Changed the light collision structure uploaded to the GPU to be a binary tree using AABBs for the nodes instead of a BSP plane

This commit is contained in:
Magnus Norddahl 2017-03-07 15:58:22 +01:00
parent 8515f9720a
commit 6df3b3fbca
6 changed files with 294 additions and 157 deletions

View file

@ -1,6 +1,6 @@
//
//---------------------------------------------------------------------------
// Doom BSP tree on the GPU
// 2D collision tree for 1D shadowmap lights
// Copyright(C) 2017 Magnus Norddahl
// All rights reserved.
//
@ -25,6 +25,7 @@
#include "gl/dynlights/gl_lightbsp.h"
#include "gl/system/gl_interface.h"
#include "r_state.h"
#include "g_levellocals.h"
int FLightBSP::GetNodesBuffer()
{
@ -32,10 +33,10 @@ int FLightBSP::GetNodesBuffer()
return NodesBuffer;
}
int FLightBSP::GetSegsBuffer()
int FLightBSP::GetLinesBuffer()
{
UpdateBuffers();
return SegsBuffer;
return LinesBuffer;
}
void FLightBSP::UpdateBuffers()
@ -49,55 +50,19 @@ void FLightBSP::UpdateBuffers()
void FLightBSP::GenerateBuffers()
{
if (!Shape)
Shape.reset(new Level2DShape());
UploadNodes();
UploadSegs();
}
void FLightBSP::UploadNodes()
{
TArray<GPUNode> gpunodes;
gpunodes.Resize(numnodes);
for (int i = 0; i < numnodes; i++)
{
const auto &node = nodes[i];
auto &gpunode = gpunodes[i];
float a = -FIXED2FLOAT(node.dy);
float b = FIXED2FLOAT(node.dx);
float c = 0.0f;
float d = -(a * FIXED2FLOAT(node.x) + b * FIXED2FLOAT(node.y));
gpunode.plane[0] = a;
gpunode.plane[1] = b;
gpunode.plane[2] = c;
gpunode.plane[3] = d;
for (int j = 0; j < 2; j++)
{
bool isNode = (!((size_t)node.children[j] & 1));
if (isNode)
{
node_t *bsp = (node_t *)node.children[j];
gpunode.children[j] = (int)(ptrdiff_t)(bsp - nodes);
gpunode.linecount[j] = -1;
}
else
{
subsector_t *sub = (subsector_t *)((BYTE *)node.children[j] - 1);
if (sub->numlines > 0)
gpunode.children[j] = (int)(ptrdiff_t)(sub->firstline - segs);
else
gpunode.children[j] = 0;
gpunode.linecount[j] = sub->numlines;
}
}
}
#if 0
if (gpunodes.Size() > 0)
if (Shape->nodes.Size() > 0)
{
FILE *file = fopen("nodes.txt", "wb");
fwrite(&gpunodes[0], sizeof(GPUNode) * gpunodes.Size(), 1, file);
fwrite(&Shape->nodes[0], sizeof(GPUNode) * Shape->nodes.Size(), 1, file);
fclose(file);
}
#endif
@ -107,7 +72,7 @@ void FLightBSP::UploadNodes()
glGenBuffers(1, (GLuint*)&NodesBuffer);
glBindBuffer(GL_SHADER_STORAGE_BUFFER, NodesBuffer);
glBufferData(GL_SHADER_STORAGE_BUFFER, sizeof(GPUNode) * gpunodes.Size(), &gpunodes[0], GL_STATIC_DRAW);
glBufferData(GL_SHADER_STORAGE_BUFFER, sizeof(GPUNode) * Shape->nodes.Size(), &Shape->nodes[0], GL_STATIC_DRAW);
glBindBuffer(GL_SHADER_STORAGE_BUFFER, oldBinding);
NumNodes = numnodes;
@ -115,28 +80,24 @@ void FLightBSP::UploadNodes()
void FLightBSP::UploadSegs()
{
TArray<GPUSeg> gpusegs;
gpusegs.Resize(numsegs);
for (int i = 0; i < numsegs; i++)
TArray<GPULine> gpulines;
gpulines.Resize(level.lines.Size());
for (unsigned int i = 0; i < level.lines.Size(); i++)
{
const auto &seg = segs[i];
auto &gpuseg = gpusegs[i];
const auto &line = level.lines[i];
auto &gpuseg = gpulines[i];
gpuseg.x = (float)seg.v1->fX();
gpuseg.y = (float)seg.v1->fY();
gpuseg.dx = (float)seg.v2->fX() - gpuseg.x;
gpuseg.dy = (float)seg.v2->fY() - gpuseg.y;
gpuseg.bSolid = (seg.backsector == nullptr) ? 1.0f : 0.0f;
gpuseg.padding1 = 0.0f;
gpuseg.padding2 = 0.0f;
gpuseg.padding3 = 0.0f;
gpuseg.x = (float)line.v1->fX();
gpuseg.y = (float)line.v1->fY();
gpuseg.dx = (float)line.v2->fX() - gpuseg.x;
gpuseg.dy = (float)line.v2->fY() - gpuseg.y;
}
#if 0
if (gpusegs.Size() > 0)
if (gpulines.Size() > 0)
{
FILE *file = fopen("segs.txt", "wb");
fwrite(&gpusegs[0], sizeof(GPUSeg) * gpusegs.Size(), 1, file);
FILE *file = fopen("lines.txt", "wb");
fwrite(&gpulines[0], sizeof(GPULine) * gpulines.Size(), 1, file);
fclose(file);
}
#endif
@ -144,9 +105,9 @@ void FLightBSP::UploadSegs()
int oldBinding = 0;
glGetIntegerv(GL_SHADER_STORAGE_BUFFER_BINDING, &oldBinding);
glGenBuffers(1, (GLuint*)&SegsBuffer);
glBindBuffer(GL_SHADER_STORAGE_BUFFER, SegsBuffer);
glBufferData(GL_SHADER_STORAGE_BUFFER, sizeof(GPUSeg) * gpusegs.Size(), &gpusegs[0], GL_STATIC_DRAW);
