Imgui: Rendering issue

[or75]

http://i.imgur.com/v3oSMok.gifv

Correct )

[aggsol]

Please describe your issue in detail including steps to reproduce and solutions you tried already. Also attach screenshots here at github. Thank you.

[or75]

1) Open in demo app -> (Widgets Header)
2) Open Word Wrapping
3) Open Selectables
4) Open Selectables -> In Columns
5) Open Selectables -> Grid
6) Open Filtered Text Input
7) Open Multi-Line Text Input
8) Scroll, it's all

directx9_example - bug
directx11_example - bug
opengl_example - Works because no "Selectables -> In Columns"
opengl2_example - bug
opengl3_example - bug

Tested on:
OS: win10 x64 pro (1703 15063.332)
Video card: gtx 1050

[ocornut]

That's super strange. I couldn't repro it here.

Do you get the bug with the precompiled binaries?
http://www.miracleworld.net/imgui/binaries/imgui-demo-binaries-20161113.zip

[or75]

Yes, it's the same here.

[paulsapps]

Must be buggy drivers?

[or75]

Perhaps, yes, but why it happens with some widgets only when they are open ?
Word Wrapping,
Selectables -> In Columnsm,
Selectables -> Grid,
Filtered Text Input,
Multi-Line Text Input

[ocornut]

What are your OS and graphics cards?
There are more vertices now.
You can try opening the Metrics window to browse the generated vertices, and see if there is a visible problem with the vertex being created?

[or75]

OS: win10 x64 pro (1703 15063.332)
Graphics card: gtx 1050

(Directx9-11) -> Something incomprehensible in general:

(Opengl2-3) -> (Only the vertices disappear here)

[dpethes]

Do you have any issues in other 3D apps? That direct3d screen looks a bit like issues with overclocking/overheating.

[paulsapps]

I agree, looks like it might be some sort of hardware damage? Do games etc work OK?

[or75]

Everything works fine (Games,App,Etc)

[ocornut]

The metrics window looks good so the bug isn't in imgui core (but however unlikely, might be something in the examples app).

The render loop are using scissoring which is rather infrequently used by apps. What happen if you disable scissor enable from the examples renderers?

I wonder if some of your installed software could be hooking into the DX/OGL rendering API and doing something funky as well?

[dmitryhryppa]

I have the same issue as well while scrolling ImGui Demo window.
gtx970, Windows 10 x64

http://i.imgur.com/udOqIrc.gifv

[ocornut]

@or75 @dmitryhryppa Could you try all the precompiled binaries package to confirm that it once worked (a while ago) and got broken sometimes? http://www.miracleworld.net/imgui/binaries/
Thanks!

[dmitryhryppa]

@ocornut ,
imgui-demo-binaries-20160410 <-- working fine
imgui-demo-binaries-20161113 <-- issue exist here

Latest build contains "In columns" section inside "Selectables", whereas previous build does not contains it.
So, looks like "In columns" conflicting with some another example components.
Hope it helps.

[or75]

@dmitryhryppa Yes, that is right.

@ocornut @dmitryhryppa

1. Word Wrapping
2. Selectables -> In Columnsm
3. Selectables -> Grid,
4. Filtered Text Input
5. Multi-Line Text Input

If something is closed from this list, everything is fine :)

[Rubonnek]

I was able to reproduce this too while I was converting the ImDrawVert struct to one that was compatible with the Cocos2D-X renderer. I cannot see any graphical glitches running the examples alone, but I can see them through the Cocos2D-X renderer.

There seems to be a very slight memory corruption in the vertex buffer and can be seen by changing the ImDrawVert to the following:

#define IMGUI_OVERRIDE_DRAWVERT_STRUCT_LAYOUT \
struct ImDrawVert \
{ \
    ImVec2 pos; \
    float z = 1000.f; \
    ImU32 col; \
    ImVec2 uv; \
};

One would expect z here to always be 1000.f, but turns out it's not:

Z is: 0
Z is: 0
Z is: 3.87324e-38
Z is: 3.87324e-38
Z is: 2.24208e-43
Z is: 2.60603e-38
Z is: 7.21133e-26
Z is: 0
Z is: 3.64338e-43
Z is: 0
Z is: 3.58732e-43
Z is: 2.602e-38
Z is: 0
Z is: 1.56945e-43
Z is: 1.56945e-43
Z is: 1.56945e-43
Z is: 1.56945e-43

I tested that using the OpenGL 2 example in Arch Linux, and this test can be found here: https://github.com/Rubonnek/imgui/commit/aa1e159537eb7bc514220900d63677f0e4e75846

I just tried the same with the OpenGL 3 example, and I see z changing values too.

