Putting some more fuel into SMS homebrew

Here I am again... well, almost another year has passed since my last words here. "Doesn't time fly when you're enjoying yourself?"[*]

I spent most of my time these last months writing code and building tools aimed to the homebrew on the SEGA Master System (SMS for short) and SEGA Game Gear.
Writing homebrew games/programs on SMS still means writing ASM code almost from scratch every time.
Yes, of course you can reuse some of the code you've already written, but still there isn't a big deal of shared ready-to-use code, snippets apart. Even if it features background music and sound effects support, the previously only existing audio library (the very good Mod2PSG2, which plays music written on the tracker with the same name) unfortunately never provided a way to export sound effects from its tracker, thus forcing developers to build their sound effects with hex editors.

So I wrote PSGlib, and the tools to convert VGM files into tunes and sound effects that the homebrewer wants to use in their program. VGM files can be produced by some well know trackers such as DefleMask and the very same Mod2PSG2 tracker too. After you convert and compress them into the PSG format, everything you have to do is just to start them (music and SFX) at the right time. Now the majority of SMS homebrew arising these days uses it, the complete list is here.

This happened mostly in 2014, still. And even if writing Z80 ASM code can be very entertaining (seriously!), I decided to try to build a development kit to write SMS homebrew in C.
Of course, first I needed a compiler. SDCC turned out to be the choice. It's a free open source optimizing C compiler that also targets the Zilog Z80 processor, among many others. So the core component I needed was already available.
Since the processor is only one of many components of the SMS, I needed to write code and tools to make it possible to use the SMS as the target, and to write a library on top of that development kit to enable programmers to use the underlying hardware straight from their C programs.

So last January devkitSMS (the kit) and SMSlib (the library) were born.
The library includes functions to handle the display hardware features, such as hardware scrolling for example, background and sprites, and supports software sprite clipping based on a user defined window. It has functions to handle colors and palettes, tiles and tilemaps, both normal keypad and Genesis/MegaDrive 3/6 buttons pad, the pause key and it also has ROM mappers support.
PSGlib then also incarnated as a additional C library, so that it also can be used with the devkit.

And that's pretty much it.

One year with no (posted) news

Hello!

Yes, I'm alive and kicking :)

Sorry I may have appeared MIA... I just had nothing interesting to post; it's been a year since I virtually don't have anything going on involving my lovely DS or the little GBA.

Anyway a few months ago I recovered my younger brother's SEGA Master System II and also his SEGA Game Gear, both of which were forgotten long ago... fortunately, they both are still almost perfectly working.

In case you've never heard of them, they're basically the same hardware, even if the latter is a portable system featuring a color LCD display, more colors and stereo audio. What we're talking about here are 3.5 MHz Zilog Z80 powered 8 bit systems, not exactly something you can compare to GBA 16.5 MHz and DS 66 MHz 32 bit ARM processors horsepower.

Still it's already proving to be very intriguing to write code for these consoles.

First, you have to deal with the processor. The Z80 is an 8 bit CPU, as I said, so actually it can handle only rather small numbers. It even has no instruction to perform multiplication, not to mention division. Shifting a register requires twice the time it takes to make an addition, and you can shift bits left or right by one position only. The fastest operations, such as the addition, require 4 clock cycles (here's the complete instruction set). Many basic operations are available on selected registers only.

The memory. The system features 8 KB of RAM, built into the console. However, the code runs from ROM: there is a chip inside each and every cartridge, so there are also no loading times. ROM size can be up to 1 MB, but everything bigger than 48 KB requires bank switching to access the upper part.

There's virtually no other option but to code in assembler. WLA DX is currently the assembler of choice.

Then there's the hardware responsible for the graphic, the Video Display Processor (VDP), which has very limited capability, again I mean compared to the DS/GBA. Basically here you've got a single background made up of a grid of 32x24 tiles, each with 16 colors either from the first or the second palette, which is the one that is also used by the sprites. The hardware also isn't capable of displaying more than 64 different colors, and I don't mean at the same moment. Finally, up to 64 16-colors sprites (each 8x8 or 8x16 pixels dimensions) are available, but only up to 8 will be drawn on the same scanline. Sprites unfortunately cannot be flipped neither horizontally nor vertically. The VDP still has the quite powerful feature of supporting hardware scrolling of the background in both X and Y directions.

