidia). 256 * 16 total compared to the minimum 256 bytes for size alignment.

When robustness2 is not supported or the alignment does not work out (It does work on Nvidia), we just copy everything like we did before.

Fixes #2264

[K0bin]

TODO:

• fallback to SSBOs when robustness2 isnt supported
• always use UBOs when it is (the huge SWVP buffer has been split up so it's well within the limits)
• make sure we don't go out of bounds anywhere due to robustness2 alignment (256bytes)
• remove swvp game hacks
• store the robustness2 alignment somewhere to reduce pointer chasing
• clean it up
• figure out if it's worth it doing that for regular shaders

EDIT:
Done

[pchome]

Pls do not forget about dxvk.conf

dxvk/dxvk.conf

Lines 442 to 451 in 67391a7

	# SWVP Constant Counts
	#
	# The number of constants available when using Software Vertex Processing.
	#
	# Supported values:
	# - Integers
	# d3d9.swvpFloatCount = 8192
	# d3d9.swvpIntCount = 2048
	# d3d9.swvpBoolCount = 2048

[K0bin]

Nice catch, will remove that soon.

[K0bin]

Thanks for the review @DadSchoorse. Implemented all your feedback.

[K0bin]

I decided to remove the SSBO fallback when robustness2 is not supported. Instead we just copy everything like we did before. Same thing also happens when any of the constant buffers does not match the robustness2 alignment. Luckily, it all works out in practice with the 256 byte alignment on Nvidia hardware.