glGenBuffers(1, (GLuint*)&LinesBuffer);
glBindBuffer(GL_SHADER_STORAGE_BUFFER, LinesBuffer);
glBufferData(GL_SHADER_STORAGE_BUFFER, sizeof(GPULine) * gpulines.Size(), &gpulines[0], GL_STATIC_DRAW);
glBindBuffer(GL_SHADER_STORAGE_BUFFER, oldBinding);
NumSegs = numsegs;
@ -159,9 +120,145 @@ void FLightBSP::Clear()
glDeleteBuffers(1, (GLuint*)&NodesBuffer);
NodesBuffer = 0;
}
if (SegsBuffer != 0)
if (LinesBuffer != 0)
{
glDeleteBuffers(1, (GLuint*)&SegsBuffer);
SegsBuffer = 0;
glDeleteBuffers(1, (GLuint*)&LinesBuffer);
LinesBuffer = 0;
}
Shape.reset();
}
/////////////////////////////////////////////////////////////////////////////
Level2DShape::Level2DShape()
{
TArray<int> lines;
TArray<FVector2> centroids;
for (unsigned int i = 0; i < level.lines.Size(); i++)
{
if (level.lines[i].backsector)
{
centroids.Push(FVector2(0.0f, 0.0f));
continue;
}
lines.Push(i);
FVector2 v1 = { (float)level.lines[i].v1->fX(), (float)level.lines[i].v1->fY() };
FVector2 v2 = { (float)level.lines[i].v2->fX(), (float)level.lines[i].v2->fY() };
centroids.Push((v1 + v2) * 0.5f);
}
TArray<int> work_buffer;
work_buffer.Resize(lines.Size() * 2);
root = subdivide(&lines[0], (int)lines.Size(), &centroids[0], &work_buffer[0]);
}
int Level2DShape::subdivide(int *lines, int num_lines, const FVector2 *centroids, int *work_buffer)
{
if (num_lines == 0)
return -1;
// Find bounding box and median of the lines
FVector2 median = FVector2(0.0f, 0.0f);
FVector2 aabb_min, aabb_max;
aabb_min.X = (float)level.lines[lines[0]].v1->fX();
aabb_min.Y = (float)level.lines[lines[0]].v1->fY();
aabb_max = aabb_min;
for (int i = 0; i < num_lines; i++)
{
float x1 = (float)level.lines[lines[i]].v1->fX();
float y1 = (float)level.lines[lines[i]].v1->fY();
float x2 = (float)level.lines[lines[i]].v2->fX();
float y2 = (float)level.lines[lines[i]].v2->fY();
aabb_min.X = MIN(aabb_min.X, x1);
aabb_min.X = MIN(aabb_min.X, x2);
aabb_min.Y = MIN(aabb_min.Y, y1);
aabb_min.Y = MIN(aabb_min.Y, y2);
aabb_max.X = MAX(aabb_max.X, x1);
aabb_max.X = MAX(aabb_max.X, x2);
aabb_max.Y = MAX(aabb_max.Y, y1);
aabb_max.Y = MAX(aabb_max.Y, y2);
median += centroids[lines[i]];
}
median /= (float)num_lines;
if (num_lines == 1) // Leaf node
{
nodes.Push(GPUNode(aabb_min, aabb_max, lines[0]));
return (int)nodes.Size() - 1;
}
// Find the longest axis
float axis_lengths[2] =
{
aabb_max.X - aabb_min.X,
aabb_max.Y - aabb_min.Y
};
int axis_order[2] = { 0, 1 };
FVector2 axis_plane[2] = { FVector2(1.0f, 0.0f), FVector2(0.0f, 1.0f) };
std::sort(axis_order, axis_order + 2, [&](int a, int b) { return axis_lengths[a] > axis_lengths[b]; });
// Try split at longest axis, then if that fails the next longest, and then the remaining one
int left_count, right_count;
FVector2 axis;
for (int attempt = 0; attempt < 2; attempt++)
{
// Find the split plane for axis
FVector2 axis = axis_plane[axis_order[attempt]];
FVector3 plane(axis, -(median | axis));
// Split lines into two
left_count = 0;
right_count = 0;
for (int i = 0; i < num_lines; i++)
{
int line_index = lines[i];
float side = FVector3(centroids[lines[i]], 1.0f) | plane;
if (side >= 0.0f)
{
work_buffer[left_count] = line_index;
left_count++;
}
else
{
work_buffer[num_lines + right_count] = line_index;
right_count++;
}
}
if (left_count != 0 && right_count != 0)
break;
}
// Check if something went wrong when splitting and do a random split instead
if (left_count == 0 || right_count == 0)
{
left_count = num_lines / 2;
right_count = num_lines - left_count;
}
else
{
// Move result back into lines list:
for (int i = 0; i < left_count; i++)
lines[i] = work_buffer[i];
for (int i = 0; i < right_count; i++)
lines[i + left_count] = work_buffer[num_lines + i];
}
// Create child nodes:
int left_index = -1;
int right_index = -1;
if (left_count > 0)
left_index = subdivide(lines, left_count, centroids, work_buffer);
if (right_count > 0)
right_index = subdivide(lines + left_count, right_count, centroids, work_buffer);
nodes.Push(GPUNode(aabb_min, aabb_max, left_index, right_index));
return (int)nodes.Size() - 1;
}

