Both of these issues can be fixed without switching away from forwards texture mapping.
We have at least one examples of this being implemented in real hardware at around the same time: Nvidia's NV1 (their first attempt at a GPU) was a quad based GPU that used forwards texture mapping. And it had both UV mapping and translucent polygons. The sega saturn titles that were ported to the NV1 look absolutely great. And I the Sega Model 3 also implemented both UV mapping and transparency on quads (I'm just not 100% sure it was using forwards texture mapping).
I'm not suggesting the Saturn should have implemented the full perspective correct quadratic texture mapping that the NV1 had. That would be overkill graphically (especially since the ps1 didn't have perspective correct texture mapping either), and probably would have taken up too many transistors.
But even a simple scheme that ignored perspective correctness were the artist provided UV coord for three of the vertices (and the forth was implicitly derived by assuming the quad and it's UVs were flat) would have provided most of the texture mapping capabilities that artists needed. And when using malformed quads to emulate triangles, the 3 UV coords would match.
And most of the hardware is already there. VDP1 already has hardware to calculate the slopes all four quad edges from vertex coords and an edge walker to walk though all pixels in screen space. It also has similar hardware for calculating the texture space height/width deltas and walking though texture space. It's just that texture space is assumed to always have slopes of zero.
So that's the only changes we make. VDP1 now calculates proper texture space slopes for each quad, and the texture coord walker is updated to take non-zero slopes. I'm hopeful such an improvement could have been made without massively increasing the transistor count of VDP1, and suddenly the saturn's texture mapping capabilities are roughly equal with the PS1.
> I don't consider this to be a significant problem. The checkerboard mesh feature provides a workable pseudo-half-transparent effect
I think you are underestimating just how useful proper transparency is for early 3D hardware.
(by proper transparency, I mean supporting the add/subtract blend modes)
And I haven't put it on my list of "important missing features" because sometimes games want to render transparent objects. As you said, the mesh (checkerboard) mode actually works for a lot of those cases. And if you are doing particles, then you can often get away with using non-distorted quads, which does work with the half-transparent mode. (though, from an artistic perspective, particles would really look a lot better if they could instead use add/subtract modes)
No. Proper transparency is on my list because it's an essential building block for emulating advanced features that aren't natively supported on the hardware.
One great example is the environment mapping effect that is found in a large number of ps1 racing games. Those shiny reflections on your car that sell the effect of your car being made out of metal and glass. The ps1 doesn't have native support for environment mapping or any kind of multitexturing, so how do these games do it?
Well, they actually render your car twice. Once with the texture representing the body paint color, and then a second pass with a texture representing the current environment UV mapped over your car. The second pass is rendered with a transparency mode that neatly adds the reflections over the body paint. The player doesn't even notice that transparency was used.
Environment mapping is perhaps the most visually impressive of these advanced rendering tricks using translucency, but you can find dozens of variations spread out though the entire PS1 and N64 libraries.
Probably the most pervasive of these tricks are decals. Decals are great for providing visual variation to large parts of level geometry. Rather than your walls being a single repeating texture, the artist can just slap random decals every so often. A light switch here, a patch of torn wallpaper there, blood spatters here and graffiti over there. It the kind of thing the player doesn't really notice except when they are missing.
And decals are implemented with transparency. Just draw the surface with it's normal repeating texture and then use transparency to blend the decals over top in a second pass. Sometimes they Add, sometimes they Subtract. Sometimes they just want to skip transparent texels.
And because the Saturn doesn't support transparency on 3D quads, it can't do proper decals. So artists need to manually model any decal-like elements into the level geometry itself, which wastes artist time, wastes texture memory and wastes polygons.
For some reason, VDP1 doesn't even support transparent texels on 3D quads, the manual says they only work for non-distorted quads.
> Side note but forward mapping is also the reason why half-transparency does not work for distorted sprites. Forward mapping means that the same pixel may be overdrawn. If a half-transparent pixel is blended twice, the result is a corrupt pixel.
No... This isn't a limitation of forwards texture mapping. Just a limitation of VDP1's underwhelming implementation of forwards texture mapping.
I don't think it would be that hard to adjust the edge walking algorithm to deal with "holes" in a way that avoids double writes.
My gut says this is basically the same problem that triangle-based GPUs run into when two triangles share an edge. The naive approach results in similar issues with both holes and double-written pixels (which breaks transparency). So triangle rasterizers commonly use the top-left rule to ensure every pixel is written exactly once by one of the two triangles based on which side of the edge the triangle is on.
----------
Though that fix only allows 3D quads to correctly use the half-transparency mode.
I'm really confused as to why VDP1 has such limited blending capabilities. The SNES had proper configurable blending in the 90s, and VDP2 also has proper blending between it's layers.
And it's not like proper blending is only useful for 3D games. Proper blending is very useful for 2D games too, and certain effects in cross platform 2D games can end up looking washed out on the Saturn version because they they couldn't use VDP2 for that effect and were limited to just VDP1's 50% transparency on VDP1 instead of add/subtract.
