diff options
author | Dave Parks <davep@lindenlab.com> | 2011-10-07 03:12:11 -0500 |
---|---|---|
committer | Dave Parks <davep@lindenlab.com> | 2011-10-07 03:12:11 -0500 |
commit | 79912f9d3f0807529183521f69f989f947c1cff1 (patch) | |
tree | eebfd21524f2e70d550397edfccdfcaefd6ab290 /indra | |
parent | c834bdd05a134d6b3442a31f351a94f21965d4e9 (diff) |
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
Diffstat (limited to 'indra')
-rw-r--r-- | indra/llrender/llcubemap.cpp | 6 | ||||
-rw-r--r-- | indra/llrender/llrender.cpp | 252 | ||||
-rw-r--r-- | indra/llrender/llrender.h | 10 | ||||
-rw-r--r-- | indra/newview/llhudrender.cpp | 12 | ||||
-rw-r--r-- | indra/newview/llselectmgr.cpp | 2 | ||||
-rw-r--r-- | indra/newview/llviewercamera.cpp | 58 | ||||
-rw-r--r-- | indra/newview/pipeline.cpp | 28 | ||||
-rw-r--r-- | indra/newview/pipeline.h | 2 |
8 files changed, 213 insertions, 157 deletions
diff --git a/indra/llrender/llcubemap.cpp b/indra/llrender/llcubemap.cpp index a75f01c708..45a3b18179 100644 --- a/indra/llrender/llcubemap.cpp +++ b/indra/llrender/llcubemap.cpp @@ -265,9 +265,9 @@ void LLCubeMap::setMatrix(S32 stage) gGL.getTexUnit(stage)->activate(); } - LLVector3 x(LLVector3d(gGLModelView+0)); - LLVector3 y(LLVector3d(gGLModelView+4)); - LLVector3 z(LLVector3d(gGLModelView+8)); + LLVector3 x(gGLModelView+0); + LLVector3 y(gGLModelView+4); + LLVector3 z(gGLModelView+8); LLMatrix3 mat3; mat3.setRows(x,y,z); diff --git a/indra/llrender/llrender.cpp b/indra/llrender/llrender.cpp index 942ad87566..c5a6859016 100644 --- a/indra/llrender/llrender.cpp +++ b/indra/llrender/llrender.cpp @@ -38,10 +38,10 @@ LLRender gGL; // Handy copies of last good GL matrices -F64 gGLModelView[16]; -F64 gGLLastModelView[16]; -F64 gGLLastProjection[16]; -F64 gGLProjection[16]; +F32 gGLModelView[16]; +F32 gGLLastModelView[16]; +F32 gGLLastProjection[16]; +F32 gGLProjection[16]; S32 gGLViewport[4]; U32 LLRender::sUICalls = 0; @@ -1163,41 +1163,41 @@ void LLRender::syncMatrices() LLGLSLShader* shader = LLGLSLShader::sCurBoundShaderPtr; + static glh::matrix4f cached_mvp; + static U32 cached_mvp_mdv_hash = 0xFFFFFFFF; + static U32 cached_mvp_proj_hash = 0xFFFFFFFF; + + static glh::matrix4f cached_normal; + static U32 cached_normal_hash = 0xFFFFFFFF; + if (shader) { - llassert(shader); bool do_normal = false; - bool do_mvp = false; + bool do_mvp = false; - for (U32 i = 0; i < NUM_MATRIX_MODES; ++i) - { - if (mMatHash[i] != shader->mMatHash[i]) - { - shader->uniformMatrix4fv(name[i], 1, GL_FALSE, mMatrix[i][mMatIdx[i]].m); - shader->mMatHash[i] = mMatHash[i]; + bool mvp_done = false; - if (i == MM_MODELVIEW) - { - do_normal = true; - do_mvp = true; - } - else if (i == MM_PROJECTION) - { - do_mvp = true; - } - } - } + U32 i = MM_MODELVIEW; + if (mMatHash[i] != shader->mMatHash[i]) + { //update modelview, normal, and MVP + glh::matrix4f& mat = mMatrix[i][mMatIdx[i]]; - if (do_normal) - { + shader->uniformMatrix4fv(name[i], 1, GL_FALSE, mat.m); + shader->mMatHash[i] = mMatHash[i]; + + //update normal matrix S32 loc = shader->getUniformLocation("normal_matrix"); if (loc > -1) { - U32 i = MM_MODELVIEW; + if (cached_normal_hash != mMatHash[i]) + { + cached_normal = mat.inverse().transpose(); + cached_normal_hash = mMatHash[i]; + } - glh::matrix4f norm = mMatrix[i][mMatIdx[i]].inverse().transpose(); + glh::matrix4f& norm = cached_normal; F32 norm_mat[] = { @@ -1208,23 +1208,65 @@ void LLRender::syncMatrices() shader->uniformMatrix3fv("normal_matrix", 1, GL_FALSE, norm_mat); } - } - if (do_mvp) - { - S32 loc = shader->getUniformLocation("modelview_projection_matrix"); + //update MVP matrix + mvp_done = true; + loc = shader->getUniformLocation("modelview_projection_matrix"); if (loc > -1) { - U32 mv = MM_MODELVIEW; U32 proj = MM_PROJECTION; - glh::matrix4f mvp = mMatrix[mv][mMatIdx[mv]]; - mvp.mult_left(mMatrix[proj][mMatIdx[proj]]); - - shader->uniformMatrix4fv("modelview_projection_matrix", 1, GL_FALSE, mvp.m); + if (cached_mvp_mdv_hash != mMatHash[i] || cached_mvp_proj_hash != mMatHash[MM_PROJECTION]) + { + cached_mvp = mat; + cached_mvp.mult_left(mMatrix[proj][mMatIdx[proj]]); + cached_mvp_mdv_hash = mMatHash[i]; + cached_mvp_proj_hash = mMatHash[MM_PROJECTION]; + } + + shader->uniformMatrix4fv("modelview_projection_matrix", 1, GL_FALSE, cached_mvp.m); + } + } + + + i = MM_PROJECTION; + if (mMatHash[i] != shader->mMatHash[i]) + { //update projection matrix, normal, and MVP + glh::matrix4f& mat = mMatrix[i][mMatIdx[i]]; + + shader->uniformMatrix4fv(name[i], 1, GL_FALSE, mat.m); + shader->mMatHash[i] = mMatHash[i]; + + if (!mvp_done) + { + //update MVP matrix + S32 loc = shader->getUniformLocation("modelview_projection_matrix"); + if (loc > -1) + { + if (cached_mvp_mdv_hash != mMatHash[i] || cached_mvp_proj_hash != mMatHash[MM_PROJECTION]) + { + U32 mdv = MM_MODELVIEW; + cached_mvp = mat; + cached_mvp.mult_right(mMatrix[mdv][mMatIdx[mdv]]); + cached_mvp_mdv_hash = mMatHash[MM_MODELVIEW]; + cached_mvp_proj_hash = mMatHash[MM_PROJECTION]; + } + + shader->uniformMatrix4fv("modelview_projection_matrix", 1, GL_FALSE, cached_mvp.m); + } } } + for (i = MM_TEXTURE0; i < NUM_MATRIX_MODES; ++i) + { + if (mMatHash[i] != shader->mMatHash[i]) + { + shader->uniformMatrix4fv(name[i], 1, GL_FALSE, mMatrix[i][mMatIdx[i]].m); + shader->mMatHash[i] = mMatHash[i]; + } + } + + if (shader->mFeatures.hasLighting || shader->mFeatures.calculatesLighting) { //also sync light