From f5e4a85d6157f9e889fc9c85794379899ae4e1a8 Mon Sep 17 00:00:00 2001
From: Marten Richter <marten.richter@freenet.de>
Date: Sun, 8 Jul 2012 16:28:51 +0200
Subject: [PATCH] MoComp nearly bug free
---
glmocoshader.cc | 118 ++++++++++++++++++++++++++++++++++++++----------
osdopengl.cc | 2 +-
vfeed.cc | 2 +-
videovpeogl.cc | 16 +++++--
4 files changed, 109 insertions(+), 29 deletions(-)
mode change 100644 => 100755 vfeed.cc
diff --git a/glmocoshader.cc b/glmocoshader.cc
index 5f0cd6c..cbeabce 100755
--- a/glmocoshader.cc
+++ b/glmocoshader.cc
@@ -72,19 +72,19 @@ const GLchar moco_vertex_shader[] =
"void main()\n"
"{\n"
// calculate cbp
- " vec4 cbp_vec=vec4(cbp,cbp,cbp,cbp);\n"
+ //" vec4 cbp_vec=vec4(cbp,cbp,cbp,cbp);\n"
// old cbp code
// " vec4 fmod_cbp_y=mod(cbp_vec,y_cbp);\n"
- // " vec2 fmod_cbp_uv=mod(cbp_vec.xy,uv_cbp);\n"
+ // " vec2 fmod_cbp_uv=mod(cbp_vec.xy,uv_cbp.xy);\n"
// " fmod_cbp_y.x=0.;\n"
// " fmod_cbp_y=sign(vec4(fmod_cbp_y.yzw,fmod_cbp_uv.x)-fmod_cbp_y);\n"
// " fmod_cbp_uv=sign(vec2(fmod_cbp_uv.y,cbp)-fmod_cbp_uv);\n"
// resulting vector should indicate with block is present
// now transform it to a sum (TODO vectorize this)
-
- " vec4 fmod_cbp_y=mod(cbp_vec,vec4(uv_cbp.zyx,y_cbp.w));\n"
+//this code is better but nor really working
+ /* " vec4 fmod_cbp_y=mod(cbp_vec,vec4(uv_cbp.zyx,y_cbp.w));\n"
" vec2 fmod_cbp_uv=mod(cbp_vec.xy,y_cbp.zy);\n"
" fmod_cbp_y=vec4(lessThanEqual(vec4(uv_cbp.yx,y_cbp.wz),fmod_cbp_y));\n"
" fmod_cbp_uv=vec2(lessThanEqual(y_cbp.yx,fmod_cbp_uv));\n"
@@ -95,16 +95,63 @@ const GLchar moco_vertex_shader[] =
" fmod_cbp_y2.z+=fmod_cbp_y2.y;\n"
" fmod_cbp_y2.w+=fmod_cbp_y2.z;\n"
" fmod_cbp_uv2.x+=fmod_cbp_y2.w;\n"
- " fmod_cbp_uv2.y+=fmod_cbp_uv2.x;\n"
+ " fmod_cbp_uv2.y+=fmod_cbp_uv2.x;\n"*/
+
+
+ " float tcbp=cbp;\n"
+ " float tpos=czero;\n"
+ " vec4 fmod_cbp_y=vec4(czero,czero,czero,czero);\n"
+ " vec2 fmod_cbp_uv=vec2(czero,czero);\n"
+ " vec4 fmod_cbp_y2=vec4(czero,czero,czero,czero);\n"
+ " vec2 fmod_cbp_uv2=vec2(czero,czero);\n"
+ " if (tcbp>=uv_cbp.y){\n"
+ " fmod_cbp_y.x=one;\n"
+ " fmod_cbp_y2.x=tpos;\n"
+ " tpos++;\n"
+ " tcbp-=uv_cbp.y;\n"
+ " }\n"
+ " if (tcbp>=uv_cbp.x){\n"
+ " fmod_cbp_y.y=one;\n"
+ " fmod_cbp_y2.y=tpos;\n"
+ " tpos++;\n"
+ " tcbp-=uv_cbp.x;\n"
+ " }\n"
+ " if (tcbp>=y_cbp.w){\n"
+ " fmod_cbp_y.z=one;\n"
+ " fmod_cbp_y2.z=tpos;\n"
+ " tpos++;\n"
+ " tcbp-=y_cbp.w;\n"
+ " }\n"
+ " if (tcbp>=y_cbp.z){\n"
+ " fmod_cbp_y.w=one;\n"
+ " fmod_cbp_y2.w=tpos;\n"
+ " tpos++;\n"
+ " tcbp-=y_cbp.z;\n"
+ " }\n"
+ " if (tcbp>=y_cbp.y){\n"
+ " fmod_cbp_uv.x=one;\n"
+ " fmod_cbp_uv2.x=tpos;\n"
+ " tpos++;\n"
+ " tcbp-=y_cbp.y;\n"
+ " }\n"
+ " if (tcbp>=y_cbp.x){\n"
+ " fmod_cbp_uv.y=one;\n"
+ " fmod_cbp_uv2.y=tpos;\n"
+ // " tpos++;\n"
+ // " tcbp-=y_cbp.y;\n"
+ " }\n"
+
+
+
+
+
+
+
// now calculate their position inside the short array
