MPlayer 1.0pre4 IDCT Optimisation (LONG POST!)

[lewis]

Dexter's patch didn't make as much difference as I'd hoped. Probably gcc sucks. A few extra FPS is still welcome, though :)

Vegac, as you rightly pointed out on #nekochan, the YCrCb extension doesn't work with glTexSubImage2D() so one can't have hardware colourspace conversion && textured zooming on O2... but actually you can, with some messing about with UMA. If you know the PBuffer-as-texture trick then disregard this, otherwise this is ripped wholesale from something else I'm writing so it should work, ish:

Obviously... link with -ldmedia too.

Code: Select all

#include <dmedia.h>


Make and init a texture, nothing new here, move along...

Code: Select all

int tex;

glEnable(GL_TEXTURE_2D);
glGenTextures(1, &tex);
glBindTexture(GL_TEXTURE_2D, tex);
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_LINEAR); 
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
glTexEnvi(GL_TEXTURE_ENV, GL_TEXTURE_ENV_MODE, GL_DECAL);
glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA, 1024, 1024, 0, GL_RGBA,
        GL_UNSIGNED_BYTE, NULL);


Make a DMbufferpool suitable for GL with just one buffer in it:

Code: Select all

DMparams *texPoolParams, *texImageParams;
DMbufferpool texPool;
DMbuffer texDMBuffer;

dmParamsCreate(&texImageParams);
dmSetImageDefaults(texImageParams, 1024, 1024, DM_IMAGE_PACKING_RGBA);
dmParamsSetEnum(texImageParams, DM_IMAGE_LAYOUT, DM_IMAGE_LAYOUT_GRAPHICS);

dmParamsCreate(&texPoolParams);
dmBufferSetPoolDefaults(texPoolParams, 1, 1024 * 1024 * 4, DM_FALSE, DM_FALSE);
dmBufferGetGLPoolParams(texImageParams, texPoolParams);

dmBufferCreatePool(texPoolParams, &texPool);
dmBufferAllocate(texPool, texDMBuffer);


Make a PBuffer:

Code: Select all

GLXFBConfig *fbConfigs;
GLXPbuffer texPBuffer;
int configCount;
int fbAttrs[] = {GLX_DRAWABLE_TYPE, GLX_PBUFFER_BIT,
            GLX_DOUBLEBUFFER, False,
            GLX_RED_SIZE, 8,
            GLX_GREEN_SIZE, 8,
            GLX_BLUE_SIZE, 8,
            GLX_ALPHA_SIZE, 8,
            None};
int pbAttrs[] = {GLX_PBUFFER_WIDTH, 1024,
            GLX_PBUFFER_HEIGHT, 1024,
            GLX_PRESERVED_CONTENTS, True,
            GLX_DIGITAL_MEDIA_PBUFFER_SGIX, True,
            None};

fbConfigs = glXChooseFBConfig(xDisplay, DefaultScreen(xDisplay), fbAttrs,
                         &configCount);
texPBuffer = glXCreatePbuffer(xDisplay, fbConfigs[0], pbAttrs);


Associate the PBuffer with the DMBuffer. The glTexSubImage2D() copies by reference (hurrah!) if everything above has worked, so you only need to do it once. Thereafter drawing into the PBuffer draws straight into the texture.

Code: Select all

glXAssociateDMPbufferSGIX(xDisplay, texPBuffer, texImageParams, texDMBuffer);
glXMakeContextCurrent(xDisplay, glxPBuffer, glxPBuffer, glxContext);
glCopyTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, 0, 0, 1024, 1024);


Now on each frame, you draw into the texture, then draw a textured quad as big as you like, keeping aspect would be a good plan:

Code: Select all

glXMakeContextCurrent(xDisplay, texPBuffer, texPBuffer, glxContext);
glRasterPos2i(0, height);
glPixelZoom(1.0, -1.0);
glDrawPixels(width, height, GL_YCRCB_422_SGIX, GL_UNSIGNED_BYTE, pixels);

glXMakeContextCurrent(xDisplay, glxWindow, glxWindow, glxContext);
glBegin(GL_QUADS);
 glTexCoord2f(0.0, 0.0);
 glVertex2i(0, 0);

 glTexCoord2f(0.0, (float)height/1024.0);
 glVertex2i(0, 1024);

 glTexCoord2f((float)width/1024.0, (float)height/1024.0);
 glVertex2i(1280, 1024);

 glTexCoord2f((float)width/1024.0, 0.0);
 glVertex2i(1280, 0);
glEnd();
glXSwapBuffers(xDisplay, glxWin);


I'm sure you get the idea... some checking of return vales and liberal use of XFree() might be wise :) Crime can handle the fill rates up to PAL sizes. Should go at about the same speed as using glTexSubImage2D() on every frame in the normal textured-zooming manner. Wouldn't cope with zooming PAL up to 1600x1024, perhaps, that would be about 51MPixels/sec, probably not achievable on mips4 O2s and surely not on the broken 175Mhz R10k.

Enjoy...

[vegac]

hmm...
This is something I'm going to have to get going...
As is, YCrCb pixel format makes YUV conversion VERY fast, but the problem with glDrawPixels is that when scaling, it has to do the conversion per drawn pixel not per source pixel, which makes fullscreen mode a bit slower...

However if I can get this working as you've written, I'd have the faster texture zooming AND the faster conversion on o2...

I'll report back after some tests and benchmarking

[lewis]

Wow man, does it really do the conversion after the zoom? That is rather dumb. Sure it's not just the using non-accelerated zoom factors thing? Is glPixelZoom(2.0, 2.0) fast, for example? It should be, but glPixelZoom(1.5, 1.5) would be done totally in software and would suck... you're allowed factors taken from the set {1, 2, 3, ... 15, 16} or {1/16, 2/16, 3/16, ... 15/16, 1}. So the Tuning part of OpenGL On Silicon Graphics Systems tells me, anyway. Keep up the good work :)

[GeneratriX]

Wouldn't cope with zooming PAL up to 1600x1024, perhaps, that would be about 51MPixels/sec, probably not achievable on mips4 O2s and surely not on the broken 175Mhz R10k.

Enjoy...

...Those thinking on get drived a progressive-scan "Barco" projector at 1600x1024 @ 120fps now are broking their opium pipes!

Now seriously: very interesting post!, I'm arranging/adapting some new C++ modules to test your suggested code on the VideoTrackerRT framework...