View file

@ -1,19 +1,37 @@
#pragma once
#include <memory>
struct GPUNode
{
float plane[4];
int children[2];
int linecount[2];
GPUNode(const FVector2 &aabb_min, const FVector2 &aabb_max, int line_index) : aabb_left(aabb_min.X), aabb_top(aabb_min.Y), aabb_right(aabb_max.X), aabb_bottom(aabb_max.Y), left(-1), right(-1), line_index(line_index) { }
GPUNode(const FVector2 &aabb_min, const FVector2 &aabb_max, int left, int right) : aabb_left(aabb_min.X), aabb_top(aabb_min.Y), aabb_right(aabb_max.X), aabb_bottom(aabb_max.Y), left(left), right(right), line_index(-1) { }
float aabb_left, aabb_top;
float aabb_right, aabb_bottom;
int left;
int right;
int line_index;
int padding;
};
struct GPUSeg
struct GPULine
{
float x, y;
float dx, dy;
float bSolid;
float padding1, padding2, padding3;
};
class Level2DShape
{
public:
Level2DShape();
TArray<GPUNode> nodes;
int root;
private:
int subdivide(int *lines, int num_lines, const FVector2 *centroids, int *work_buffer);
};
class FLightBSP
@ -23,7 +41,7 @@ public:
~FLightBSP() { Clear(); }
int GetNodesBuffer();
int GetSegsBuffer();
int GetLinesBuffer();
void Clear();
private:
@ -36,7 +54,9 @@ private:
FLightBSP &operator=(FLightBSP &) = delete;
int NodesBuffer = 0;
int SegsBuffer = 0;
int LinesBuffer = 0;
int NumNodes = 0;
int NumSegs = 0;
std::unique_ptr<Level2DShape> Shape;
};