- " float m_index=index.x/*+index.y*c65536*/-one;\n" //Not Endian save
+ " float m_index=index.x/*+index.y*c65536*/;\n" //Not Endian save
" fmod_cbp_y2+=m_index;\n" //shift to the blocks
" fmod_cbp_uv2+=m_index;\n"
- " out_yblockpos_x=(mod(fmod_cbp_y2,c32));\n"
- " out_uvblockpos_xy.xy=(mod(fmod_cbp_uv2,c32));\n"
- " out_yblockpos_y=floor((fmod_cbp_y2)/c32);\n"
- " out_uvblockpos_xy.zw=floor((fmod_cbp_uv2)/c32);\n"
" out_yblockpos_y=floor((fmod_cbp_y2)*c1over32);\n"
" out_uvblockpos_xy.zw=floor((fmod_cbp_uv2)*c1over32);\n"
@@ -117,14 +164,14 @@ const GLchar moco_vertex_shader[] =
// " out_yblockpos_y-=sign(fmod_cbp_y-one)*c1024;\n"
// " out_uvblockpos_xy.zw-=sign(fmod_cbp_uv-one)*c1024;\n"
- " out_yblockpos_y=mix(vec4(c1024,c1024,c1024,c1024)-one,out_yblockpos_y,fmod_cbp_y);\n"
- " out_uvblockpos_xy.zw=mix(vec2(c1024,c1024)-one,out_uvblockpos_xy.zw,fmod_cbp_uv);\n"
+ " out_yblockpos_y=mix(vec4(c1024,c1024,c1024,c1024),out_yblockpos_y,fmod_cbp_y);\n"
+ " out_uvblockpos_xy.zw=mix(vec2(c1024,c1024),out_uvblockpos_xy.zw,fmod_cbp_uv);\n"
// set start positions
- " vec4 motion_vertical=vec4(block_types.z,block_types.z,block_types.z,block_types.z);\n"
- " vec4 mod_motion_vertical=mod(motion_vertical,y_cbp);\n"
- " mod_motion_vertical=sign(vec4(mod_motion_vertical.yzw,block_types.z)-mod_motion_vertical);\n"
+ // " vec4 motion_vertical=vec4(block_types.z,block_types.z,block_types.z,block_types.z);\n"
+ // " vec4 mod_motion_vertical=mod(motion_vertical,vec4(y_cbp.yzw,uv_cbp.x));\n"
+ // " mod_motion_vertical=vec4(lessThanEqual(y_cbp,mod_motion_vertical));\n"
" out_moco_pos.xy=(block_pos.xy+block_edge.xy)*uv_cbp.x/*16*/+one;"
" out_unsupported=czero;\n"
@@ -150,18 +197,40 @@ const GLchar moco_vertex_shader[] =
" out_moco_backward=vec4(PMV1.zw,PMV1.zw);\n"
//" out_unsupported=2.;\n"
" } else if (block_types.y==y_cbp.x /*1*/) {" //Field
- " out_moco_forward=vec4(PMV1.xy,PMV1.xy);\n"
- " out_moco_backward=vec4(PMV1.zw,PMV1.zw);\n"
- " out_moco_forward.y+=mod_motion_vertical.x;\n"
+ " out_moco_forward=vec4(PMV1.xy,PMV2.xy);\n"
+ " out_moco_backward=vec4(PMV1.zw,PMV2.zw);\n"
+
+ " float motion_vert=block_types.z;\n"
+ " out_moco_forward.w-=one;\n"
+ " out_moco_backward.w-=one;\n"
+ " if (motion_vert>= y_cbp.w) {\n" //second_backward
+ " motion_vert-=y_cbp.w;\n"
+ " out_moco_backward.w+=one;\n"
+ " }\n"
+ " if (motion_vert>= y_cbp.z) {\n" //second_forward
+ " motion_vert-=y_cbp.z;\n"
+ " out_moco_forward.w+=one;\n"
+ " }\n"
+ " if (motion_vert>= y_cbp.y) {\n" //first_backward
+ " motion_vert-=y_cbp.y;\n"
+ " out_moco_backward.y+=one;\n"
+ " }\n"
+ " if (motion_vert>= y_cbp.x) {\n" //second_forward
+ //" motion_vert-=y_cbp.z;\n"
+ " out_moco_forward.y+=one;\n"
+ " }\n"
+ /* " out_moco_forward.y+=mod_motion_vertical.x;\n"
" out_moco_backward.y+=mod_motion_vertical.y;\n"
" out_moco_forward.w-=one-mod_motion_vertical.z;\n"
- " out_moco_backward.w-=one-mod_motion_vertical.w;\n"
