SH-2031 Profile guided optimization of matrix ops

- don't use F64 except where needed (should really factor out calls to gluProject)
- get rid of sorting by texture in favor of sorting by matrix (no sort needed, geometry is already matrix sorted as a result of frustum cull tree traversal order)
- unroll matrix sync inner loop and cache MVP and normal matrices

5 files changed, 61 insertions, 41 deletions

diff --git a/indra/newview/llhudrender.cpp b/indra/newview/llhudrender.cpp
index 607f7f7f4b..122711a86d 100644
--- a/indra/newview/llhudrender.cpp
+++ b/indra/newview/llhudrender.cpp
@@ -107,8 +107,18 @@ void hud_render_text(const LLWString &wstr, const LLVector3 &pos_agent,
 	viewport[1] = world_view_rect.mBottom;
 	viewport[2] = world_view_rect.getWidth();
 	viewport[3] = world_view_rect.getHeight();
+
+	F64 mdlv[16];
+	F64 proj[16];
+
+	for (U32 i = 0; i < 16; i++)
+	{
+		mdlv[i] = (F64) gGLModelView[i];
+		proj[i] = (F64) gGLProjection[i];
+	}
+
 	gluProject(render_pos.mV[0], render_pos.mV[1], render_pos.mV[2],
-				gGLModelView, gGLProjection, (GLint*) viewport,
+				mdlv, proj, (GLint*) viewport,
 				&winX, &winY, &winZ);
 		
 	//fonts all render orthographically, set up projection``
diff --git a/indra/newview/llselectmgr.cpp b/indra/newview/llselectmgr.cpp
index 3a50eeb1c9..2971ee710a 100644
--- a/indra/newview/llselectmgr.cpp
+++ b/indra/newview/llselectmgr.cpp
@@ -4838,7 +4838,7 @@ void LLSelectMgr::processForceObjectSelect(LLMessageSystem* msg, void**)
 	LLSelectMgr::getInstance()->highlightObjectAndFamily(objects);
 }
 
-extern LLGLdouble	gGLModelView[16];
+extern F32	gGLModelView[16];
 
 void LLSelectMgr::updateSilhouettes()
 {
diff --git a/indra/newview/llviewercamera.cpp b/indra/newview/llviewercamera.cpp
index c0ef4da1ff..a437a8b3b5 100644
--- a/indra/newview/llviewercamera.cpp
+++ b/indra/newview/llviewercamera.cpp
@@ -218,8 +218,15 @@ void LLViewerCamera::calcProjection(const F32 far_distance) const
 void LLViewerCamera::updateFrustumPlanes(LLCamera& camera, BOOL ortho, BOOL zflip, BOOL no_hacks)
 {
 	GLint* viewport = (GLint*) gGLViewport;
-	GLdouble* model = gGLModelView;
-	GLdouble* proj = gGLProjection;
+	F64 model[16];
+	F64 proj[16];
+
+	for (U32 i = 0; i < 16; i++)
+	{
+		model[i] = (F64) gGLModelView[i];
+		proj[i] = (F64) gGLProjection[i];
+	}
+
 	GLdouble objX,objY,objZ;
 
 	LLVector3 frust[8];
@@ -420,7 +427,6 @@ void LLViewerCamera::setPerspective(BOOL for_selection,
 	if (!for_selection && mZoomFactor == 1.f)
 	{
 		// Save GL matrices for access elsewhere in code, especially project_world_to_screen
-		//glGetDoublev(GL_MODELVIEW_MATRIX, gGLModelView);
 		for (U32 i = 0; i < 16; i++)
 		{
 			gGLModelView[i] = modelview.m[i];
@@ -428,14 +434,6 @@ void LLViewerCamera::setPerspective(BOOL for_selection,
 	}
 
 	updateFrustumPlanes(*this);
-
-	/*if (gSavedSettings.getBOOL("CameraOffset"))
-	{
-		gGL.matrixMode(LLRender::MM_PROJECTION);
-		gGL.translatef(0,0,-50);
-		gGL.rotatef(20.0,1,0,0);
-		gGL.matrixMode(LLRender::MM_MODELVIEW);
-	}*/
 }
 
 
@@ -443,11 +441,20 @@ void LLViewerCamera::setPerspective(BOOL for_selection,
 // screen coordinates to the agent's region.
 void LLViewerCamera::projectScreenToPosAgent(const S32 screen_x, const S32 screen_y, LLVector3* pos_agent) const
 {
-
 	GLdouble x, y, z;
+
+	F64 mdlv[16];
+	F64 proj[16];
+
+	for (U32 i = 0; i < 16; i++)
+	{
+		mdlv[i] = (F64) gGLModelView[i];
+		proj[i] = (F64) gGLProjection[i];
+	}
+
 	gluUnProject(
 		GLdouble(screen_x), GLdouble(screen_y), 0.0,
-		gGLModelView, gGLProjection, (GLint*)gGLViewport,
+		mdlv, proj, (GLint*)gGLViewport,
 		&x,
 		&y,
 		&z );
@@ -484,8 +491,17 @@ BOOL LLViewerCamera::projectPosAgentToScreen(const LLVector3 &pos_agent, LLCoord
 	viewport[2] = world_view_rect.getWidth();
 	viewport[3] = world_view_rect.getHeight();
 
