Provide double width 4bpp output to lcd_draw_line

[ccawley2011]

By packing the palette into a single nibble and expanding it when the palette register as is, it's possible to output double-width 4bpp graphics data without much overhead. This is useful on RISC OS for rendering in modes 12 or 27, where the resulting buffer can be copied directly to the screen as-is.

This overlaps with the changes in the glfw_gles2 and sdl2-swrenderer branches, so it should probably be combined with one of them before merging.

[ccawley2011]

This has been made optional, so it should be ready for review.

[deltabeard]

Thank you for this! Have you run any benchmarks before and after these changes on RISC OS using peanut-benchmark to verify the speed improvement? I'll have a thorough look at this pull request when I can. 👍

[ccawley2011]

Thank you for this! Have you run any benchmarks before and after these changes on RISC OS using peanut-benchmark to verify the speed improvement? I'll have a thorough look at this pull request when I can. 👍

I did some tests with this, and enabling PEANUT_GB_USE_DOUBLE_WIDTH_PALETTE plus using memcpy is noticeably faster on a StrongARM RiscPC compared to expanding each byte in lcd_draw_line.

[deltabeard]

enabling PEANUT_GB_USE_DOUBLE_WIDTH_PALETTE plus using memcpy is noticeably faster on a StrongARM RiscPC

I'm trying to understand the usefulness of PEANUT_GB_USE_DOUBLE_WIDTH_PALETTE. If each byte given to lcd_draw_line consists of same two nibbles for a pixel, how are you using memcpy to increase the performance of lcd_draw_line? I think that PEANUT_GB_USE_NIBBLE_FOR_PALETTE is only required for an 8-bit paletted bitmap that only uses 12 colours which can be copied using memcpy as was done in the glfw_gles2 branch. Is PEANUT_GB_USE_DOUBLE_WIDTH_PALETTE supposed to allow for pushing two pixels at a time to a 4-bit paletted texture?

Would you kindly be able to provide example code as to how PEANUT_GB_USE_DOUBLE_WIDTH_PALETTE is supposed to be used please?

[ccawley2011]

Is PEANUT_GB_USE_DOUBLE_WIDTH_PALETTE supposed to allow for pushing two pixels at a time to a 4-bit paletted texture?

That's essentially it - the only difference is that since the RiscPC only has a basic display controller rather than a full GPU, there are limited options for scaling the final image.

I've put my current code on GitHub if you want to take a look. It's a bit primitive at the moment, but it can run games in mode 12 (640x256x4, with aspect ratio correction using older monitors) and mode 27 (640x480x4, with lines doubled in software). I may add support for mode 28 as well (640x480x8) for compatibility with newer devices like the Raspberry Pi.

https://github.com/ccawley2011/Peanut-GB-riscos

[deltabeard]

That's essentially it - the only difference is that since the RiscPC only has a basic display controller rather than a full GPU, there are limited options for scaling the final image.

I see now. So the double width allows for easy scaling the output image to 320x288 by making taking advantage of the fact that the output is a 4-bit palette. It might also then be possible to have the output as a 2-bit palette for scaling to 640x576 (albeit with less colours), but that can be added later if anyone is interested. Without PEANUT_GB_USE_DOUBLE_WIDTH_PALETTE defined, the output is technically an 8-bit pixel, even though only 12 colours are ever used at the moment (this may change when CGB support is added).

I'll do some light testing when I get the time (should be this week) and then I'll merge. 🙂

[deltabeard]

Why was this changed to 4 * 4 instead of 3 * 4 here and in other declarations?

[ccawley2011]

The increased size is a safety precaution to make sure that the table lookup on line 551 always results in a valid read even if there's an invalid combination of flags. I can reduce the size if you think it shouldn't be a problem, though.

[deltabeard]

Just some minor comments then I'll be happy to merge. Sorry for the delay.

[deltabeard]

If PEANUT_GB_USE_NIBBLE_FOR_PALETTE is defined to 0 by the user (e.g. compile time define) but the user enabled PEANUT_GB_USE_DOUBLE_WIDTH_PALETTE, should there be an error here?
The comment should state that PEANUT_GB_USE_NIBBLE_FOR_PALETTE is required if PEANUT_GB_USE_DOUBLE_WIDTH_PALETTE is enabled.

[ccawley2011]

The requirement is mentioned in the description for PEANUT_GB_USE_DOUBLE_WIDTH_PALETTE, and it's checked further down on line 504. I'm not sure on how to word this as part of the PEANUT_GB_USE_NIBBLE_FOR_PALETTE, but I'm open to suggestions if this is what you'd prefer.

[deltabeard]

What is the purpose of changing [3][4] to [3*4]?

[ccawley2011]

It's so that the index doesn't have to be split up when reading from the table on line 184.