Dream Club portable graphic issue

[sum2012]

v0.9.8-1000-ga675c30 no problem

v0.9.8-1018-g318f641 human some problem
v0.9.8-1020-gcda9833 cannot see human
v0.9.8-1061-gcda9833 cannot see human
v0.9.8-1142-g1415532-windows-amd64\ bg full black
v0.9.8-1206-g9503603 cannot see human
v0.9.8-1219-gf664979 cannot see human
v0.9.8-1235-gd1357e5 semi problem on human
v0.9.8-1282-g935f1cb semi problem on human
v0.9.8-1538-gbe0340b semi problem on human

v0.9.8-1018-g318f641 log:
https://gist.github.com/sum2012/13c11126d5b9add2ef81
v0.9.8-1538-gbe0340b log:
https://gist.github.com/sum2012/b8d15983366d98e19a7e

[sum2012]

I find a solution in v0.9.8-1538-gbe0340b
Windows size set from 2x to 1x
then no problem
edit: (can set back 2x)

[unknownbrackets]

What does the GE debugger show when drawing the missing elements?

-[Unknown]

[sum2012]

v0.9.8-1000-ga675c30 does draw

v0.9.8-1538-gbe0340b doesn't

[unknownbrackets]

Darn, I thought it'd be related to framebuf size. Does it help if you disable any of the tests during drawing? Like alpha, stencil, depth? We did change the value we initially had framebuffers set to... weird that it has gone through so many iterations...

-[Unknown]

[sum2012]

After change the flag's value ,press resume ?
Also , when do change flag's value ? In the https://cloud.githubusercontent.com/assets/2177532/3484243/c45734fa-03a3-11e4-9697-27492fb5fd5d.png ?

. Does it help if you disable any of the tests during drawing

btw:
Press Escape twice also fix the missing elements

[sum2012]

Press Escape twice log

[unknownbrackets]

Pressing escape fixes it, huh? That implies that it's something to do with the stencil, depth, or alpha since it will reset the buffers. Maybe we are not detecting a memcpy.

504 width is a strange number.

-[Unknown]

[sum2012]

I re-doing the log (ignore sceunulity)
https://gist.github.com/sum2012/79cbbdad9e09f020e025
The 504 is from
46:34:434 idle0 I[SCEGE]: GLES\Framebuffer.cpp:823 Resizing FBO for 00000000 : 504 x 272 x 1
47:05:708 idle0 I[SCEGE]: GLES\Framebuffer.cpp:823 Resizing FBO for 00044000 : 504 x 272 x 1
47:22:408 idle0 I[SCEGE]: GLES\Framebuffer.cpp:2038 Decimating FBO for 00044000 (504 x 272 x 1), age 6

Debug log (ignore sceunulity)
https://drive.google.com/file/d/0B3OaSdeV0L8kY1MtbnpNZEpmR2M/edit?usp=sharing

[sum2012]

Change stencil, depth, and alpha 's value to disable,Press "Resume" not work .

[sum2012]

Re-start once the girl appear with flash
Press two times "step Frame" result differuclt

[sum2012]

I review the issue.
edit :The first bisect result (from good to nothing) is
32439bc

[sum2012]

Hmm,Another good test

edit v0.9.8-1224-g43bddd8 semi problem on human
edit :v0.9.8-1228-gf0f8f72 semi problem on human
v0.9.8-1235-gd1357e5 semi problem on human

edit:The changes are
v0.9.8-1224-g43bddd8...v0.9.8-1228-gf0f8f72

[sum2012]

@unknownbrackets git bisect result:
a54cefc
Hope you have idea for correct fix.

Anyway,v1.0.1-857-g08b340f gpu debugger

[sum2012]

No,The git bisect result is wrong :(

[unknownbrackets]

Has anything improved here recently?

-[Unknown]

[daniel229]

No.

[sum2012]

v1.1.1-760-g07a1670 same problem
soft gpu no missing element but ugly

Can we try some code of invalidate to fix it ?

[unknownbrackets]

Haha, wow that is bad.

I still really wonder if this is depth related. If you change the StateMapping.cpp file, and find:

            glstate.depthFunc.set(ztests[gstate.getDepthTestFunction()]);

And replace it with:

            glstate.depthFunc.set(GL_ALWAYS);

Does it draw all parts of the character? It'll probably draw them wrong, but I'm wondering if it will at least draw everything.

-[Unknown]

[sum2012]

Yes.It draw all parts of the character.
Just a little ugly

[unknownbrackets]

So, the reason estimation affected this is because we only blit the depth if the render width and height matched.

Find this in Framebuffer.cpp:

void FramebufferManager::BlitFramebufferDepth(VirtualFramebuffer *src, VirtualFramebuffer *dst) {

And add below:

if (src->z_address == dst->z_address && src->z_stride != 0 && dst->z_stride != 0) {
    NOTICE_LOG(HLE, "Framebuffer switch with same depth, %dx%d -> %dx%d", src->renderWidth, src->renderHeight, dst->renderWidth, dst->renderHeight);
}

What does it say? Ideally, use the GE debugger to click "Step Frame" twice. That will show only one frame of the log, which will be most ideal.

