From 6c59ca80ead7210c58c96234fb1f2a951335266e Mon Sep 17 00:00:00 2001
From: Rye Mutt <rye@alchemyviewer.org>
Date: Sun, 31 Oct 2021 16:59:10 -0400
Subject: [PATCH] Optimize setupSpotlight
---
indra/llmath/alglmath.h | 77 ++++++++++++++++++++++++++++++++++++++
indra/newview/pipeline.cpp | 50 +++++++++++++------------
2 files changed, 104 insertions(+), 23 deletions(-)
diff --git a/indra/llmath/alglmath.h b/indra/llmath/alglmath.h
index 0ad8cb30756..8f63c276f3a 100644
--- a/indra/llmath/alglmath.h
+++ b/indra/llmath/alglmath.h
@@ -7,6 +7,12 @@
namespace ALGLMath
{
+ inline static const LLMatrix4a TRANS_MAT = LLMatrix4a(
+ LLVector4a(.5f, 0, 0, 0),
+ LLVector4a(0, .5f, 0, 0),
+ LLVector4a(0, 0, .5f, 0),
+ LLVector4a(.5f, .5f, .5f, 1.f));
+
inline LLMatrix4a genRot(const float a, const LLVector4a& axis)
{
F32 r = a * DEG_TO_RAD;
@@ -44,6 +50,77 @@ namespace ALGLMath
inline LLMatrix4a genRot(const float a, const float x, const float y, const float z) { return genRot(a, LLVector4a(x, y, z)); }
+ inline LLMatrix4a genOrtho(const GLfloat& left, const GLfloat& right, const GLfloat& bottom, const GLfloat& top, const GLfloat& zNear, const GLfloat& zFar)
+ {
+ LLMatrix4a ortho_mat;
+ ortho_mat.setRow<0>(LLVector4a(2.f / (right - left), 0, 0));
+ ortho_mat.setRow<1>(LLVector4a(0, 2.f / (top - bottom), 0));
+ ortho_mat.setRow<2>(LLVector4a(0, 0, -2.f / (zFar - zNear)));
+ ortho_mat.setRow<3>(LLVector4a(-(right + left) / (right - left), -(top + bottom) / (top - bottom), -(zFar + zNear) / (zFar - zNear), 1));
+
+ return ortho_mat;
+ }
+
+ inline LLMatrix4a genPersp(const GLfloat& fovy, const GLfloat& aspect, const GLfloat& zNear, const GLfloat& zFar)
+ {
+ GLfloat f = 1.f / tanf(DEG_TO_RAD * fovy / 2.f);
+
+ LLMatrix4a persp_mat;
+ persp_mat.setRow<0>(LLVector4a(f / aspect, 0, 0));
+ persp_mat.setRow<1>(LLVector4a(0, f, 0));
+ persp_mat.setRow<2>(LLVector4a(0, 0, (zFar + zNear) / (zNear - zFar), -1.f));
+ persp_mat.setRow<3>(LLVector4a(0, 0, (2.f * zFar * zNear) / (zNear - zFar), 0));
+
+ return persp_mat;
+ }
+
+ inline LLMatrix4a genLook(const LLVector3& pos_in, const LLVector3& dir_in, const LLVector3& up_in)
+ {
+ const LLVector4a pos(pos_in.mV[VX], pos_in.mV[VY], pos_in.mV[VZ], 1.f);
+ LLVector4a dir(dir_in.mV[VX], dir_in.mV[VY], dir_in.mV[VZ]);
+ const LLVector4a up(up_in.mV[VX], up_in.mV[VY], up_in.mV[VZ]);
+
+ LLVector4a left_norm;
+ left_norm.setCross3(dir, up);
+ left_norm.normalize3fast();
+ LLVector4a up_norm;
+ up_norm.setCross3(left_norm, dir);
+ up_norm.normalize3fast();
+ LLVector4a& dir_norm = dir;
+ dir.normalize3fast();
+
+ LLVector4a left_dot;
+ left_dot.setAllDot3(left_norm, pos);
+ left_dot.negate();
+ LLVector4a up_dot;
+ up_dot.setAllDot3(up_norm, pos);
+ up_dot.negate();
+ LLVector4a dir_dot;
+ dir_dot.setAllDot3(dir_norm, pos);
+
+ dir_norm.negate();
+
+ LLMatrix4a lookat_mat;
+ lookat_mat.setRow<0>(left_norm);
+ lookat_mat.setRow<1>(up_norm);
+ lookat_mat.setRow<2>(dir_norm);
+ lookat_mat.setRow<3>(LLVector4a(0, 0, 0, 1));
+
+ lookat_mat.getRow<0>().copyComponent<3>(left_dot);
+ lookat_mat.getRow<1>().copyComponent<3>(up_dot);
+ lookat_mat.getRow<2>().copyComponent<3>(dir_dot);
+
+ lookat_mat.transpose();
+
+ return lookat_mat;
+ }
+
+ inline const LLMatrix4a& genNDCtoWC()
+ {
+ return TRANS_MAT;
+ }
+
+
inline bool projectf(const LLVector3& object, const LLMatrix4a& modelview, const LLMatrix4a& projection, const LLRect& viewport, LLVector3& windowCoordinate)
{
//Begin SSE intrinsics
diff --git a/indra/newview/pipeline.cpp b/indra/newview/pipeline.cpp
index 7045e6d6b1f..ee648d2ae70 100644
--- a/indra/newview/pipeline.cpp
+++ b/indra/newview/pipeline.cpp
@@ -29,6 +29,7 @@
#include "pipeline.h"
// library includes
+#include "alglmath.h"
#include "llaudioengine.h" // For debugging.