The background tiles, the map, the sprites graphics and the Sprite Attribute Table (SAT) all share the same VRAM space, which is only 16 KB total. The most troublesome issue here is that you can write to VRAM in specific moments only (when video is disabled or during VBlank). If you write to VRAM at the wrong moment, your data will simply be discarded, so it's very easy to end up with corrupted graphics.

Finally, Japanese version of the SEGA Master System apart, the system generates music and sound effects from its PSG chip (the Texas Instruments SN76489 Programmable Sound Generator), which has 4 mono audio channels. Each of the first three channels can output a true square wave (50% duty cycle) of a given frequency, while the fourth channel can output noise only. Volume of each channel can be set to one of 16 attenuation levels on a logarithmic scale.

There are some ready-to-use tools and libraries to compose and replay modules using the PSG chip, but none of these libraries is currently supporting both music and sound effects. So I decided to try implementing a solution to be able to have sound effects over background music, even with the very limited number of available channels.

The result is PSGlib. It plays VGM tunes written for the SN76489 (the tunes need to be converted to a specific format), and it supports sound effects on the third square wave channel and/or on the noise channel, avoiding any collision with the music that would be probably trying to use the same channels. More details about the library may follow in a separate post, eventually.

And... that's pretty much everything so far.

sverx in GBA homebrew land...

(I wrote this post quite a long time ago, during the port of Waimanu Daring Slides to GBA. It has been sitting here for a while, sorry for the delay. You can find the result of my pretty hard work here, on Disjointed Studio blog)

When Nintendo were planning the DS, they decided that the new console would be compatible with the previous one, the Game Boy Advance (GBA for short). So they have put in the same processor - an ARM7TDMI, which is the only processor in the GBA and the 'secondary' processor in the DS when running DS native code. They have also put in an evolution of the same 2D core, giving it many new powerful features. So the GBA, from the point of view of a DS homebrewer, is not completely different, but there are lots of differences that you should keep in mind if you decide to adventure yourself into GBA homebrew land.
So here's an overview of the GBA 2D core features comparing them to the DS... of course, without mentioning things you may easily notice such as that the GBA has got one screen only with a resolution of 240x160 pixels, whereas DS has two 256x192 pixels screens and 2 separate 2D cores. Of course, the GBA has no 3D core at all.
Please note that the list isn't comprehensive, and I'm describing differences pertaining to the graphical 2D core only.

- The DS supports up to 4 backgrounds at the same time. Two of them can only be 'normal' backgrounds (no rotation and scaling is supported on these backgrounds), but you can choose how you want the other two backgrounds to be. So your options are to have all 4 normal backgrounds or you can have 3 normal backgrounds and 1 that supports rotation and scaling  (known as 'rotscale' or 'affine' background) or even have 2 normal and 2 rotscale backgrounds. With GBA, you can have 4 normal backgrounds too, but if you need rotscale backgrounds, you have to give up two normal backgrounds for each rotscale background you want to use. So you'll eventually have 2 normal backgrounds and just one rotscale background or 2 rotscale backgrounds with no other backgrounds at all.
- The DS also features 'extended' rotscale backgrounds, which are rotscale backgrounds supporting up to 1024 different tiles, and each tile can be eventually flipped horizontally/vertically and/or use one of 16 separate 16-color or 256-color palettes. On the GBA there's no such 'extended' rotscale thing, and 'regular' rotscale backgrounds are 256 colors backgrounds that supports up to 256 different tiles only, with no flipping and of course no palette selection.
- Extended rotscale backgrounds on DS can also become bitmap backgrounds, making it possible to have bitmaps over (or under!) text/rotscale backgrounds, or even a bitmap over another. On the other hand, the GBA has very few bitmap oriented features. You can have only a single bitmap background, even if you have 3 choices of what to show in it. You can show a 240x160 15bpp (32 thousand colors) bitmap with a single framebuffer since there's not enough VRAM to have two of them - such a bitmap requires 75 KB. The second choice is a 240x160 256 colors bitmap with double framebuffer, and the last choice is quite a bizarre 160x128 15bpp double framebuffer bitmap background.
- The DS main 2D core (but not the sub core) has a 'large bitmap mode', featuring a single 1024x512 bitmap background. Of course, on the GBA that doesn't exists.
- Palettes: the DS has 16 256-color additional palettes (known as 'extended' palettes) for backgrounds plus another 16 for sprites, besides the regular 256 colors palette for the backgrounds and the regular one for the sprites. Both of these regular palettes also be used as if they were 16 separate 16-color palettes, for 16-color tiles and sprites. The GBA features the very same regular palettes, but there are no 'extended' palettes.
- The GBA has only a total of 96 KB of video RAM, of which 64 KB are dedicated to background maps and tiles, and 32 KB dedicated to sprites. This means, for example, that only 512 different 256-color tiles for sprites can be stored here, even if the GBA 2D core could use up to 1024 different tiles. Also, when choosing a bitmap mode, only 16 KB are left for sprites as the first 80 KB of VRAM are bound to the background bitmap framebuffer(s).
- On the GBA the sprites will always overlap with each other according to their order in the OAM (Object Attribute Memory). This means that sprite number 0 will be always 'on top' of sprite number 1, even if the latter is bearing higher priority than the former. On the contrary, on the DS the priority also works sprite-on-sprite, not simply sprite-on-background.
- Bitmap Objects (also known as 15bpp sprites) don't exist on GBA.

