Culling the Battlefield: Data Oriented Design in Practice

Culling the Battlefield
Daniel Collin (DICE)

Tuesday, March 8, 2011

Overview
›Background
›Why rewrite it?
›Requirements
›Target hardware
›Details
›Software Occlusion
›Conclusion


Background of the old culling
›Hierarchical Sphere Trees
›StaticCullTree
›DynamicCullTree


Background of the old culling


Why rewrite it?
›DynamicCullTree scaling
›Sub-levels
›Pipeline dependencies
›Hard to scale
›One job per frustum


Job graph (Old Culling)
Job 0 DynamicCullJob (Shadow 1, 2, View frustum)

Shadow 1
Job 1 frustum
Merge Job

Shadow 2
Job 2 frustum
Merge Job

Job 3 View frustum Merge Job

Bitmasks


Requirements for new system

›Better scaling
›Destruction
›Real-time editing
›Simpler code
›Uniﬁcation of sub-systems


Target hardware


What doesn’t work well on these systems?
›Non-local data
›Branches
›Switching between register types (LHS)
›Tree based structures are usually branch heavy
›Data is the most important thing to address


What does work well on these systems?
›Local data
›(SIMD) Computing power
›Parallelism


The new culling
›Our worlds usually has max ~15000 objects
›First try was to just use parallel brute force
›3x times faster than the old culling
›1/5 code size
›Easier to optimize even further


The new culling
›Linear arrays scale great
›Predictable data
›Few branches
›Uses the computing power


The new culling


Performance numbers (no occlusion)
15000 Spheres
Platform 1 Job 4 Jobs

Xbox 360 1.55 ms (2.10 ms / 4) = 0.52

x86 (Core i7, 2.66 GHz) 1.0 ms (1.30 ms / 4) = 0.32

Playstation 3 0.85 ms ((0.95 ms / 4) = 0.23

Playstation 3 (SPA) 0.63 ms (0.75 ms / 4) = 0.18


Details of the new culling
›Improve performance with a simple grid
›Really an AABB assigned to a “cell” with spheres
›Separate grids for
›
› Rendering: Static
› Rendering: Dynamic
› Physics: Static
› Physics: Dynamic


Data layout
EntityGridCell Block
Block* Pointer Pointer Pointer positions x, y, z, r x, y, z, r …
u8 Count Count Count entityInfo handle, Handle, …
u32 Total Count transformData … … …

struct TransformData
{
half rotation[4];
half minAabb[3];
half pad[1];
half maxAabb[3];
half scale[3];
};


Adding objects
• Pre-allocated array that we can grab data from
AtomicAdd(…) to “alloc” new block
4k 4k …

EntityGridCell

Block* Pointer Pointer Pointer
u8 Count Count Count
u32 Total Count


Adding objects

Cell 0 Cell 1
Cell 4
Cell 2 Cell 3


Removing objects
›Use the “swap trick”
›Data doesn’t need to be sorted
›Just swap with the last entry and decrease the count


Rendering culling
• Let’s look at what the rendering expects
struct EntityRenderCullInfo
{
Handle entity; // handle to the entity
u16 visibleViews; // bits of which frustums that was visible
u16 classId; // type of mesh
float screenArea; // at which screen area entity should be culled
};


Culling code
while (1)
{

uint blockIter = interlockedIncrement(currentBlockIndex) - 1;

if (blockIter >= blockCount) break;

u32 masks[EntityGridCell::Block::MaxCount] = {}, frustumMask = 1;

block = gridCell->blocks[blockIter];

foreach (frustum in frustums, frustumMask <<= 1)

{

for (i = 0; i < gridCell->blockCounts[blockIter]; ++i)

{

u32 inside = intersect(frustum, block->postition[i]);

masks[i] |= frustumMask & inside;

}

}

for (i = 0; i < gridCell->blockCounts[blockIter]; ++i)

{

// ﬁlter list here (if masks[i] is zero it should be skipped)

// ...