VDP1's 50% transparency already pays the cost of doing a full read-write-modify on the framebuffer, so why not implement the add/substract modes? A single adder wouldn't be that many transistors? That would be enough to more or less match the playstation, which only had four blending modes (50/50, add, subtract and 25% of background + 100% of foreground).
Though, you can go a lot further than just equality with the PS1. The N64 had a fully configurable blender that could take alpha values from either vertex colors or texture colors (or a global alpha value).
> VDP2 can also provide half-transparent effects in various circumstances - this video provides a comprehensive look at various methods which were used to deliver this effect.
> No. Proper transparency is on my list because it's an essential building block for emulating advanced features that aren't natively supported on the hardware. One great example is the environment mapping effect that is found in a large number of ps1 racing games.
That's a good point which I had not considered. However, just adding support for half-transparency wouldn't have been enough to bring the Saturn up to speed in this regard. Half-transparency is also really damn slow on the Saturn. Quoting the VDP1 User's Manual:
> Color calculation of half-transparency is performed on the pixel data of the original
graphic and the pixel data read from the write coordinates. Drawing in this case
slows down, so use caution—it takes six times longer than when color calculation is
not performed.
This compounds with the Saturn VDP1 already having a considerable reduction in fillrate compared to the Playstation GPU. Playstation wasn't just capable of transparency, it was also damn fast at it, so emulating effects with transparency was feasible.
My proposed changes only really bring the Saturn's 3D graphics into roughly the same technical capabilities as the Playstation.
They don't do anything to help with the performance issues or the higher manufacturing costs. Or the string of bad moves from SEGA management that alienated everyone.
But an alternative history with these minor changes might have been enough to stop developers complaining about the Saturns non-standard 3D graphics and keep a healthy market of cross-platform games. It might have also prevented commenters on the internet from making overally broad claims about the Saturn's lack of 3D capabilities.
---------------
Fixing the performance limitations and BoM cost issues would have required a larger redesign.
I'm thinking something along the lines of eliminating most specialised 2D capabilities (especially all of VDP2). Unify what's left to a single chip and simplify the five blocks of video memory down to a single block on a 32bit wide bus. This would significantly increase the bandwidth to the framebuffer, compared to VDP1's framebuffer which uses a 16bit bus, while decreasing part count and overall cost.
The increased performance of this new 3D focused VDP should hopefully be powerful enough to emulate the lost VDP2 background functionality as large 3D quads.
The design you're describing sounds a lot more like the Playstation. I find most of these hypotheticals about "fixing" the Saturn work out that way :)
Playstation indeed had unified video memory on a 32 bit bus, but the bus is only saturated when drawing 2D elements or untextured polygons[1].
Sprite-heavy 2D games like DoDonPachi already run better on Playstation than Saturn due to the faster framebuffer, while games like Spyro the Dragon, Crash Bandicoot and Final Fantasy 7 demonstrate the effective strategic use of untextured polygons.
Yeah. It essentially ends up as the Playstation GPU, but with quads and forwards texturing.
It's basically inevitable. Any proposal for fixing the Saturn that slaps on a "optimise for BoM" requirement will be pushed towards unified video memory. And the only contemporary example of such a GPU was the playstation. Not to mention, the design of the playstation GPU appears to be pretty close to optimal for it's era and price point.
And if you were to also switch your hypothetical Saturn design to triangles and inverse texturing, the resulting design looks even more like the Playstation GPU.
> but the bus is only saturated when drawing 2D elements or untextured polygons
The Playstation's bus will also be saturated when loading the texture cache, or when scanning out the previous frame (only on the later SGRAM versions of the GPU). Even with textured triangles framebuffer writes are limited to 16bit bandwidth, less time is wasted on those other tasks which periodically interrupt rasterisation.
Though, now that I think about it, inverse texturing probably gives the playstation a massive performance advantage over and forwards texturing. Inverse texturing allows triangles to be rasterized in strict left-to-right then top-to-bottom order, which is also the order that the framebuffer is laid out in memory. This order will minimise the amount of time spent waiting for the next DRAM row to be prefeched.
The order of framebuffer writes with forwards texturing is somewhat chaotic and probably wastes a bunch more time closing and opening DRAM rows.
The memory access pattern of inverse texturing is just as chaotic when it comes to reading texels, which is why the Playstation's texture cache plays an important role in its performance characteristics.
Saturn doesn't have a texture cache, since texel reads should be nice and sequential. But then it pays a penalty when writing to the framebuffer.
We have at least one examples of this being implemented in real hardware at around the same time: Nvidia's NV1 (their first attempt at a GPU) was a quad based GPU that used forwards texture mapping. And it had both UV mapping and translucent polygons. The sega saturn titles that were ported to the NV1 look absolutely great. And I the Sega Model 3 also implemented both UV mapping and transparency on quads (I'm just not 100% sure it was using forwards texture mapping).
I'm not suggesting the Saturn should have implemented the full perspective correct quadratic texture mapping that the NV1 had. That would be overkill graphically (especially since the ps1 didn't have perspective correct texture mapping either), and probably would have taken up too many transistors.