View file

@ -55,7 +55,7 @@ void FShadowMap::Update()
GLRenderer->mShadowMapShader->Bind();
glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 4, mLightList);
glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 2, mLightBSP.GetNodesBuffer());
glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 3, mLightBSP.GetSegsBuffer());
glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 3, mLightBSP.GetLinesBuffer());
glViewport(0, 0, 1024, 1024);
GLRenderer->RenderScreenQuad();

View file

@ -772,6 +772,11 @@ void FGLRenderBuffers::CreateShadowMap()
glGetIntegerv(GL_FRAMEBUFFER_BINDING, &frameBufferBinding);
mShadowMapTexture = Create2DTexture("ShadowMap", GL_R32F, 1024, 1024);
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
mShadowMapFB = CreateFrameBuffer("ShadowMapFB", mShadowMapTexture);
glBindTexture(GL_TEXTURE_2D, textureBinding);

View file

@ -158,7 +158,7 @@ float shadowmapAttenuation(vec4 lightpos, float shadowIndex)
else
u = dir.x / dir.y * 0.125 + (0.50 + 0.125);
}
dir -= sign(dir); // margin, to remove wall acne
dir -= sign(dir) * 5.0; // margin, to remove wall acne
float dist2 = dot(dir, dir);
return texture(ShadowMap, vec2(u, v)).x > dist2 ? 1.0 : 0.0;
}

