|
Something you mentioned earlier comes back to mind. Maybe changing your point of view is what's needed.
Switch over to writing a use-case for this API. What's the most useful behavior for what's being returned? If only one 'thing' can be active, and the returned item is no longer the active 'thing', then what about making the returned object read-only? You can return information from it that either doesn't require the active state, or is a copy of the last value held while the 'thing' was in the active state.
Software Zen: delete this;
|
|
|
|
|
The thing is if anything, write only, because it's a screen mode. You can draw to it to draw to a screen. (some screens support reading, but all screens support writing or they wouldn't be screens!  )
Basically you instantiate a device, but the device itself has several screen modes. because of the way this code works, they must different actual types.
Hence the design complication.
Real programmers use butterflies
|
|
|
|
|
Whenever I've been in this state, pounding my head on the desk trying to decide something, it's always been the case I was looking at the problem incorrectly.
I'll admit I've used a brute-force solution fairly often of making the simplest decision and implementing it, knowing that it wasn't adequate and would fail. Iterating through this again and again, while looking wasteful, was a way of eroding the mental block. The act of implementing an idea, making it concrete, puts the brakes on the fruitless "nope, nope, nope, nope" cycles. Getting down to the details of each inadequate solution helped me learn about the actual shape and edges of the problem.
This also lets you throw something at the problem other than your peace (piece?) of mind, which is useful in my environment. Good luck.
Software Zen: delete this;
|
|
|
|
|
Absolutely. I've been telling myself, just code it, and then deal with it, knowing that if I choose poorly, I'll have to rewrite 90% of the driver code because of how it operates. The design decision has ramifications that impact the entire driver's codebase.
It's just tough to get that motivation to continue going feeling like I'll probably have to do it over again. Or at least I think that's part of it.
I'll work through it eventually. I feel a little obligated to make it a priority based on me accepting goods in exchange for my coding, even if it was informal. Just a little pressure though, not a lot.
Real programmers use butterflies
|
|
|
|
|
honey the codewitch wrote: I can't put myself in the shoes of a person using my code
As far a I'm concerned, you're already ahead of the game by just considering another person using your code. I see far too many implementations where "oh, someone else might use this code besides me" clearly was never considered. Particularly the way API endpoints are coded. Definitely a WTF, as I see endpoints that are unusable except for the very narrow use case of a specific front-end implementation.
|
|
|
|
|
I always try to design my APIs like I'm going to be forced to use them. 
Real programmers use butterflies
|
|
|
|
|
Without code to go on, this is very speculative. Even code might not help, because of the time that would be needed to study it, trying to infer the design rationale before providing what might have a shot at being useful unsolicited advice.
But why should any of that be a barrier? 
1. How about returning void? The user must know they're switching modes, so they should adjust accordingly.
2. If switching modes can fail, even because of bad arguments, then a bool would be better.
3. If the previous mode can no longer be used, how about returning the new one? Can't it actually be returned?
4. There's also a "draw target". How about returning that?
5. I assume there's a "handle" to which clients write, and that it's polymorphic. So can the new one be returned?
|
|
|
|
|
There is no polymorphism.
Draw targets are those handles you write to.
draw::point(my_draw_target,spoint16(10,15),color_max::red);
Here, my_draw_target is what gets drawn to. above draws a point at (10,16), color of red
my_draw_target in this case, is say, screen mode 0
but I now want to switch screen modes
auto my_draw_target2 = driver.mode<1>();
draw::line(my_draw_target2,srect16(0,0,9,14),color_max::antique_white);
In the above case, once you have my_draw_target2, my_draw_target becomes persona non grata. How that is handled is the question, meaning what happens if i then try to draw again to my_draw_target?
PS: It may appear that my_draw_target and my_draw_target2 are polymorphic but they are not. They do sort of have a shared interface, but they don't. It's complicated. I bind using templates, not virtual base classes
Real programmers use butterflies
|
|
|
|
|
So how about wrapping it in a kind of de-reference handler for the user to use.
The user only creates one Context object, which contains the current draw_target.
Changing modes doesn't create a new object for them, but sets the internal draw_target of the Context.
The draw::line methods accept the Context instead of the draw_target and access the internal draw_target.
There is much I don't understand about what you're doing with the template binding, so this may not be at all feasible. It's also adding an abstraction layer, and I know you're working in resource constrained environments. However, that abstraction layer may be what you need to present a better API to the user.
|
|
|
|
|
That's not possible because there is no common base to use.
Different draw targets are entirely distinct types.
Furthermore, each draw target is typed by the type of its pixel.
So for example, if you have a bitmap<rgb_pixel<16>> and bitmap<rgb_pixel<24>>, the former indicating a bitmap with 16-bit color RGB pixels and the latter indicating 24-bit RGB, each bitmap is a different type. And it must be because the drawing code for a 16-bit pixel is different than that of a 24-bit pixel because the layout in memory of what it's drawing is different.
A bitmap<> is an example of a draw target, but each screen you can draw to is a draw target as well. A draw target is anything you can draw on.
Ergo, each screen mode is also a draw target, and if it has a different type of pixel than the next, it must be a different type.
There can be no common base since 90% of the drawing methods will take a pixel and that has a particular type. So even if i made a base type, that type would have to be a template.