-[Unknown]

[unknownbrackets]

Actually, how about this function instead:

void FramebufferManager::BlitFramebufferDepth(VirtualFramebuffer *src, VirtualFramebuffer *dst) {
    bool matchingDepthBuffer = src->z_address == dst->z_address && src->z_stride != 0 && dst->z_stride != 0;
    bool matchingSize = src->width == dst->width && src->height == dst->height;
    if (matchingDepthBuffer && matchingSize) {
        int w = std::min(src->renderWidth, dst->renderWidth);
        int h = std::min(src->renderHeight, dst->renderHeight);

        if (gstate_c.Supports(GPU_SUPPORTS_ARB_FRAMEBUFFER_BLIT | GPU_SUPPORTS_NV_FRAMEBUFFER_BLIT)) {
            // Only use NV if ARB isn't supported.
            bool useNV = !gstate_c.Supports(GPU_SUPPORTS_ARB_FRAMEBUFFER_BLIT);

            // Let's only do this if not clearing depth.
            fbo_bind_for_read(src->fbo);
            glstate.scissorTest.force(false);

            if (useNV) {
#if defined(USING_GLES2) && defined(ANDROID)  // We only support this extension on Android, it's not even available on PC.
                glBlitFramebufferNV(0, 0, w, h, 0, 0, w, h, GL_DEPTH_BUFFER_BIT, GL_NEAREST);
#endif // defined(USING_GLES2) && defined(ANDROID)
            } else {
                glBlitFramebuffer(0, 0, w, h, 0, 0, w, h, GL_DEPTH_BUFFER_BIT, GL_NEAREST);
            }
            // If we set dst->depthUpdated here, our optimization above would be pointless.

            glstate.scissorTest.restore();
        }
    }
}

-[Unknown]

[sum2012]

Talk in the first change first

Find this in Framebuffer.cpp:

void FramebufferManager::BlitFramebufferDepth(VirtualFramebuffer *src, VirtualFramebuffer *dst) {

And add below:

if (src->z_address == dst->z_address && src->z_stride != 0 && dst->z_stride != 0) {
    NOTICE_LOG(HLE, "Framebuffer switch with same depth, %dx%d -> %dx%d", src->renderWidth, src->renderHeight, dst->renderWidth, dst->renderHeight);
}

The GE debugger graphic not same as the first time (Step Frame)

There is invalid address in the log (in your change)!!!

[sum2012]

In your second change,

void FramebufferManager::BlitFramebufferDepth(VirtualFramebuffer *src, VirtualFramebuffer *dst) {
    bool matchingDepthBuffer = src->z_address == dst->z_address && src->z_stride != 0 && dst->z_stride != 0;
    bool matchingSize = src->width == dst->width && src->height == dst->height;
    if (matchingDepthBuffer && matchingSize) {
        int w = std::min(src->renderWidth, dst->renderWidth);
        int h = std::min(src->renderHeight, dst->renderHeight);

        if (gstate_c.Supports(GPU_SUPPORTS_ARB_FRAMEBUFFER_BLIT | GPU_SUPPORTS_NV_FRAMEBUFFER_BLIT)) {
            // Only use NV if ARB isn't supported.
            bool useNV = !gstate_c.Supports(GPU_SUPPORTS_ARB_FRAMEBUFFER_BLIT);

            // Let's only do this if not clearing depth.
            fbo_bind_for_read(src->fbo);
            glstate.scissorTest.force(false);

            if (useNV) {
#if defined(USING_GLES2) && defined(ANDROID)  // We only support this extension on Android, it's not even available on PC.
                glBlitFramebufferNV(0, 0, w, h, 0, 0, w, h, GL_DEPTH_BUFFER_BIT, GL_NEAREST);
#endif // defined(USING_GLES2) && defined(ANDROID)
            } else {
                glBlitFramebuffer(0, 0, w, h, 0, 0, w, h, GL_DEPTH_BUFFER_BIT, GL_NEAREST);
            }
            // If we set dst->depthUpdated here, our optimization above would be pointless.

            glstate.scissorTest.restore();
        }
    }
}

Well done,it fixed

[sum2012]

git bisect result e14437c

[hrydgard]

Fixed it for OpenGL again. Moving further work to 1.12.

[sum2012]

Re-test in v1.10.3-1324-g226a8be1c
Yes,Remain Directx 9 and Direct 11 have problem

[unknownbrackets]

The frame dump here appears to work for me in Direct3D 11 and Direct3D 9. Is this potentially fixed? Or does it not reproduce in the dump now?

-[Unknown]

[sum2012]

in v1.12.3-607-g2d7a7fd34
Opengl and Vulkan fixed.
Dx9 and DX 11 girl disappear

[sum2012]

My harddisk die again :(
Now I focus in making game in renpy

[unknownbrackets]

With the recent changes to depth buffer copying, I think this is even more likely to be fixed. The INTZ change may have helped D3D9, potentially.

-[Unknown]

[sum2012]

OpenGL , DX9,Valkun no problem
DX11 and Soft gpu have problem
DX11:
https://img.chkaja.com/92bde1cfd326de86.jpg
Softgpu a little problem:

Update frame dump
ULJS00333.zip

[unknownbrackets]

The correct rendering:

Software renders:

There's some precision issue with some coordinates. But it's actually pretty close.

D3D11 does better with the accuracy:

The frame dump works on Windows 10 / Direct3D 11 using both NVIDIA and Intel. I wonder if some drivers don't support depth readback or require it in a different format.

-[Unknown]

[sum2012]

@unknownbrackets some graphic problem may only see by part of iso (umd bufer)
jpcsp made it in jpcsp/jpcsp@c2d7c80
Do you interest it ?

[unknownbrackets]

I'm not really sure what that is, but it seems like a tool to see what sectors of the UMD the game read from. I don't think it's relevant here?

-[Unknown]

[sum2012]

@unknownbrackets I only mean that easier to reproduce the problem

[sum2012]

Test in v1.16.3-14-gc4ad32420 of AMD cpu 7900 of AMD Integrated graphics card in windows 11
All render mode okay now.