+ " out_moco_backward.w-=one-mod_motion_vertical.w;\n"*/
" out_misc_info.y=one;\n" // step to for field based
// " out_unsupported=2.;\n"
" } else {\n"
" out_unsupported=1.;\n"
" }\n"
" }\n"
+ // " if (cbp==63. || cbp==0.) out_unsupported=2.;\n"
+ // " if (fmod_cbp_y.w==1. && mod(cbp,8.)>=4.) out_unsupported=2.;\n"
@@ -461,7 +530,7 @@ int GLMocoShader::uploadDataBlocks(short* blocks,unsigned int num_blocks, XvMCMa
-// Log::getInstance()->log("GLMocoShader", Log::WARN, "uploadDataBlocks mark1 glerror %x %d",glGetError(),height);
+ Log::getInstance()->log("GLMocoShader", Log::WARN, "uploadDataBlocks mark1 %d %d %d",height,num_blocks,num_m_blocks);
glTexSubImage2D(GL_TEXTURE_2D,0,0,0,
BLOCK_TEXTURE_WIDTH>>1,height,
GL_RGBA,GL_UNSIGNED_BYTE,
@@ -484,7 +553,10 @@ int GLMocoShader::uploadDataBlocks(short* blocks,unsigned int num_blocks, XvMCMa
XvMCMacroBlock *m_blocks_end=m_blocks+num_m_blocks;
while (m_blocks_run!=m_blocks_end) {
m_blocks_run->pad0=0xFF00;
+ // Log::getInstance()->log("GLMocoShader", Log::WARN, "uploadDataBlocks coded block %x %x %d",
+ // m_blocks_run->coded_block_pattern, m_blocks_run->macroblock_type,m_blocks_run->index);
m_blocks_run++;
+
}
// debug
/* m_blocks[0].x=0;
@@ -605,7 +677,7 @@ int GLMocoShader::doMoCo(VPEOGLFrame *target,VPEOGLFrame *forward,VPEOGLFrame *b
glVertexAttribPointer(2, 4, GL_UNSIGNED_BYTE, GL_FALSE,sizeof(XvMCMacroBlock), (const void *) &(helperp->macroblock_type));
glEnableVertexAttribArray(2);
- glVertexAttribPointer(3, 1, GL_SHORT, GL_FALSE,sizeof(XvMCMacroBlock), (const void *) &(helperp->coded_block_pattern));
+ glVertexAttribPointer(3, 1, GL_UNSIGNED_SHORT, GL_FALSE,sizeof(XvMCMacroBlock), (const void *) &(helperp->coded_block_pattern));
glEnableVertexAttribArray(3);
glVertexAttribPointer(4, 2, GL_UNSIGNED_SHORT, GL_FALSE,sizeof(XvMCMacroBlock), (const void *) &(helperp->index));
glEnableVertexAttribArray(4);
@@ -630,7 +702,7 @@ int GLMocoShader::doMoCo(VPEOGLFrame *target,VPEOGLFrame *forward,VPEOGLFrame *b
if (forward) {
- Log::getInstance()->log("GLMocoShader", Log::WARN, "doMoCo forward %x",forward->textures[0]);
+ // Log::getInstance()->log("GLMocoShader", Log::WARN, "doMoCo forward %x",forward->textures[0]);
glActiveTexture(GL_TEXTURE1);
glBindTexture(GL_TEXTURE_2D,forward->textures[0]);
glUniform1i(forward_pic_loc,1);
@@ -643,7 +715,7 @@ int GLMocoShader::doMoCo(VPEOGLFrame *target,VPEOGLFrame *forward,VPEOGLFrame *b
if (backward) {
glActiveTexture(GL_TEXTURE2);
- Log::getInstance()->log("GLMocoShader", Log::WARN, "doMoCo backward %x",backward->textures[0]);
+ // Log::getInstance()->log("GLMocoShader", Log::WARN, "doMoCo backward %x",backward->textures[0]);
glBindTexture(GL_TEXTURE_2D,backward->textures[0]);
glUniform1i(backward_pic_loc,2);
glTexParameteri(GL_TEXTURE_2D,GL_TEXTURE_MIN_FILTER,GL_NEAREST);
diff --git a/osdopengl.cc b/osdopengl.cc
index f0bc96b..9ba5a4e 100755