}
}


Intersection Code
bool intersect(const Plane* frustumPlanes, Vec4 pos)
{
ﬂoat radius = pos.w;

if (distance(frustumPlanes[Frustum::Far], pos) > radius)

return false;

if (distance(frustumPlanes[Frustum::Near], pos) > radius)

return false;

if (distance(frustumPlanes[Frustum::Right], pos) > radius)

return false;

if (distance(frustumPlanes[Frustum::Left], pos) > radius)

return false;

if (distance(frustumPlanes[Frustum::Upper], pos) > radius)

return false;

if (distance(frustumPlanes[Frustum::Lower], pos) > radius)

return false;

return true;
}


See “Typical C++ Bullshit” by
@mike_acton


Intersection Code
LHS!
bool intersect(const Plane* frustumPlanes, Vec4 pos) Float branch!
{
float radius = pos.w;
return So what do the consoles
false;
think?
LHS! return false;
if (distance(frustumPlanes[Frustum::Right], pos) >Float branch!

:(
radius)
return false;
return false;
LHS! return false;
return false; Float branch!

return true;
} LHS! Float branch!


Intersection Code
bool intersect(const Plane* frustumPlanes, Vec4 pos)
{
float radius = pos.w;
return false;
return false;
if (distance(frustumPlanes[Frustum::Right], pos) > radius)
return false;
return false;
return false;
return false;

return true;
}


Intersection Code
›How can we improve this?
›Dot products are not very SIMD friendly
›Usually need to shuffle data around to get result
›(x0 * x1 + y0 * y1 + z0 * z1 + w0 * w1)


Intersection Code
›Rearrange the data from AoS to SoA
Vec 0 X0 Y0 Z0 W0 VecX X0 X1 X2 X3
Vec 1 X1 Y1 Z1 W1 VecY Y0 Y1 Y2 Y3
Vec 2 X2 Y2 Z2 W2 VecZ Z0 Z1 Z2 Z3
Vec 3 X3 Y3 Z3 W3 VecW W0 W1 W2 W3

›Now we only need 3 instructions for 4 dots!


Rearrange the frustum planes
X0 X1 X2 X3
Plane 0 X0 Y0 Z0 W0 Y0 Y1 Y2 Y3
Plane 1 X1 Y1 Z1 W1 Z0 Z1 Z2 Z3
Plane 2 X2 Y2 Z2 W2 W0 W1 W2 W3
Plane 3 X3 Y3 Z3 W3
Plane 4 X4 Y4 Z4 W4 X4 X5 X4 X5
Plane 5 X5 Y5 Z5 W5 Y4 Y5 Y4 Y5
Z4 Z5 Z4 Z5
W4 W5 W4 W5


New intersection code
›Two frustum vs Sphere intersections per loop
›4 * 3 dot products with 9 instructions
›Loop over all frustums and merge the result


New intersection code (1/4)

Vec posA_xxxx = vecShuffle<VecMask::_xxxx>(posA);
Vec posA_yyyy = vecShuffle<VecMask::_yyyy>(posA);
Vec posA_zzzz = vecShuffle<VecMask::_zzzz>(posA);
Vec posA_rrrr = vecShuffle<VecMask::_wwww>(posA);

// 4 dot products

dotA_0123 = vecMulAdd(posA_zzzz, pl_z0z1z2z3, pl_w0w1w2w3);
dotA_0123 = vecMulAdd(posA_yyyy, pl_y0y1y2y3, dotA_0123);
dotA_0123 = vecMulAdd(posA_xxxx, pl_x0x1x2x3, dotA_0123);



Vec posB_xxxx = vecShuffle<VecMask::_xxxx>(posB);
Vec posB_yyyy = vecShuffle<VecMask::_yyyy>(posB);
Vec posB_zzzz = vecShuffle<VecMask::_zzzz>(posB);
Vec posB_rrrr = vecShuffle<VecMask::_wwww>(posB);

// 4 dot products

dotB_0123 = vecMulAdd(posB_zzzz, pl_z0z1z2z3, pl_w0w1w2w3);
dotB_0123 = vecMulAdd(posB_yyyy, pl_y0y1y2y3, dotB_0123);
dotB_0123 = vecMulAdd(posB_xxxx, pl_x0x1x2x3, dotB_0123



Vec posAB_xxxx = vecInsert<VecMask::_0011>(posA_xxxx, posB_xxxx);
Vec posAB_yyyy = vecInsert<VecMask::_0011>(posA_yyyy, posB_yyyy);
Vec posAB_zzzz = vecInsert<VecMask::_0011>(posA_zzzz, posB_zzzz);
Vec posAB_rrrr = vecInsert<VecMask::_0011>(posA_rrrr, posB_rrrr);