But even a simple scheme that ignored perspective correctness were the artist provided UV coord for three of the vertices (and the forth was implicitly derived by assuming the quad and it's UVs were flat) would have provided most of the texture mapping capabilities that artists needed. And when using malformed quads to emulate triangles, the 3 UV coords would match.
And most of the hardware is already there. VDP1 already has hardware to calculate the slopes all four quad edges from vertex coords and an edge walker to walk though all pixels in screen space. It also has similar hardware for calculating the texture space height/width deltas and walking though texture space. It's just that texture space is assumed to always have slopes of zero.
So that's the only changes we make. VDP1 now calculates proper texture space slopes for each quad, and the texture coord walker is updated to take non-zero slopes. I'm hopeful such an improvement could have been made without massively increasing the transistor count of VDP1, and suddenly the saturn's texture mapping capabilities are roughly equal with the PS1.
> I don't consider this to be a significant problem. The checkerboard mesh feature provides a workable pseudo-half-transparent effect
I think you are underestimating just how useful proper transparency is for early 3D hardware.
(by proper transparency, I mean supporting the add/subtract blend modes)
And I haven't put it on my list of "important missing features" because sometimes games want to render transparent objects. As you said, the mesh (checkerboard) mode actually works for a lot of those cases. And if you are doing particles, then you can often get away with using non-distorted quads, which does work with the half-transparent mode. (though, from an artistic perspective, particles would really look a lot better if they could instead use add/subtract modes)
No. Proper transparency is on my list because it's an essential building block for emulating advanced features that aren't natively supported on the hardware.
One great example is the environment mapping effect that is found in a large number of ps1 racing games. Those shiny reflections on your car that sell the effect of your car being made out of metal and glass. The ps1 doesn't have native support for environment mapping or any kind of multitexturing, so how do these games do it?
Well, they actually render your car twice. Once with the texture representing the body paint color, and then a second pass with a texture representing the current environment UV mapped over your car. The second pass is rendered with a transparency mode that neatly adds the reflections over the body paint. The player doesn't even notice that transparency was used.
Environment mapping is perhaps the most visually impressive of these advanced rendering tricks using translucency, but you can find dozens of variations spread out though the entire PS1 and N64 libraries.
Probably the most pervasive of these tricks are decals. Decals are great for providing visual variation to large parts of level geometry. Rather than your walls being a single repeating texture, the artist can just slap random decals every so often. A light switch here, a patch of torn wallpaper there, blood spatters here and graffiti over there. It the kind of thing the player doesn't really notice except when they are missing.
And decals are implemented with transparency. Just draw the surface with it's normal repeating texture and then use transparency to blend the decals over top in a second pass. Sometimes they Add, sometimes they Subtract. Sometimes they just want to skip transparent texels. And because the Saturn doesn't support transparency on 3D quads, it can't do proper decals. So artists need to manually model any decal-like elements into the level geometry itself, which wastes artist time, wastes texture memory and wastes polygons.
For some reason, VDP1 doesn't even support transparent texels on 3D quads, the manual says they only work for non-distorted quads.
> Side note but forward mapping is also the reason why half-transparency does not work for distorted sprites. Forward mapping means that the same pixel may be overdrawn. If a half-transparent pixel is blended twice, the result is a corrupt pixel.
No... This isn't a limitation of forwards texture mapping. Just a limitation of VDP1's underwhelming implementation of forwards texture mapping.
I don't think it would be that hard to adjust the edge walking algorithm to deal with "holes" in a way that avoids double writes.
My gut says this is basically the same problem that triangle-based GPUs run into when two triangles share an edge. The naive approach results in similar issues with both holes and double-written pixels (which breaks transparency). So triangle rasterizers commonly use the top-left rule to ensure every pixel is written exactly once by one of the two triangles based on which side of the edge the triangle is on.
----------
Though that fix only allows 3D quads to correctly use the half-transparency mode.
I'm really confused as to why VDP1 has such limited blending capabilities. The SNES had proper configurable blending in the 90s, and VDP2 also has proper blending between it's layers.
And it's not like proper blending is only useful for 3D games. Proper blending is very useful for 2D games too, and certain effects in cross platform 2D games can end up looking washed out on the Saturn version because they they couldn't use VDP2 for that effect and were limited to just VDP1's 50% transparency on VDP1 instead of add/subtract.
VDP1's 50% transparency already pays the cost of doing a full read-write-modify on the framebuffer, so why not implement the add/substract modes? A single adder wouldn't be that many transistors? That would be enough to more or less match the playstation, which only had four blending modes (50/50, add, subtract and 25% of background + 100% of foreground).
Though, you can go a lot further than just equality with the PS1. The N64 had a fully configurable blender that could take alpha values from either vertex colors or texture colors (or a global alpha value).
> VDP2 can also provide half-transparent effects in various circumstances - this video provides a comprehensive look at various methods which were used to deliver this effect.
> https://www.youtube.com/watch?v=f_OchOV_WDg
Yeah, that's a great video. And it's the main reason why I started digging into the capabilities of the Saturn.