#include "llerror.h"
#include "llviewercontrol.h"
@@ -9376,12 +9377,17 @@ void LLPipeline::setupSpotLight(LLGLSLShader& shader, LLDrawable* drawablep)
LLVector3 origin = np - at_axis*dist;
//matrix from volume space to agent space
- LLMatrix4 light_mat(quat, LLVector4(origin,1.f));
+ LLMatrix4 light_mat_(quat, LLVector4(origin,1.f));
- glh::matrix4f light_to_agent((F32*) light_mat.mMatrix);
- glh::matrix4f light_to_screen = get_current_modelview() * light_to_agent;
+ LLMatrix4a cur_modelview;
+ cur_modelview.loadu(gGLModelView);
- glh::matrix4f screen_to_light = light_to_screen.inverse();
+ LLMatrix4a light_mat;
+ light_mat.loadu(light_mat_.mMatrix[0]);
+ LLMatrix4a light_to_screen;
+ light_to_screen.setMul(cur_modelview,light_mat);
+ LLMatrix4a screen_to_light = light_to_screen;
+ screen_to_light.invert();
F32 s = volume->getLightRadius()*1.5f;
F32 near_clip = dist;
@@ -9392,30 +9398,28 @@ void LLPipeline::setupSpotLight(LLGLSLShader& shader, LLDrawable* drawablep)
F32 fovy = fov * RAD_TO_DEG;
F32 aspect = width/height;
- glh::matrix4f trans(0.5f, 0.f, 0.f, 0.5f,
- 0.f, 0.5f, 0.f, 0.5f,
- 0.f, 0.f, 0.5f, 0.5f,
- 0.f, 0.f, 0.f, 1.f);
+ LLVector4a p1(0, 0, -(near_clip+0.01f));
+ LLVector4a p2(0, 0, -(near_clip+1.f));
- glh::vec3f p1(0, 0, -(near_clip+0.01f));
- glh::vec3f p2(0, 0, -(near_clip+1.f));
+ LLVector4a screen_origin(LLVector4a::getZero());
- glh::vec3f screen_origin(0, 0, 0);
+ light_to_screen.affineTransform(p1,p1);
+ light_to_screen.affineTransform(p2,p2);
+ light_to_screen.affineTransform(screen_origin,screen_origin);
- light_to_screen.mult_matrix_vec(p1);
- light_to_screen.mult_matrix_vec(p2);
- light_to_screen.mult_matrix_vec(screen_origin);
+ LLVector4a n;
+ n.setSub(p2,p1);
+ n.normalize3fast();
- glh::vec3f n = p2-p1;
- n.normalize();
-
F32 proj_range = far_clip - near_clip;
- glh::matrix4f light_proj = gl_perspective(fovy, aspect, near_clip, far_clip);
- screen_to_light = trans * light_proj * screen_to_light;
- shader.uniformMatrix4fv(LLShaderMgr::PROJECTOR_MATRIX, 1, FALSE, screen_to_light.m);
- shader.uniform3fv(LLShaderMgr::PROJECTOR_P, 1, p1.v);
- shader.uniform3fv(LLShaderMgr::PROJECTOR_N, 1, n.v);
- shader.uniform3fv(LLShaderMgr::PROJECTOR_ORIGIN, 1, screen_origin.v);
+ LLMatrix4a light_proj = ALGLMath::genPersp(fovy, aspect, near_clip, far_clip);
+ light_proj.setMul(ALGLMath::genNDCtoWC(),light_proj);
+ screen_to_light.setMul(light_proj,screen_to_light);
+
+ shader.uniformMatrix4fv(LLShaderMgr::PROJECTOR_MATRIX, 1, FALSE, screen_to_light.getF32ptr());
+ shader.uniform3fv(LLShaderMgr::PROJECTOR_P, 1, p1.getF32ptr());
+ shader.uniform3fv(LLShaderMgr::PROJECTOR_N, 1, n.getF32ptr());
+ shader.uniform3fv(LLShaderMgr::PROJECTOR_ORIGIN, 1, screen_origin.getF32ptr());
shader.uniform1f(LLShaderMgr::PROJECTOR_RANGE, proj_range);
shader.uniform1f(LLShaderMgr::PROJECTOR_AMBIANCE, params.mV[2]);
--
GitLab