[thedmd]

@Rubonnek ImDrawVert does not support additional data, just reordering.

Vertex buffer is represented using ImVector<ImDrawVert>. ImVector::reserve() allocate memory and keep it uninitialized. Therefore every field which is not filled by ImGui will contain random trash. __I had to deal with this error and I chose to always initialize allocated memory to zero.__ You can examine changes there.

Alternative approaches (which failed):

• Adding default constructor to ImDrawVert does not work because ImVector works only with POD types and does not call constructors.
• Modifying ImDrawList:: PrimWriteVtx to initialize extra data in ImDrawVert works in theory. However, text rendering and a few other primitives does not use this function and access vertex buffer to batch data generation.

[paulsapps]

Running with valgrind will catch corruptions. I don't think this is a corruption, just uninitialized memory as thedmd said.

[Rubonnek]

@thedmd @paulsapps Thanks a lot for the information!

[ocornut]

The code intentionally avoid initializing memory or calling constructor because imgui doesn't rely on any of those for maximum perf.

I think we could change the ImDrawList code specifically for the purpose of calling either memset or the constructor after calling reserve() or resize(), and that only if ImDrawVtx is defined and its sizeof() is != sizeof of the default ImDrawVtx.

[ratchetfreak]

You can also use std::is_trivially_default_constructible from #include to check whether the initializer actually does something.

[thedmd]

This quick and hack should do the trick. Remember to include <type_traits>.

    inline void                 reserve(int new_capacity)
    {
        constexpr auto should_clear_memory = std::is_same<T, ImDrawVert>::value
            && sizeof(ImDrawVert) != 20;

        if (new_capacity <= Capacity) return;
        const size_t new_capacity_bytes = (size_t)new_capacity * sizeof(value_type);
        const size_t         size_bytes = (size_t)Size * sizeof(value_type);
        T* new_data = (value_type*)ImGui::MemAlloc(new_capacity_bytes);
        if (Data)
        {
            memcpy(new_data, Data, size_bytes);
            if (should_clear_memory)
                memset(new_data + Size, 0, new_capacity_bytes - size_bytes);
        }
        else if (should_clear_memory)
            memset(new_data, 0, new_capacity_bytes);
        ImGui::MemFree(Data);
        Data = new_data;
        Capacity = new_capacity;
    }

[paulsapps]

I think ImGui is using C++98, so probably can't use this.

[thedmd]

ImGui does not this that snippet. If @ocornut decide it may benefit an library, appropriate solution can be found.
It is a quick & dirty hack, and quick & dirty hacks are using any tool that gets the job done. They do not care to put label 'hot' on Starbucks coffee cup. : )

[ocornut]

Moved that discussion to #1231

[or75]

The problem is not yet solved in the last build (imgui-demo-binaries-20170723) all the same. (Even on Vulkan).
This problem still exists even on a laptop with (Intel HD Graphics 500).

[ocornut]

@or75 Have you tried updating your graphics driver?
(btw #1231 has nothing to do with your problem, it was a different issue that @Rubonnek brought to this thread)

[or75]

@ocornut I have everything updated (I do not know why such a problem).
I'll try to do something with the driver settings, if something happens, I'll tell you :)

[dtugend]

I am having the same issue with imgui-demo-binaries-20171013.

Interestingly it only seems to happen for me, when I expand "Filtered Text Input", can expand pretty much everything else, so maybe the problem is somehow related to that.

Edit: Seems to be related to ImGuiInputTextFlags_Password still investigating with Debugger.

Edit2: Seems to be related to PushFont / PopFont somehow.

Computer Information:
    Manufacturer:  LENOVO
    Model:  90B6000AGE
    Form Factor: Desktop
    No Touch Input Detected

Processor Information:
    CPU Vendor:  GenuineIntel
    CPU Brand:  Intel(R) Core(TM) i7-4790 CPU @ 3.60GHz
    CPU Family:  0x6
    CPU Model:  0x3c
    CPU Stepping:  0x3
    CPU Type:  0x0
    Speed:  3592 Mhz
    8 logical processors
    4 physical processors
    HyperThreading:  Supported
    FCMOV:  Supported
    SSE2:  Supported
    SSE3:  Supported
    SSSE3:  Supported
    SSE4a:  Unsupported
    SSE41:  Supported
    SSE42:  Supported
    AES:  Supported
    AVX:  Supported
    CMPXCHG16B:  Supported
    LAHF/SAHF:  Supported
    PrefetchW:  Supported

