From 2e3d77e8f2b78e1570c214857dd013ae27c78eed Mon Sep 17 00:00:00 2001
From: twinaphex <autechre1024@hotmail.com>
Date: Thu, 28 Mar 2013 16:21:32 +0100
Subject: [PATCH] (RGL PS3) Optimize FIFO out of space callback

---
 ps3/rgl/src/ps3/include/rgl-externs.h |  1 -
 ps3/rgl/src/ps3/rgl_ps3.cpp           | 28 +++------------------------
 ps3/rgl/src/ps3/rgl_ps3_raster.cpp    | 14 ++++++++++++--
 3 files changed, 15 insertions(+), 28 deletions(-)

diff --git a/ps3/rgl/src/ps3/include/rgl-externs.h b/ps3/rgl/src/ps3/include/rgl-externs.h
index 1465905a26..e8046fe583 100644
--- a/ps3/rgl/src/ps3/include/rgl-externs.h
+++ b/ps3/rgl/src/ps3/include/rgl-externs.h
@@ -2,7 +2,6 @@ void rglGcmFifoFinish (void *data);
 GLboolean rglGcmFifoReferenceInUse (void *data, GLuint reference);
 GLuint rglGcmFifoPutReference (void *data);
 void rglGcmFifoFlush (void *data);
-uint32_t *rglGcmFifoWaitForFreeSpace (void *data, GLuint spaceInWords);
 void rglGcmGetTileRegionInfo (void *data, GLuint *address, GLuint *size);
 GLboolean rglGcmTryResizeTileRegion( GLuint address, GLuint size, void *data);
 
diff --git a/ps3/rgl/src/ps3/rgl_ps3.cpp b/ps3/rgl/src/ps3/rgl_ps3.cpp
index 7f74f1ecf6..0961fcbc8e 100644
--- a/ps3/rgl/src/ps3/rgl_ps3.cpp
+++ b/ps3/rgl/src/ps3/rgl_ps3.cpp
@@ -1500,19 +1500,6 @@ GLboolean rglGcmFifoReferenceInUse (void *data, GLuint reference)
    return GL_TRUE;
 }
 
-// Wait until the requested space is available.
-// If not currently available, will call the out of space callback
-
-uint32_t * rglGcmFifoWaitForFreeSpace (void *data, GLuint spaceInWords)
-{
-   rglGcmFifo *fifo = (rglGcmFifo*)data;
-
-   if ( fifo->current + spaceInWords + 1024 > fifo->end )
-      rglOutOfSpaceCallback( fifo, spaceInWords );
-
-   return rglGcmState_i.fifo.current;
-}
-
 void rglGcmFifoInit (void *data, void *dmaControl, unsigned long dmaPushBufferOffset, uint32_t*dmaPushBuffer,
       GLuint dmaPushBufferSize )
 {
@@ -2065,15 +2052,10 @@ int32_t rglOutOfSpaceCallback (void *data, uint32_t spaceInWords)
    // If the current end isn't the same as the full fifo end we 
    // aren't at the end.  Just go ahead and set the next begin and end 
    if(fifo->end != fifo->dmaPushBufferEnd)
-   {
       nextbegin = (uint32_t *)fifo->end + 1; 
-      nextend = nextbegin + fifo->fifoBlockSize/sizeof(uint32_t) - 1;
-   }
    else
-   {
       nextbegin = (uint32_t *)fifo->dmaPushBufferBegin;
-      nextend = nextbegin + (fifo->fifoBlockSize)/sizeof(uint32_t) - 1;
-   }
+   nextend = nextbegin + (fifo->fifoBlockSize)/sizeof(uint32_t) - 1;
 
    cellGcmAddressToOffset(nextbegin, &nextbeginoffset);
    cellGcmAddressToOffset(nextend, &nextendoffset);
@@ -2087,8 +2069,6 @@ int32_t rglOutOfSpaceCallback (void *data, uint32_t spaceInWords)
    fifo->current = nextbegin;
    fifo->end = nextend;
 
-   const GLuint nopsAtBegin = 8;
-
    //if Gpu busy with the new area, stall and flush
    uint32_t get = fifo->dmaControl->Get;
 
@@ -2106,17 +2086,15 @@ int32_t rglOutOfSpaceCallback (void *data, uint32_t spaceInWords)
    while(((get >= nextbeginoffset) && (get <= nextendoffset)) 
          || ((0 <= get) && (get < 0x10000))) 
    {
-      // Don't be a ppu hog ;)
-      sys_timer_usleep(30);
       get = fifo->dmaControl->Get;
       cellGcmIoOffsetToAddress( get, &getEA ); 
    }
 
    // need to add some nops here at the beginning for a issue with the get and the put being at the 
    // same position when the fifo is in GPU memory. 
-   for ( GLuint i = 0; i < nopsAtBegin; i++ )
+   for ( GLuint i = 0; i < 8; i++ )
    {
-      fifo->current[0] = RGLGCM_NOP();
+      fifo->current[0] = 0;
       fifo->current++;
    }
 
diff --git a/ps3/rgl/src/ps3/rgl_ps3_raster.cpp b/ps3/rgl/src/ps3/rgl_ps3_raster.cpp
index c1b891d092..b557577457 100644
--- a/ps3/rgl/src/ps3/rgl_ps3_raster.cpp
+++ b/ps3/rgl/src/ps3/rgl_ps3_raster.cpp
@@ -1766,7 +1766,14 @@ static inline void rglValidateStates (GLuint mask)
       conf.registerCount = program->header.vertexProgram.registerCount;
       conf.attributeInputMask = program->header.attributeInputMask;
 
-      rglGcmFifoWaitForFreeSpace( &rglGcmState_i.fifo, 7 + 5 * conf.instructionCount );
+      rglGcmFifo *fifo = (rglGcmFifo*)&rglGcmState_i.fifo;
+      GLuint spaceInWords = 7 + 5 * conf.instructionCount;
+
+      // Push a CG program onto the current command buffer
+
+      // make sure there is space for the pushbuffer + any nops we need to add for alignment  
+      if ( fifo->current + spaceInWords + 1024 > fifo->end )
+         rglOutOfSpaceCallback( fifo, spaceInWords );
 
       GCM_FUNC( cellGcmSetVertexProgramLoad, &conf, program->ucode );
       GCM_FUNC( cellGcmSetUserClipPlaneControl, 0, 0, 0, 0, 0, 0 );
@@ -1896,11 +1903,14 @@ static inline void rglValidateStates (GLuint mask)
    if (RGL_LIKELY(needValidate & RGL_VALIDATE_VERTEX_CONSTANTS) || validate_vertex_consts)
    {
       _CGprogram *cgprog = LContext->BoundVertexProgram;
+      rglGcmFifo *fifo = (rglGcmFifo*)&rglGcmState_i.fifo;
+      GLuint spaceInWords = cgprog->constantPushBufferWordSize + 4 + 32;
 
       // Push a CG program onto the current command buffer
 
       // make sure there is space for the pushbuffer + any nops we need to add for alignment  
-      rglGcmFifoWaitForFreeSpace( &rglGcmState_i.fifo,  cgprog->constantPushBufferWordSize + 4 + 32); 
+      if ( fifo->current + spaceInWords + 1024 > fifo->end )
+         rglOutOfSpaceCallback( fifo, spaceInWords );
 
       // first add nops to get us the next alligned position in the fifo 
       // [YLIN] Use VMX register to copy