All that said, please don't let this scare you. It's really a lot of fun to code on that little neat machine, and it will surely give you lots of satisfaction.

One hundred twenty-seven shades of grey

Nothing to deal with the bestseller, assured.
As you may already know, the DS 2D core supports paletted and direct colors, both of which are expressed using 5 bits per primary color (red, green, blue). This is 15bpp, also known as HiColor mode.
Having 'only' five bits per primary color means that there are just 32 different shades of gray that can be defined including the darkest - black, and the brightest - white.

Back in 2008, while reading GBATek specifications, I had found out that the DS screens were 18bpp LCD panels, so I started a topic on gbadev forum suggesting that there might be a way to exploit this. It turned out that using the hardware alpha blending capabilities of the 2D core you can indeed force the hardware to show real 18bpp images. Fellow forum member Cydrak has even posted a very good demo then, and here's his original forum post with the link to his 18bpp demo.
So, having now 6 bits per primary color, it's possible to display 64 shades of grey. The improvement is significant, even if banding is still noticeable.

Pushing this thing further has been tickling me since then, so very recently I decided to give it a try. Of course, the hardware limit of 18bpp can't be overcome, and the display can't show more colors than what it's capable of. However, having a 2D core that can generate 60 frames per second, we can exploit the human eye persistence of vision. The idea here is that if we display two slightly different images alternatively at a sufficient high rate (60 times per second is surely high enough), our retinas will just perceive a sort of an average of the two images. And that's what we get: 127 shades of grey. Even if you know that the additional 63 shades aren't really there, they're simply the result of our perception.

Note: the second and third image in this post are fake: there are no emulators that can show 18bpp and, of course, there's no way of making any emulator show the image that your eyes perceive. So, I suggest that you test the demo yourself on your DS. You can download it here.

...more tricks, more speed!

Very recently, I've further optimized the VGA smooth scaling that DSx86 uses, and Pate already announced on his blog the next release will feature the faster code. If you haven't read about the first wave of optimizations, please read about it in my August posts.

This time I exploited two more tricks, and the result is that the speed increased from 79% faster to 114% faster compared to the original code. Of course, this isn't bad at all.

The former optimization comes from the observation that there's no way to use a base register plus a shifted register offset addressing scheme when accessing halfwords, whereas this is very common when accessing words instead. This means we need a separate shift instruction to calculate the offset if we want to access a halfword in a lookup table while, on the contrary, we can access a word in a lookup table using a single instruction. Thus, if we define a 256-words temporary space in the stack (it takes 1 KB) for the lookup table and copy there each palette RGB values as whole words, we will later save one instruction per input pixel when we access them. So this fragment of code:

ldrb r3, [r1], #1  @ read first pixel value
ldrb r4, [r1], #1  @ read second pixel value
ldrb r5, [r1], #1  @ read third pixel value
lsl r3, #1         @ calculate offset (1st pixel)
ldrh r3, [r11, r3] @ read first pixel RGB color
lsl r4, #1         @ calculate offset (2nd pixel)
ldrh r4, [r11, r4] @ read second pixel RGB color
lsl r5, #1         @ calculate offset (3rd pixel)
ldrh r5, [r11, r5] @ read third pixel RGB color

turns into this shorter one:

ldrb r3, [r1], #1  @ read first pixel value
ldrb r4, [r1], #1  @ read second pixel value
ldrb r5, [r1], #1  @ read third pixel value
ldr r3, [r11, r3, lsl #2] @ read first pixel RGB color
ldr r4, [r11, r4, lsl #2] @ read second pixel RGB color
ldr r5, [r11, r5, lsl #2] @ read second pixel RGB color

Since the code performs 5 lookups per loop in total, this optimization saved 5 instructions, shortening the whole loop to 27 instructions only and increasing the speed to 98%.