state syncLightState(); @@ -1261,116 +1303,121 @@ void LLRender::translatef(const GLfloat& x, const GLfloat& y, const GLfloat& z) { flush(); - glh::matrix4f trans_mat(1,0,0,x, - 0,1,0,y, - 0,0,1,z, - 0,0,0,1); + { + glh::matrix4f trans_mat(1,0,0,x, + 0,1,0,y, + 0,0,1,z, + 0,0,0,1); - mMatrix[mMatrixMode][mMatIdx[mMatrixMode]].mult_right(trans_mat); - mMatHash[mMatrixMode]++; + mMatrix[mMatrixMode][mMatIdx[mMatrixMode]].mult_right(trans_mat); + mMatHash[mMatrixMode]++; + } } void LLRender::scalef(const GLfloat& x, const GLfloat& y, const GLfloat& z) { flush(); - glh::matrix4f scale_mat(x,0,0,0, - 0,y,0,0, - 0,0,z,0, - 0,0,0,1); + { + glh::matrix4f scale_mat(x,0,0,0, + 0,y,0,0, + 0,0,z,0, + 0,0,0,1); - mMatrix[mMatrixMode][mMatIdx[mMatrixMode]].mult_right(scale_mat); - mMatHash[mMatrixMode]++; + mMatrix[mMatrixMode][mMatIdx[mMatrixMode]].mult_right(scale_mat); + mMatHash[mMatrixMode]++; + } } void LLRender::ortho(F32 left, F32 right, F32 bottom, F32 top, F32 zNear, F32 zFar) { flush(); - glh::matrix4f ortho_mat(2.f/(right-left),0,0, -(right+left)/(right-left), - 0,2.f/(top-bottom),0, -(top+bottom)/(top-bottom), - 0,0,-2.f/(zFar-zNear), -(zFar+zNear)/(zFar-zNear), - 0,0,0,1); + { + + glh::matrix4f ortho_mat(2.f/(right-left),0,0, -(right+left)/(right-left), + 0,2.f/(top-bottom),0, -(top+bottom)/(top-bottom), + 0,0,-2.f/(zFar-zNear), -(zFar+zNear)/(zFar-zNear), + 0,0,0,1); - mMatrix[mMatrixMode][mMatIdx[mMatrixMode]].mult_right(ortho_mat); - mMatHash[mMatrixMode]++; + mMatrix[mMatrixMode][mMatIdx[mMatrixMode]].mult_right(ortho_mat); + mMatHash[mMatrixMode]++; + } } void LLRender::rotatef(const GLfloat& a, const GLfloat& x, const GLfloat& y, const GLfloat& z) { flush(); - - F32 r = a * DEG_TO_RAD; - F32 c = cosf(r); - F32 s = sinf(r); + { + F32 r = a * DEG_TO_RAD; + + F32 c = cosf(r); + F32 s = sinf(r); - F32 ic = 1.f-c; + F32 ic = 1.f-c; - glh::matrix4f rot_mat(x*x*ic+c, x*y*ic-z*s, x*z*ic+y*s, 0, - x*y*ic+z*s, y*y*ic+c, y*z*ic-x*s, 0, - x*z*ic-y*s, y*z*ic+x*s, z*z*ic+c, 0, - 0,0,0,1); + glh::matrix4f rot_mat(x*x*ic+c, x*y*ic-z*s, x*z*ic+y*s, 0, + x*y*ic+z*s, y*y*ic+c, y*z*ic-x*s, 0, + x*z*ic-y*s, y*z*ic+x*s, z*z*ic+c, 0, + 0,0,0,1); - mMatrix[mMatrixMode][mMatIdx[mMatrixMode]].mult_right(rot_mat); - mMatHash[mMatrixMode]++; + mMatrix[mMatrixMode][mMatIdx[mMatrixMode]].mult_right(rot_mat); + mMatHash[mMatrixMode]++; + } } void LLRender::pushMatrix() { flush(); - if (mMatIdx[mMatrixMode] < LL_MATRIX_STACK_DEPTH-1) { - mMatrix[mMatrixMode][mMatIdx[mMatrixMode]+1] = mMatrix[mMatrixMode][mMatIdx[mMatrixMode]]; - ++mMatIdx[mMatrixMode]; - } - else - { - llwarns << "Matrix stack overflow." << llendl; + if (mMatIdx[mMatrixMode] < LL_MATRIX_STACK_DEPTH-1) + { + mMatrix[mMatrixMode][mMatIdx[mMatrixMode]+1] = mMatrix[mMatrixMode][mMatIdx[mMatrixMode]]; + ++mMatIdx[mMatrixMode]; + } + else + { + llwarns << "Matrix stack overflow." << llendl; + } } } void LLRender::popMatrix() { flush(); - if (mMatIdx[mMatrixMode] > 0) - { - --mMatIdx[mMatrixMode]; - mMatHash[mMatrixMode]++; - } - else { - llwarns << "Matrix stack underflow." << llendl; + if (mMatIdx[mMatrixMode] > 0) + { + --mMatIdx[mMatrixMode]; + mMatHash[mMatrixMode]++; + } + else + { + llwarns << "Matrix stack underflow." << llendl; + } } } void LLRender::loadMatrix(const GLfloat* m) { flush(); - mMatrix[mMatrixMode][mMatIdx[mMatrixMode]].set_value((GLfloat*) m); - mMatHash[mMatrixMode]++; -} - -void LLRender::loadMatrix(const GLdouble* dm) -{ - F32 m[16]; - for (U32 i = 0; i < 16; i++) { - m[i] = (F32) dm[i]; + mMatrix[mMatrixMode][mMatIdx[mMatrixMode]].set_value((GLfloat*) m); + mMatHash[mMatrixMode]++; } - - loadMatrix(m); } void LLRender::multMatrix(const GLfloat* m) { flush(); - - glh::matrix4f mat((GLfloat*) m); + { + glh::matrix4f mat((GLfloat*) m); - mMatrix[mMatrixMode][mMatIdx[mMatrixMode]].mult_right(mat); - mMatHash[mMatrixMode]++; + mMatrix[mMatrixMode][mMatIdx[mMatrixMode]].mult_right(mat); + mMatHash[mMatrixMode]++; + } } void LLRender::matrixMode(U32 mode) @@ -1384,23 +1431,14 @@ void LLRender::matrixMode(U32 mode) mMatrixMode = mode; } -void LLRender::multMatrix(const GLdouble* dm) -{ - F32 m[16]; - for (U32 i = 0; i < 16; i++) - { - m[i] = (F32) dm[i]; - } - - multMatrix(m); -} - void LLRender::loadIdentity() { flush(); - mMatrix[mMatrixMode][mMatIdx[mMatrixMode]].make_identity(); - mMatHash[mMatrixMode]++; + { + mMatrix[mMatrixMode][mMatIdx[mMatrixMode]].make_identity(); + mMatHash[mMatrixMode]++; + } } const glh::matrix4f& LLRender::getModelviewMatrix() diff --git a/indra/llrender/llrender.h b/indra/llrender/llrender.h index 61e503d384..7581b9f908 100644 --- a/indra/llrender/llrender.h +++ b/indra/llrender/llrender.h @@ -343,10 +343,8 @@ public: void pushMatrix(); void popMatrix(); void loadMatrix(const GLfloat* m); - void loadMatrix(const GLdouble* m); void loadIdentity(); void multMatrix(const GLfloat* m); - void multMatrix(const GLdouble* m); void matrixMode(U32 mode); const glh::matrix4f& getModelviewMatrix(); @@ -470,10 +468,10 @@ private: }; -extern F64 gGLModelView[16]; -extern F64 gGLLastModelView[16]; -extern F64 gGLLastProjection[16]; -extern F64 gGLProjection[16]; +extern F32 gGLModelView[16]; +extern F32 gGLLastModelView[16]; +extern F32 gGLLastProjection[16]; +extern F32 gGLProjection[16]; extern S32 gGLViewport[4]; extern LLRender gGL; 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 ¢er, 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(); |