Operating System Version:
    Windows 10 (64 bit)
    NTFS:  Supported
    Crypto Provider Codes:  Supported 311 0x0 0x0 0x0

Video Card:
    Driver:  NVIDIA GeForce GTX 745
    DirectX Driver Name:  nvd3dum.dll
    Driver Version:  21.21.13.7688
    DirectX Driver Version:  21.21.13.7688
    Driver Date: 2 10 2017
    OpenGL Version: 4.5
    Desktop Color Depth: 32 bits per pixel
    Monitor Refresh Rate: 60 Hz
    DirectX Card: NVIDIA GeForce GTX 745
    VendorID:  0x10de
    DeviceID:  0x1382
    Revision:  0xa2
    Number of Monitors:  1
    Number of Logical Video Cards:  1
    No SLI or Crossfire Detected
    Primary Display Resolution:  1920 x 1080
    Desktop Resolution: 1920 x 1080
    Primary Display Size: 18.78" x 10.55" (21.54" diag)
                                            47.7cm x 26.8cm (54.7cm diag)
    Primary Bus: PCI Express 16x
    Primary VRAM: 2047 MB
    Supported MSAA Modes:  2x 4x 8x 

Sound card:
    Audio device: Lautsprecher (Realtek High Defi

Memory:
    RAM:  8140 Mb

I know the GPU driver is not the most recent, but was pretty satisfied with that one so far.

[dtugend]

There seems to be something wrong with the merging logic in UpdateTextureID:
https://github.com/ocornut/imgui/blob/09f6f564d91ae061bfe35763c8e738e9f7255774/imgui_draw.cpp#L225

I suppose curr_cmd->ElemCount == 0 && is missing there, because compare with the logic in UpdateClipRect:
https://github.com/ocornut/imgui/blob/09f6f564d91ae061bfe35763c8e738e9f7255774/imgui_draw.cpp#L206

Confirmed it's the merging logic by detouring the UpdateTextureID in the debugger to always use AddDrawCmd() branch of the if.

Edit: Also tested the Windows build of the pull request bellow, with fixed code issue doesn't seem to be reproducable anymore.

[ocornut]

WOW, really amazing and weird bug! I managed to reproduce it now, it was there all along! Sorry @or75!
It required a very specific combination of nodes opened and a specific clipping configuration:
edit minor edits.

Word Wrapping
Selectables -> In Columns
Filtered Text Input
+ enough stuff below to be able to scroll exactly until the first two are not possible.

• The code in Word-Wrapping render a custom rectangle by appending manually to the ImDrawList without testing for visibility, so those vertices are always appended into the draw list.

• The columns code using the ImDrawList channel splitting API which in ChannelsMerge() calls AddDrawCmd(), that itself is an error, it should call UpdateClipRect(). By calling AddDrawCmd it would always add a new command after merging. But in this case the Columns have no borders, and when the contents is out of screen the entire columns set produces zero vertices. So previously it would add an unnecessary extra draw command into the stream.

• The Password input text demo calls PushFont() as you pointed out, which calls UpdateClipRect() to give it a chance to either add a new draw command or merge them.

So what happened with that specific configuration and scrolling is that the word-wrapping demo code adds vertices, then the fully clipped columns adds a extraneous draw command, which then gets filled with the vertices below the columns area. So at this point in the draw list there are TWO draw commands with exactly the same clipping rectangle and texture, which shouldn't happen. And because it wasn't meant to ever happen the code in UpdateClipRect() essentially tries to apply the merge that was only supposed to happen when the second command is empty.

Note that if the Columns weren't clipped it wouldn't happen because there would be different draw command in-between for the columns contents, so merging wouldn't trigger. Also if the Word-Wrapping code wasn't submitting vertices regardless of its visibility there would haven't been two consecutive draw commands with the same clipping rectangle.

So, there are two fixes here:

• I fixed ChannelMerge() from creating a new command w/ commit: https://github.com/ocornut/imgui/commit/e32569dd5aa008448b10d9be11e5193502f5368b
• Your fix in #1368 is also correct.

Either of those fix will make the issue go away.
Because AddDrawCmd() is exposed to the user your fix and the similar test in curr_cmd->ElemCount == 0 are necessary and correct. If AddDrawCmd() wasn't exposed technically wouldn't be required, but it's saner and more logical to just test for that.

We are also reducing the number of draw calls when columns are fully clipped.

Thank you @or75 , thank you @ripieces ! And thanks @dmitryhryppa, @paulsapps, @aggsol , @dpethes.