The latter optimization done uses the very well known trick of the loop unrolling. Since the results are so good, I thought it was worth spending some code space. The loop has been unrolled 8 times, thus now it processes 40 input pixel each iteration before encountering the costly (3 cycles) branch instruction. Even this simple optimization proved to be very effective in terms of performance improvement.

Hardware generated smooth scaling

In my August posts I focused on the improvements done with DSx86's ARM ASM smooth scaling routine where I did my best to make it as fast as possible, knowing that every CPU cycle saved there would turn useful in the emulation main loop. Then it took me a few more months to realize that actually the same result can be achieved by properly programming the NDS 2D graphical core. So here's how I did it.

The smooth scaling routine takes groups of five 256-color pixels on the same line and turns them into four 32K-color pixels on the DS screen by performing many palette lookups and regular/weighted averages, as we've seen already. The DS 2D core, on the other hand, can perform alpha blending between two backgrounds, without requiring any effort from the CPU. This alpha blending feature can achieve nothing less than an average between each pixel of the first background and the corresponding pixel on the second background, returning a 32K-color image.(1) Additionally, the 2D core can also perform background scaling. We need to exploit both these features.

Let's define the 5 original pixels as p0-p4, and the resulting 4 output pixels as r0-r3. What we need to get is:

r0 as the sum of 3/4 p0 and 1/4 p1
r1 as the sum of 1/2 p1 and 1/2 p2
r2 as the sum of 1/4 p2 and 3/4 p3
r3 as p4

If we could blend 4 backgrounds together we could simply copy specific pixels in the 4 backgrounds to obtain this (please check that each output pixel is exactly as expected):

BG0: p0 p1 p2 p4
BG1: p0 p1 p3 p4
BG2: p0 p2 p3 p4
BG3: p1 p2 p3 p4

Since the 2D core can do background scaling, we don't even need to copy specific pixels. Each background can be generated the way we need it starting from the unmodified original image stored in Video RAM using the scaling features. Thus, we program the 2D core to skip one source pixel each group of five, and choose which pixel has to be skipped.
For example, to generate each of the backgrounds (the code does that for BG2), we have to program the background affine matrix to scale a 320-pixel wide image in a 256-pixel wide background:

REG_BG2PA = (320 << 8) / 256;
REG_BG2PB = 0;
REG_BG2PC = 0;
REG_BG2PD = (1 << 8);

Then we should tell the 2D core to skip pixel p1. This is accomplished by using the reference point X coordinate register:

REG_BG2X = (3 << 8) / 4;

You can think of this register as if it was a sort of a counter of the fractional part. We initialize it to a precise value (3/4, in this case) and after each output pixel has been generated, 1/4 gets added to this counter. (It's because 320 divided by 256 gives 1 plus a fractional part of 1/4). When the counter reaches the unit, the scaling process skips one pixel of the original image, and in this case this will happen after processing one pixel. We can also tell the 2D core to use the same 320x200 bitmap for all the backgrounds, then program different reference point X coordinate values for each background.

Unfortunately, what we can't ask the 2D core is to blend all 4 backgrounds at the same time. However, we can make it blend 2 of these backgrounds each frame, and blend the other 2 backgrounds the next frame, at 60 frames per second.(2) The LCD screen and our retinas will average the 2 generated images, providing in fact the expected result.

DSx86 actually uses a slightly different implementation. It performs vertical scaling at the same time (200 lines down to 192 in VGA "Mode 13h" and 240 lines down to 192 in VGA "Mode X", using different affine matrices) in the so-called 'Jitter' mode.