+	F64 mdlv[16];
+	F64 proj[16];
+
+	for (U32 i = 0; i < 16; i++)
+	{
+		mdlv[i] = (F64) gGLModelView[i];
+		proj[i] = (F64) gGLProjection[i];
+	}
+
 	if (GL_TRUE == gluProject(pos_agent.mV[VX], pos_agent.mV[VY], pos_agent.mV[VZ],
-								gGLModelView, gGLProjection, (GLint*)viewport,
+								mdlv, proj, (GLint*)viewport,
 								&x, &y, &z))
 	{
 		// convert screen coordinates to virtual UI coordinates
@@ -587,9 +603,19 @@ BOOL LLViewerCamera::projectPosAgentToScreenEdge(const LLVector3 &pos_agent,
 	viewport[2] = world_view_rect.getWidth();
 	viewport[3] = world_view_rect.getHeight();
 	GLdouble	x, y, z;			// object's window coords, GL-style
+
+	F64 mdlv[16];
+	F64 proj[16];
+
+	for (U32 i = 0; i < 16; i++)
+	{
+		mdlv[i] = (F64) gGLModelView[i];
+		proj[i] = (F64) gGLProjection[i];
+	}
+
 	if (GL_TRUE == gluProject(pos_agent.mV[VX], pos_agent.mV[VY],
-							  pos_agent.mV[VZ], gGLModelView,
-							  gGLProjection, (GLint*)viewport,
+							  pos_agent.mV[VZ], mdlv,
+							  proj, (GLint*)viewport,
 							  &x, &y, &z))
 	{
 		x /= gViewerWindow->getDisplayScale().mV[VX];
diff --git a/indra/newview/pipeline.cpp b/indra/newview/pipeline.cpp
index d28bb1f64e..42873dbca8 100644
--- a/indra/newview/pipeline.cpp
+++ b/indra/newview/pipeline.cpp
@@ -201,13 +201,10 @@ U32 nhpo2(U32 v)
 	return r;
 }
 
-glh::matrix4f glh_copy_matrix(GLdouble* src)
+glh::matrix4f glh_copy_matrix(F32* src)
 {
 	glh::matrix4f ret;
-	for (U32 i = 0; i < 16; i++)
-	{
-		ret.m[i] = (F32) src[i];
-	}
+	ret.set_value(src);
 	return ret;
 }
 
@@ -231,7 +228,7 @@ glh::matrix4f glh_get_last_projection()
 	return glh_copy_matrix(gGLLastProjection);
 }
 
-void glh_copy_matrix(const glh::matrix4f& src, GLdouble* dst)
+void glh_copy_matrix(const glh::matrix4f& src, F32* dst)
 {
 	for (U32 i = 0; i < 16; i++)
 	{
@@ -3171,19 +3168,6 @@ void LLPipeline::postSort(LLCamera& camera)
 		
 	if (!sShadowRender)
 	{
-		//sort by texture or bump map
-		for (U32 i = 0; i < LLRenderPass::NUM_RENDER_TYPES; ++i)
-		{
-			if (i == LLRenderPass::PASS_BUMP)
-			{
-				std::sort(sCull->beginRenderMap(i), sCull->endRenderMap(i), LLDrawInfo::CompareBump());
-			}
-			else 
-			{
-				std::sort(sCull->beginRenderMap(i), sCull->endRenderMap(i), LLDrawInfo::CompareTexturePtrMatrix());
-			}	
-		}
-
 		std::sort(sCull->beginAlphaGroups(), sCull->endAlphaGroups(), LLSpatialGroup::CompareDepthGreater());
 	}
 	llpushcallstacks ;
@@ -3503,8 +3487,8 @@ void LLPipeline::renderGeom(LLCamera& camera, BOOL forceVBOUpdate)
 
 	assertInitialized();
 
-	F64 saved_modelview[16];
-	F64 saved_projection[16];
+	F32 saved_modelview[16];
+	F32 saved_projection[16];
 
 	//HACK: preserve/restore matrices around HUD render
 	if (gPipeline.hasRenderType(LLPipeline::RENDER_TYPE_HUD))
@@ -6768,7 +6752,7 @@ void LLPipeline::bindDeferredShader(LLGLSLShader& shader, U32 light_index, U32 n
 		{
 			cube_map->enable(channel);
 			cube_map->bind();
-			F64* m = gGLModelView;
+			F32* m = gGLModelView;
 
 			
 			F32 mat[] = { m[0], m[1], m[2],
diff --git a/indra/newview/pipeline.h b/indra/newview/pipeline.h
index d24bab747b..86579261ca 100644
--- a/indra/newview/pipeline.h
+++ b/indra/newview/pipeline.h
@@ -72,7 +72,7 @@ BOOL compute_min_max(LLMatrix4& box, LLVector2& min, LLVector2& max); // Shouldn
 bool LLRayAABB(const LLVector3 &center, const LLVector3 &size, const LLVector3& origin, const LLVector3& dir, LLVector3 &coord, F32 epsilon = 0);
 BOOL setup_hud_matrices(); // use whole screen to render hud
 BOOL setup_hud_matrices(const LLRect& screen_region); // specify portion of screen (in pixels) to render hud attachments from (for picking)
-glh::matrix4f glh_copy_matrix(GLdouble* src);
+glh::matrix4f glh_copy_matrix(F32* src);
 glh::matrix4f glh_get_current_modelview();
 void glh_set_current_modelview(const glh::matrix4f& mat);
 glh::matrix4f glh_get_current_projection();