[UNOFFICIAL] Ishiiruka-Dolphin Custom Version

[kirbypuff]

(03-21-2015, 08:21 AM)Tino Wrote: please can you give more detail about the issue?

Normally, Dolphin (master/dev builds) when set to a crazy high IR (12x or more), uses VRAM [displayed in GPU-Z as 'Dedicated Memory'] until it reaches the maximum amount the GPU has on-board. When it exceeds that amount, the OS starts paging that memory to a "reserved" portion of the main system RAM known as 'Dynamic Memory'.

But the latest version of Ishiiruka does something weird: Once the VRAM exceeds the on-board limit, it doesn't "dump" that excess VRAM to the 'Dynamic Memory' pool, but begins to use system RAM (easily seen with the Task Manager) and the usage increases rapidly [even when the user isn't doing anything !] until it consumes all of the real system memory (RAM) and the virtual memory (allocated pagefile). Then you get a hard crash.
With the latest version of Ishiiruka (v310), this happens even at 1x (native) IR (!).

[kirbypuff]

@Tino: This would be an interesting "experiment":

Compile a custom build of Ishiiruka that disables the normal (SSSE3) vertex loader and enables the SSE2 fallback for ALL CPUs.
Then try that build on your PC and compare the performance to the normal Ishiiruka build.
I wonder how efficient that SSE2 fallback really is (on your AMD FX CPU)
Maybe it's even better than the normal vertex loader (who knows?)

[Craftyawesome]

(03-21-2015, 11:48 AM)kirbypuff Wrote:

(03-21-2015, 11:35 AM)Craftyawesome Wrote: 4.0-5901 OpenGL: 32 FPS
Ishiiruka v310 OpenGL: 30 FPS

master 4.0-5901 D3D11: 50~52 FPS
Ishiiruka v310 D3D11: 48~50 FPS

So Ishiiruka (with 'Fast EFB Access' turned off) is still 5%~10% slower than master. Those results perfectly match the ones from the previous benchmark.

@Craftyawesome:
Enable 'Fast EFB Access' in Ishiiruka's Graphics Settings for a nice speed boost.

Skip efb access to CPU was on.

(03-21-2015, 12:41 PM)kirbypuff Wrote:

(03-21-2015, 11:16 AM)Craftyawesome Wrote: Even more accurate than the latest Dev?
Then why is normal dolphn faster?

'more accurate' means more exact / precise / closer to the behavior of the real hardware, not faster.

I thought all possible slowdowns were given toggles.

[Craftyawesome]

(03-21-2015, 11:42 AM)Tino Wrote: Because there are still some optimizations from master that are not merged.

What if you make your changes in a format where it can be merged to prevent this.(not by official developers just so you can add all your changes to the very latest dev with all PRs)

[Archdevil456]

Some questions about the combination of options:

1. When dual core is enabled, does it make sense to check "Synchronize GPU thread" AND set Deterministic dual core to "fake-completion" in order to optimize stability and speed.

2. What is more stable/recommended for games like F-Zero:

Check "Predictive FIFO" + check "Wait for Shader Compilation"

OR

Uncheck both (without enabling "Full Async Shader Compilation")


On a completely different subject: Does anybody know why enabling V-Sync before starting Xenoblade Chronicles causes the emulation to hang at the "reading disc..." screen whereas enabling V-Sync after passing that screen does not cause any problems? Is this a known problem which can/will be fixed in the near future?

[Tino]

(03-21-2015, 09:14 PM)Craftyawesome Wrote:

(03-21-2015, 11:42 AM)Tino Wrote: Because there are still some optimizations from master that are not merged.

What if you make your changes in a format where it can be merged to prevent this.(not by official developers just so you can add all your changes to the very latest dev with all PRs)

Still cant do it in that way because ishiirukas videocommon still has some diferences from master that i need to rework

[Tino]

(03-21-2015, 09:49 PM)Archdevil456 Wrote: Some questions about the combination of options:

1. When dual core is enabled, does it make sense to check "Synchronize GPU thread" AND set Deterministic dual core to "fake-completion" in order to optimize stability and speed.

2. What is more stable/recommended for games like F-Zero:

Check "Predictive FIFO" + check "Wait for Shader Compilation"

OR

Uncheck both (without enabling "Full Async Shader Compilation")


On a completely different subject: Does anybody know why enabling V-Sync before starting Xenoblade Chronicles causes the emulation to hang at the "reading disc..." screen whereas enabling V-Sync after passing that screen does not cause any problems? Is this a known problem which can/will be fixed in the near future?

1 deterministic dual core is still nor implemented
2 predictive fifo + wait for shaders
3 the problem with xenoblade is well known the games use uncaped fps in that scene. So enabling vsync makes the boot takes a really long time

[Tino]

(03-21-2015, 05:31 PM)kirbypuff Wrote: @Tino: This would be an interesting "experiment":

Compile a custom build of Ishiiruka that disables the normal (SSSE3) vertex loader and enables the SSE2 fallback for ALL CPUs.
Then try that build on your PC and compare the performance to the normal Ishiiruka build.
I wonder how efficient that SSE2 fallback really is - on your AMD FX CPU
Maybe it's even better than the normal vertex loader (who knows?)

Already tested in nnewer fx proccesors is 5% slower. In older fx the performance is the same

[Archdevil456]

(03-21-2015, 10:24 PM)Tino Wrote:

(03-21-2015, 09:49 PM)Archdevil456 Wrote: Some questions about the combination of options:

1. When dual core is enabled, does it make sense to check "Synchronize GPU thread" AND set Deterministic dual core to "fake-completion" in order to optimize stability and speed.

2. What is more stable/recommended for games like F-Zero:

Check "Predictive FIFO" + check "Wait for Shader Compilation"

OR

Uncheck both (without enabling "Full Async Shader Compilation")


On a completely different subject: Does anybody know why enabling V-Sync before starting Xenoblade Chronicles causes the emulation to hang at the "reading disc..." screen whereas enabling V-Sync after passing that screen does not cause any problems? Is this a known problem which can/will be fixed in the near future?

1 deterministic dual core is still nor implemented
2 predictive fifo + wait for shaders
3 the problem with xenoblade is well known the games use uncaped fps in that scene. So enabling vsync makes the boot takes a really long time

Thanks, Tino! @1. In the master branch, does the combination of both options make sense?

[Tino]

Master dos not support predictive fifo.