This is generic programming and it's powerful, but it's not polymorphic in the traditional way we think about polymorphism.
It sounds like you might be a C++ dev but I don't want to assume. If you're not a C++ dev I'm not sure if there's an analogy for this style of coding in any other language.
OO principles only loosely apply here though. GP is a different animal.
Real programmers use butterflies
|
|
|
|
|
No, I've never developed anything in C++. Plenty of other languages, including plain C, but not C++. I have a vague theoretical understanding of what you're doing and would love to have time to explore it more (sadly, I don't)
Logically, if I have a screen device configured in a particular mode and I reconfigure it to a new, incompatible mode, then I am essentially removing one device and adding another.
All reference to the original device is now invalid, and any attempt to reference it is invalid.
If, as a programmer using your library, I make an invalid call, I would expect a simple error back.
The trickier question now becomes, if I reconfigure it back, am I adding the original device back? I'd say no, it is a new device with the same configuration as one that was previously removed.
|
|
|
|
|
Thank you! That's actually very helpful. I've been considering your approach but my fear was that it would be difficult or counterintuitive. Since you basically articulated to me that you expect precisely one of the avenues I was considering, it seems I have a clear winner in terms of approaches to this.
Real programmers use butterflies
|
|
|
|
|
I do not know the details of your code, only what I read in this thread, but I was thinking the same as NeverJustHere with one change.
I would probably make the class defining the device/target a singleton so that getting a new target would make the old reference point to the new target and probably no errors would have to be thrown.
If a method is added to check the current drawing mode you have all your bases covered.
Maybe!
Hope it helps. Good luck.
|
|
|
|
|
Yeah that's not doable because each mode must be a fundamentally different type. You can't represent mode 1 with a class for mode 2.
If you could, I'd just have a "set mode" method on the driver class, and then use the driver class as the draw target like i do with my single screen mode drivers.
Real programmers use butterflies
|
|
|
|
|
The Witch wrote: I bind using templates, not virtual base classes
That's your problem right there.
|
|
|
|
|
It would be impossible to do what I am doing efficiently and by hand in C++ using any other method than what I'm doing.
Generic programming has no real substitute in terms of C++ functionality. You basically *can't* replace it with virtual base classes unless you were using it where you didn't need it in the first place.
Consider this:
I have a pixel. The pixel is user-defined in terms of the number of color channels it has, their names, and what their binary layout, min and max channel values are, etc. This is too involved to compute at run time. It would kill performance
Now, for every draw method, it deals in pixels like this, but because a pixel must be typed to its functionality, literally the only way I could make it polymorphic is to use a vtbl every time i needed a pixel. That's just not acceptable in terms of performance.
Binding using templates is the solution. That way you can inline, or you can otherwise aggressively optimize away those calls.
Real programmers use butterflies
|
|
|
|
|
Wouldn't polymorphic flyweights work?
You've piqued my interest. If one of your recent articles discusses this code, please point me to it.
|
|
|
|
|
I don't have an article that explicitly talks about how the code works, just how to use it.
But..
gfx_demo/gfx_pixel.hpp at master · codewitch-honey-crisis/gfx_demo · GitHub[^]
Consider the convert<>() routine starting at line 690.
It's massive. It's ugly. It would be slow as heck to do at runtime. This will get called width*height times during a call to draw a bitmap when format conversion needs to be done.
If any of the routine did not result in aggressive compile time operations, to where this routine translates to a few shifts and multiply/divides at worst - the thing would slow to a crawl if I had any JMPs in there. There are none despite what it looks like, except sometimes for clamping, which often gets compiled to ifless code anyway.
Real programmers use butterflies
|
|
|
|
|
|
Not even a little.
Real programmers use butterflies
|
|
|
|
|
Time to look at some creational design pattern?
"In testa che avete, Signor di Ceprano?"
-- Rigoletto
|
|
|
|
|
I have no choice but to use the factory pattern for creation.
Real programmers use butterflies
|
|
|
|
|
When you're stuck on a super important issue with significant consequences and you just can't decide which of your options is the best one, use The Grand Query Of Decision Making (TGQODM).
Its MySQL version looks like this:
SELECT IF(RAND() < 0.5, 'Do option 1', 'Do option 2')
|
|
|
|
|
I'm not sure if I even fully understand the problem but it sounds a bit like you could benefit from a design change where you maintain state information for each draw mode. This would allow you to switch between modes without loosing anything since all relevant data would be saved before switching modes.
|
|
|
|
|
I don't have that kind of memory. The e-paper displays essentially require me to keep the frame buffer around in local RAM. So my size in bytes for monochrome black and white displays is width*height/8 and it gets worse when you introduce virtualization of higher bit depths through dithering, such that you can simulate 8-bit grayscale for example, but at that point your frame buffer size becomes simply width*height in bytes. Ouch. That's my state for each mode.
It's not a good idea to hold a frame buffer around when someone is not using it on these devices. The devices I'm targeting with these displays range in RAM anywhere from about 300k to 4.5MB, with the caveat that 4MBs of that is PSRAM over SPI and thus slow.
Real programmers use butterflies
|
|
|
|
|
General 
News 
Suggestion 
Question 
Bug 
Answer 
Joke 
Praise 
Rant 
Admin 
Submit an article or tip