// 4 dot products

dotA45B45 = vecMulAdd(posAB_zzzz, pl_z4z5z4z5, pl_w4w5w4w5);
dotA45B45 = vecMulAdd(posAB_yyyy, pl_y4y5y4y5, dotA45B45);
dotA45B45 = vecMulAdd(posAB_xxxx, pl_x4x5x4x5, dotA45B45);


// Compare against radius

dotA_0123 = vecCmpGTMask(dotA_0123, posA_rrrr);
dotB_0123 = vecCmpGTMask(dotB_0123, posB_rrrr);
dotA45B45 = vecCmpGTMask(dotA45B45, posAB_rrrr);

Vec dotA45 = vecInsert<VecMask::_0011>(dotA45B45, zero);
Vec dotB45 = vecInsert<VecMask::_0011>(zero, dotA45B45);

// collect the results

Vec resA = vecOrx(dotA_0123);
Vec resB = vecOrx(dotB_0123);

resA = vecOr(resA, vecOrx(dotA45));
resB = vecOr(resB, vecOrx(dotB45));

// resA = inside or outside of frustum for point A, resB for point B

Vec rA = vecNotMask(resA);
Vec rB = vecNotMask(resB);

masksCurrent[0] |= frustumMask & rA;
masksCurrent[1] |= frustumMask & rB;


SPU Pipelining Assembler (SPA)
›Like VCL (for PS2) but for PS3 SPU
›Can give you that extra boost if needed
›Does software pipelining for you
›Gives about 35% speed boost in the culling
›Not really that different from using intrinsics
›And coding assembler is fun :)


SPA Inner loop (partly)
lqd posA, -0x20(currentPos)
lqd posB, -0x10(currentPos)

shufb posA_xxxx, posA, posA, Mask_xxxx
shufb posA_yyyy, posA, posA, Mask_yyyy
shufb posA_zzzz, posA, posA, Mask_zzzz
shufb posA_rrrr, posA, posA, Mask_wwww

// 4 dot products

fma dotA_0123, posA_zzzz, pl_z0z1z2z3, pl_w0w1w2w3
fma dotA_0123, posA_yyyy, pl_y0y1y2y3, dotA_0123
fma dotA_0123, posA_xxxx, pl_x0x1x2x3, dotA_0123

shufb posB_xxxx, posB, posB, Mask_xxxx
shufb posB_yyyy, posB, posB, Mask_yyyy
shufb posB_zzzz, posB, posB, Mask_zzzz
shufb posB_rrrr, posB, posB, Mask_wwww

// 4 dot products

fma dotB_0123, posB_zzzz, pl_z0z1z2z3, pl_w0w1w2w3
fma dotB_0123, posB_yyyy, pl_y0y1y2y3, dotB_0123
fma dotB_0123, posB_xxxx, pl_x0x1x2x3, dotB_0123


SPA Inner loop
# Loop stats - frustumCull::loop
# (ims enabled, sms disabled, optimisation level 2)
# resmii : 24 (*) (resource constrained)
# recmii : 2 (recurrence constrained)
# resource usage:
# even pipe : 24 inst. (100% use) (*)
# FX[15] SP[9]
# odd pipe : 24 inst. (100% use) (*)
# SH[17] LS[6] BR[1]
# misc:
# linear schedule = 57 cycles (for information only)
# software pipelining:
# best pipelined schedule = 24 cycles (pipelined, 3 iterations in parallel)
# software pipelining adjustments:
# not generating non-pipelined loop since trip count >=3 (3)
# estimated loop performance:
# =24*n+59 cycles