View file

@ -4,26 +4,28 @@ out vec4 FragColor;
struct GPUNode
{
vec4 plane;
int children[2];
int linecount[2];
vec2 aabb_min;
vec2 aabb_max;
int left;
int right;
int line_index;
int padding;
};
struct GPUSeg
struct GPULine
{
vec2 pos;
vec2 delta;
vec4 bSolid;
};
layout(std430, binding = 2) buffer LightNodes
{
GPUNode bspNodes[];
GPUNode nodes[];
};
layout(std430, binding = 3) buffer LightSegs
layout(std430, binding = 3) buffer LightLines
{
GPUSeg bspSegs[];
GPULine lines[];
};
layout(std430, binding = 4) buffer LightList
@ -31,87 +33,100 @@ layout(std430, binding = 4) buffer LightList
vec4 lights[];
};
//===========================================================================
//
// Ray/BSP collision test. Returns where the ray hit something.
//
//===========================================================================
vec2 rayTest(vec2 from, vec2 to)
bool overlapRayAABB(vec2 ray_start2d, vec2 ray_end2d, vec2 aabb_min2d, vec2 aabb_max2d)
{
// To do: simplify test to use a 2D test
vec3 ray_start = vec3(ray_start2d, 0.0);
vec3 ray_end = vec3(ray_end2d, 0.0);
vec3 aabb_min = vec3(aabb_min2d, -1.0);
vec3 aabb_max = vec3(aabb_max2d, 1.0);
vec3 c = (ray_start + ray_end) * 0.5f;
vec3 w = ray_end - c;
vec3 h = (aabb_max - aabb_min) * 0.5f; // aabb.extents();
c -= (aabb_max + aabb_min) * 0.5f; // aabb.center();
vec3 v = abs(w);
if (abs(c.x) > v.x + h.x || abs(c.y) > v.y + h.y || abs(c.z) > v.z + h.z)
return false; // disjoint;
if (abs(c.y * w.z - c.z * w.y) > h.y * v.z + h.z * v.y ||
abs(c.x * w.z - c.z * w.x) > h.x * v.z + h.z * v.x ||
abs(c.x * w.y - c.y * w.x) > h.x * v.y + h.y * v.x)
return false; // disjoint;
return true; // overlap;
}
float intersectRayLine(vec2 ray_start, vec2 ray_end, int line_index, vec2 raydelta, float rayd, float raydist2)
{
const int max_iterations = 50;
const float epsilon = 0.0000001;
GPULine line = lines[line_index];
// Avoid wall acne by adding some margin
vec2 margin = normalize(to - from);
vec2 raydelta = to - from;
float raydist2 = dot(raydelta, raydelta);
vec2 raynormal = vec2(raydelta.y, -raydelta.x);
float rayd = dot(raynormal, from);
if (raydist2 < 1.0 || bspNodes.length() == 0)
return to;
int nodeIndex = bspNodes.length() - 1;
for (int iteration = 0; iteration < max_iterations; iteration++)
float den = dot(raynormal, line.delta);
if (abs(den) > epsilon)
{
GPUNode node = bspNodes[nodeIndex];
int side = (dot(node.plane, vec4(from, 0.0, 1.0)) > 0.0) ? 1 : 0;
int linecount = node.linecount[side];
if (linecount < 0)
float t_line = (rayd - dot(raynormal, line.pos)) / den;
if (t_line >= 0.0 && t_line <= 1.0)
{
nodeIndex = node.children[side];
}
else
{
int startLineIndex = node.children[side];
// Ray/line test each line segment.
bool hit_line = false;
for (int i = 0; i < linecount; i++)
{
GPUSeg seg = bspSegs[startLineIndex + i];
float den = dot(raynormal, seg.delta);
if (abs(den) > epsilon)
{
float t_seg = (rayd - dot(raynormal, seg.pos)) / den;
if (t_seg >= 0.0 && t_seg <= 1.0)
{
vec2 seghitdelta = seg.pos + seg.delta * t_seg - from;
if (dot(raydelta, seghitdelta) > 0.0 && dot(seghitdelta, seghitdelta) < raydist2) // We hit a line segment.
{
if (seg.bSolid.x > 0.0) // segment line is one-sided
return from + seghitdelta;
// We hit a two-sided segment line. Move to the other side and continue ray tracing.
from = from + seghitdelta + margin;
raydelta = to - from;
raydist2 = dot(raydelta, raydelta);
raynormal = vec2(raydelta.y, -raydelta.x);
rayd = dot(raynormal, from);
if (raydist2 < 1.0 || bspNodes.length() == 0)
return to;
nodeIndex = bspNodes.length() - 1;
hit_line = true;
break;
}
}
}
}
if (!hit_line)
return to;
vec2 linehitdelta = line.pos + line.delta * t_line - ray_start;
float t = dot(raydelta, linehitdelta) / raydist2;
return t > 0.0 ? t : 1.0;
}
}
return to;
return 1.0;
}
bool isLeaf(int node_index)
{
return nodes[node_index].line_index != -1;
}
float rayTest(vec2 ray_start, vec2 ray_end)
{
vec2 raydelta = ray_end - ray_start;
float raydist2 = dot(raydelta, raydelta);
vec2 raynormal = vec2(raydelta.y, -raydelta.x);
float rayd = dot(raynormal, ray_start);
if (raydist2 < 1.0)
return 1.0;
float t = 1.0;
int stack[16];
int stack_pos = 1;
stack[0] = nodes.length() - 1;
while (stack_pos > 0)
{
int node_index = stack[stack_pos - 1];
if (!overlapRayAABB(ray_start, ray_end, nodes[node_index].aabb_min, nodes[node_index].aabb_max))
{
stack_pos--;
}
else if (isLeaf(node_index))
{
t = min(intersectRayLine(ray_start, ray_end, nodes[node_index].line_index, raydelta, rayd, raydist2), t);
stack_pos--;
}
else if (stack_pos == 16)
{
stack_pos--; // stack overflow
}
else
{
stack[stack_pos - 1] = nodes[node_index].left;
stack[stack_pos] = nodes[node_index].right;
stack_pos++;
}
}
return t;
}
void main()
@ -134,8 +149,8 @@ void main()
}
pixelpos = lightpos + pixelpos * radius;
vec2 hitpos = rayTest(lightpos, pixelpos);
vec2 delta = hitpos - lightpos;
float t = rayTest(lightpos, pixelpos);
vec2 delta = (pixelpos - lightpos) * t;
float dist2 = dot(delta, delta);
FragColor = vec4(dist2, 0.0, 0.0, 1.0);