(1) The DS screen output supports 18bpp color, and alpha blending is probably performed with even more precision.
(2) Since only BG2 and BG3 support bitmap backgrounds, the code will blend these two, redefining them as needed on each frame.

a couple speed improving tricks

(This post is a follow-up on my Quick color averaging post. Please read that post first)

Developing a weighted average that uses only 7 ARM assembler instructions instead of using 8 instructions to get the same result was just the tip of the iceberg. To achieve the highest speed possible when resizing a 320 pixel wide image into a 256 pixel wide one, which effectively means converting a PC VGA "Mode 13h" (256 colors) image into a 256 pixel wide 15bpp (32K colors) image on the DS, we should try to speed up every step of the whole conversion process.

For instance, retrieving each pixel's RGB values means reading a byte from the source image that is a pixel in the VGA screen, and accessing the corresponding offset within the palette by performing a lookup table read. So the ARM assembler code for reading the first two RGB values might look like this:

ldrb r3, [r1], #1  @ read first pixel value
lsl r3, #1         @ calculate offset
ldrh r3, [r11, r3] @ read first pixel RGB color
ldrb r4, [r1], #1  @ read second pixel value
lsl r4, #1         @ calculate offset
ldrh r4, [r11, r4] @ read second pixel RGB color

The code above is correct, but it doesn't take into account register interlocking. The ARM946E processor has a 5 stages pipeline, and its loading instructions require that the Memory stage be completed before you can use the target register. This means that there would be a so-called single-cycle load-use interlock if you load a word from memory to a register and you use that register right in the next instruction. In other words, the processor needs to insert a 1-cycle 'pause' before the Execute stage of each of the lsl instructions. Unfortunately, in our code we're reading a single byte from memory instead of a whole word, and things get even worse. Loading a byte (or a halfword) from memory into a register additionally requires the Write stage, thus triggering a two-cycle load-use interlock if the following instruction needs to use the register just loaded, as it happens in our code. (see section 7.12.1 of the ARM9E-S Core Technical Reference, PDF)

Simply reordering the instructions will save us lots of wasted cycles:

ldrb r3, [r1], #1  @ read first pixel value
ldrb r4, [r1], #1  @ read second pixel value
ldrb r5, [r1], #1  @ read third pixel value - we need it later
lsl r3, #1         @ calculate offset (1st pixel)
ldrh r3, [r11, r3] @ read first pixel RGB color
lsl r4, #1         @ calculate offset (2nd pixel)
ldrh r4, [r11, r4] @ read second pixel RGB color

Another thing we have to take into account is that in the DS the color palette is stored in a rather slow memory, and non-sequential accesses to this memory are even slower. According to GBATek, a single 16-bits non-sequential access to palette RAM takes four 33.5 MHz cycles, which translates into eight CPU cycles, because the ARM9 runs at 67 MHz. Palette RAM isn't even cacheable (it's the default setting with DevKitARM; however, I don't suggest that you change this setting even if you actually can) and a lookup is needed for each pixel of the PC VGA "Mode 13h" screen. With a resolution of 320x200 pixel, this happens 64000 times per frame.

To speed up all those lookups, we can copy the palette into a faster memory right before starting our conversion routine. DTCM (Data Tightly-Coupled Memory) is just the right choice. It's a very fast memory: it has single-cycle access time even with non-sequential accesses, but it isn't very large being in fact only 16 KB total. The program's stack resides on it (again, it's a DevKitARM default setting, and once more I don't recommend changing it) but we actually need only 512 bytes to copy the 256 halfwords. So we temporarely allocate that half kilobyte on top of the stack and copy the palette there. Then we will perform all our lookups being sure there will be no slowdown. Actually, this has surely been the most effective change applied to the code in terms of performance improvement.

The last code optimization uses a peculiar kind of SIMD. ARM9 isn't a SIMD CPU, so it can't process multiple data with a single instruction unlike most processors in use nowadays. However, since we have 32 bits registers in there and we need to process 16 bits operands, we could stuff two operands per register and process double-operands as if they were normal operands. Of course, we have to be sure that we don't mix them up. This 'trick' is called SWAR - SIMD Within A Register.

Since in our code we have to perform two weighted averages for each stripe of 5 pixels that we want to convert into 4, we can actually perform the two weighted averages at the same time. Obviously, there's a little overhead: we need to move the operands together before performing the operations and separate them afterwards. This requires 4 ARM assembler instructions. So we can perform two weighted averages in just 11 instructions.

The resulting code, after all these changes, turned out to be 79% faster. Now it processes 179 pixels in the same time that it took the previous code to process only 100 pixels.

In the next post I'll tell you how to obtain the same graphical output without virtually using any CPU resources.