--- a/osdopengl.cc
+++ b/osdopengl.cc
@@ -420,7 +420,7 @@ void OsdOpenGL::threadMethod()
if (!frame) frame=video->getReadyOGLFrame();
if (frame) {
InternalRendering(frame);
- // MILLISLEEP(1000);
+ //MILLISLEEP(1000);
lastrendertime=getTimeMS();
video->returnOGLFrame(frame); //please recycle it
frame=NULL;
diff --git a/vfeed.cc b/vfeed.cc
old mode 100644
new mode 100755
index 6640a13..b0fefab
--- a/vfeed.cc
+++ b/vfeed.cc
@@ -76,7 +76,7 @@ void VFeed::threadMethod()
}
else
{
- //Log::getInstance()->log("VFeed", Log::DEBUG, "No data delay");
+ //Log::getInstance()->log("VFeed", Log::DEBUG, "No data delay FPS");
MILLISLEEP(5);
}
}
diff --git a/videovpeogl.cc b/videovpeogl.cc
index 1035ec2..2e24791 100755
--- a/videovpeogl.cc
+++ b/videovpeogl.cc
@@ -333,7 +333,7 @@ void VideoVPEOGL::recycleOGLRefFrames()
list<VPEOGLFrame*> keep;
list<VPEOGLFrame*>::iterator itty=recycle_ref_ogl_frames.begin();
while (itty!=recycle_ref_ogl_frames.end()) {
- Log::getInstance()->log("Video", Log::WARN, "recycleOGLRefFrame mark3");
+ // Log::getInstance()->log("Video", Log::WARN, "recycleOGLRefFrame mark3");
if ((*itty)->pict_num==ogl_forward_ref_frame_num
|| (*itty)->pict_num==ogl_backward_ref_frame_num)
{
@@ -446,6 +446,7 @@ void VideoVPEOGL::threadMethod()
int width,height,pixfmt;
//First get a free ogl image
VPEOGLFrame* out_frame=NULL;
+ //int msleep=0;
while (!out_frame) {
ogl_frame_mutex.Lock();
// Log::getInstance()->log("Video", Log::WARN, "threadMethod mark upload 1a %d %d %d %d %d",all_ogl_frames.size(),free_ogl_frames.size(),
@@ -460,7 +461,13 @@ void VideoVPEOGL::threadMethod()
pixfmt=libavpixfmt;
out_frame=free_ogl_frames.front();
free_ogl_frames.pop_front();
- } else MILLISLEEP(2);
+ } else {
+ ogl_frame_mutex.Unlock();
+ MILLISLEEP(2);
+ // msleep+=2;
+ ogl_frame_mutex.Lock();
+ }
+ // if (msleep)Log::getInstance()->log("Video", Log::WARN, "msleep FPS %d",msleep);
ogl_frame_mutex.Unlock();
}
bool failed=false;
@@ -658,7 +665,7 @@ void VideoVPEOGL::threadMethod()
target_time.tv_sec+=1;
}
threadWaitForSignalTimed(&target_time);
- //Log::getInstance()->log("Video", Log::WARN, "threadMethod signalled");
+ //Log::getInstance()->log("Video", Log::WARN, "threadMethod signalled FPS");
}
threadCheckExit();
}
@@ -2460,6 +2467,7 @@ int VideoVPEOGL::DecodePacketlibav()
} else {
libav_hastime=false;
dec_frame_libav_mutex.Unlock();
+ Log::getInstance()->log("Video", Log::DEBUG, "We have no free buffers 2 FPS");
threadSignal();
// No free buffers
return 0;
@@ -2508,7 +2516,7 @@ UINT VideoVPEOGL::DeliverMediaPacketlibav(MediaPacket packet,
dec_frame_libav_free.pop_front();
dec_frame_libav_mutex.Unlock();
} else {
- Log::getInstance()->log("Video", Log::DEBUG, "We have no free buffers");
+ Log::getInstance()->log("Video", Log::DEBUG, "We have no free buffers FPS");
dec_frame_libav_mutex.Unlock();
// No free buffers
return 0;
--
2.39.5