Metal: Do not reset barrier fence index after first wait.

[squidbus]

Unleashed Recompiled can do the following for resolve passes:

• Issue barriers for source and destination textures for both color and depth at once.
• Set color framebuffer and execute shader color copy draw.
• Set depth framebuffer and execute shader depth copy draw.

The current logic for barriers here will reset the fence index after the placed barrier is waited on by the next encoder. This was a mistaken optimization that does not account for when encoders are split for reasons other than a new barrier. As a result, subsequent encoders are not blocked on the barrier fence like they should be, and can run out-of-order with the operations before the barrier.

In this particular scenario, the copies are separated into different render encoders due to the framebuffer change, it forgets about the last barrier before the second encoder, and the result is the following:

• Wait for fence N and signal fence N + 1 in the color copy
• Perform no fence wait at all and signal fence N + 1 in the depth copy

As a result the depth copy may be reordered to before the barrier, causing e.g. water to flicker in Apotos with anti-aliasing turned off (because this causes both resolves to be shader copy draws).

Fix the issue by removing the lines that reset the fence index to -1 after a wait.