_local_c0de000000000002:
SPA Inner loop
fma $46,$42,$30,$29; /* +1 */ shufb $47,$44,$37,$33 /* +2 */
fcgt $57,$20,$24; /* +2 */ orx $48,$15 /* +2 */
selb $55,$37,$44,$33; /* +2 */ shufb $56,$21,$16,$33 /* +1 */
fma $52,$16,$28,$41; /* +1 */ orx $49,$57 /* +2 */
ai $4,$4,32; orx $54,$47 /* +2 */
fma $51,$19,$26,$45; /* +1 */ orx $53,$55 /* +2 */
fma $50,$56,$27,$46; /* +1 */ shufb $24,$23,$23,$34 /* +1 */
ai $2,$2,32; /* +2 */ lqd $13,-32($4)
or $69,$48,$54; /* +2 */ lqd $23,-16($4)
fma $20,$18,$26,$52; /* +1 */ lqd $12,-32($2) /* +2 */
nor $60,$69,$69; /* +2 */ lqd $43,-16($2) /* +2 */
or $62,$49,$53; /* +2 */ shufb $59,$22,$17,$33 /* +1 */
and $39,$60,$35; /* +2 */ shufb $11,$14,$24,$33 /* +1 */
nor $61,$62,$60; /* +2 */ shufb $22,$13,$13,$36
fcgt $15,$51,$14; /* +1 */ shufb $17,$23,$23,$36
and $58,$61,$35; /* +2 */ shufb $19,$13,$13,$3
fma $10,$59,$25,$50; /* +1 */ shufb $18,$23,$23,$3
fma $9,$22,$32,$31; shufb $16,$23,$23,$38
or $8,$58,$43; /* +2 */ shufb $21,$13,$13,$38
or $40,$39,$12; /* +2 */ shufb $14,$13,$13,$34
ai $7,$7,-1; /* +2 */ shufb $42,$19,$18,$33
fma $41,$17,$32,$31; stqd $8,-16($2) /* +2 */
fcgt $44,$10,$11; /* +1 */ stqd $40,-32($2) /* +2 */
fma $45,$21,$28,$9; brnz $7,_local_c0de000000000002 /* +2 */
nop ; hbrr


Additional culling
›Frustum vs AABB
›Project AABB to screen space
›Software Occlusion


Project AABB to screen space
›Calculate the area of the AABB in screen space
›If area is smaller than setting just skip it
›Due to FOV taking distance doesn’t work


Software Occlusion
›Used in Frostbite for 3 years
›Cross platform
›Artist made occluders
›Terrain


Why Software Occlusion?
›Want to remove CPU time not just GPU
›Cull as early as possible
›GPU queries troublesome as lagging behind CPU
›Must support destruction
›Easy for artists to control


Software Occlusion
›So how does it work?
›Render PS1 style geometry to a zbuffer using
software rendering
›The zbuffer is 256 x 114 ﬂoat


Software Occlusion
›Occluder triangle setup
›Terrain triangle setup
›Rasterize triangles
›Culling


Software Occlusion (Occluders)


Software Occlusion (In-game)


Culling jobs
Culling Jobs
Job 0 Occluder Triangles Rasterize Triangles Culling Z-buffer Test




Job 4 Terrain Triangles Rasterize Triangles Culling Z-buffer Test


Occluder triangles
Job 0 Inside, Project triangles

Job 1 Inside, Project triangles

Job 2 Intersecting, Clip, Project

Job 3 Outside, Skip

Output


Occluder triangles
Rasterize Triangles
Input

16 Triangles 16 Triangles 16 Triangles 16 Triangles

256 x 114 zbuffer 256 x 114 zbuffer 256 x 114 zbuffer

Job 0 Job 1 Merge step
<Todo: Fix Colors>

z-buffer testing
›Calculate screen space AABB for object
›Get single distance value
›Test the square against the z-buffer


Conclusion
›Accurate and high performance culling is essential
›Reduces pressure on low-level systems/rendering
›It’s all about data
›Simple data often means simple code
›Understanding your target hardware


Thanks to
›Andreas Fredriksson (@deplinenoise)
›Christina Coffin (@christinacoffin)
›Johan Andersson (@repi)
›Stephen Hill (@self_shadow)
›Steven Tovey (@nonchaotic)
›Halldor Fannar
›Evelyn Donis


Questions? Email: daniel@collin.com
Blog: zenic.org
Twitter: @daniel_collin

Battlefield 3 & Frostbite 2 talks at GDC’11:
Mon 1:45 DX11 Rendering in Battlefield 3 Johan Andersson
Wed 10:30 SPU-based Deferred Shading in Battlefield 3 Christina Coffin
for PlayStation 3
Wed 3:00 Culling the Battlefield: Data Oriented Design Daniel Collin
in Practice
Thu 1:30 Lighting You Up in Battlefield 3 Kenny Magnusson
Fri 4:05 Approximating Translucency for a Fast, Colin Barré-Brisebois
Cheap & Convincing Subsurface Scattering
Look

For more DICE talks: http://publications.dice.se

Culling the Battlefield: Data Oriented Design in Practice

Recommended

Recommended

More Related Content

Viewers also liked

Viewers also liked (12)

More from Electronic Arts / DICE

More from Electronic Arts / DICE (20)

Recently uploaded

Recently uploaded (20)

Culling the Battlefield: Data Oriented Design in Practice