Compilation For GPU Accelerated Ray Tracing in OptiX PDF

Compilation for GPU
Accelerated Ray Tracing in OptiX

Steven G. Parker
Tuesday, August 16, 2011
Overview
Lessons from history

OptiX overview
CUDA as a shading language
OptiX compiler
Results
RTSL: a Ray Tracing Shading

Language
Steven G. Parker
Solomon Boulos
James Bigler
Austin Robison
University of Utah
3
Foundation
Build on OpenGL Shading Language (GLSL)
Simple, clean, C-like syntax
Simple type system
float, int, bool
small vectors (2-4 elements)
small matrices (float only)
structs, fixed-length arrays, no pointers
Convenient element access syntax:

v1.xy = v2.yx; v3.xyz = v4.yyy;
Graphics oriented operators, functions
New functions
Convenience:
dominant axis of vector
vector perpendicular to vector
range test
horizontal min/max
Additional functionality:
random numbers
matrix inversion
overloaded pow for integer exponent
Ray Tracing:
trace(): Recursive ray tracing call
hit(): Report possible intersection with primitive
Deviation from GLSL

Object model with inheritance:
class Sphere : rt_Primitive;
Reference type
Different variable qualifiers
GLSL: attribute, uniform, varying, const
RTSL: public, private, scratch, const
First class color type (not always 3/4 components)
Computational Model
dow
Sha
s
Ray
Point,
Normal
Light
Frame Buffer
or
Parent Ray
Camera
Lighting
Engine
LightColor,
LightDirection,
LightDistance
RayOrigin,
RayDirection
Se
Ra con
ys da
ry
Render Loop
ScreenCoord,
LensCoord
lor
leCo
p
m
Sa
RayOrigin,
RayDirection
GeometricNormal,
ShadingNormal
HitPoint
HitDistance
Primitive
Material
TextureUV/UVW
TextureColor
RTSL
Scene Traversal
Texture
Rendering Engine
OptiX Goals
Make GPU ray tracing simpler

Function in a resource limited device
Achieve high performance
Able to express most ray tracing algorithms
Algorithm agnostic
User defined payloads

Programmable intersection
Interoperate with rasterization pipeline
Dont require a new language
Leverage CUDA architecture and existing compiler infrastructure
Life of a ray
1 Ray Generation
2
Intersection
3 Shading
Pinhole
1 Camera
Payload
float3 color
Ray-Sphere
Intersection
Lambertian
Shading
10
Life of a ray
1
Pinhole
Camera
RT_PROGRAM void pinhole_camera()

{
float2 d = make_float2
(launch_index) / make_float2
(launch_dim) * 2.f - 1.f;
float3 ray_origin = eye;
float3 ray_direction = normalize
(d.x*U + d.y*V + W);
optix::Ray ray = optix::make_Ray
(ray_origin, ray_direction,
radiance_ray_type, scene_epsilon,
RT_DEFAULT_MAX);
PerRayData_radiance prd;
rtTrace(top_object, ray, prd);
output_buffer[launch_index] =
make_color( prd.result );
}
Ray-Sphere
Intersection
RT_PROGRAM void intersect_sphere()

{
float3 O = ray.origin - center;
float3 D = ray.direction;
float b = dot(O, D);
float c = dot(O, O)-radius*radius;
float disc = b*b-c;
if(disc > 0.0f){
float sdisc = sqrtf(disc);
float root1 = (-b - sdisc);
bool check_second = true;
if( rtPotentialIntersection
( root1 ) ) {
shading_normal = geometric_normal
= (O + root1*D)/radius;
if(rtReportIntersection(0))
check_second = false;
}
if(check_second) {
float root2 = (-b + sdisc);
if( rtPotentialIntersection
( root2 ) ) {
shading_normal =
geometric_normal = (O + root2*D)/
radius;
rtReportIntersection(0);
}
}
}
Lambertian
Shading
RT_PROGRAM void closest_hit_radiance3()

{
float3 world_geo_normal
= normalize
( rtTransformNormal
( RT_OBJECT_TO_WORLD,
geometric_normal ) );
float3 world_shade_normal = normalize
( rtTransformNormal
( RT_OBJECT_TO_WORLD,
shading_normal ) );
float3 ffnormal
= faceforward
( world_shade_normal, -ray.direction,
world_geo_normal );
float3 color = Ka *
ambient_light_color;
float3 hit_point = ray.origin + t_hit
* ray.direction;
for(int i = 0; i < lights.size(); +
+i) {
BasicLight light = lights[i];
float3 L = normalize(light.pos hit_point);
float nDl = dot( ffnormal, L);
if( nDl > 0.0f ){
// cast shadow ray
PerRayData_shadow shadow_prd;
11
Program objects (shaders)

{
float2 d = make_float2
(launch_index) / make_float2
(launch_dim) * 2.f - 1.f;
float3 ray_origin = eye;
float3 ray_direction = normalize
(d.x*U + d.y*V + W);
Interconnection of programs defines the outcome
optix::Ray ray = optix::make_Ray

(ray_origin, ray_direction,
radiance_ray_type, scene_epsilon,
RT_DEFAULT_MAX);
Input language is based on CUDA C/C++
output_buffer[launch_index] =
}
Data associated with ray is programmable

No new language to learn
Powerful language features available immediately
(templates, function overloading, default values, ...)
Can also take raw PTX as input
Caveat: still need to use it responsibly to get performance
12
Abstract Execution Model

Launch
Ray Generation
Program
rtContextLaunch
Exception
Program
rtTrace
Shade
Traverse
Miss
Program
Node Graph
Traversal
Selector Visit
Program
Acceleration
Traversal
Closest Hit
Program
Intersection
Program
Any Hit
Program
13
Programmable Operations
Rasterization
Ray Tracing
Fragment
Closest Hit
Any Hit
Vertex
Intersection
Geometry
Selector
Hull/Domain (Tesselation)
Ray Generation
Miss
Exception
14
Flexible intersection
Primary Ray
Intersection (miss)
Intersection (hit)
Any Hit
Intersection (miss)
Intersection (hit)
Any hit (ignore intersection)
Closest hit
15
Flexible intersection
Shadow Ray
Intersection (hit)
Any Hit (ignore intersection)
Intersection (hit)
Any hit (terminate ray)
Closest hit
16
Geometry representation
Geometry Instance
Geometry
Intersection
Material
Material
Material
Geometry instance binds

geometry to one or more
materials
Closest Hit/Any Hit pairs per raytype (user defined)
Closest
Hit 00 Any
Hit 00
Closest
ClosestHit
Hit 0 Any
AnyHit
Hit 0
Closest
Hit 11 Any
Hit 11
Closest
ClosestHit
Hit 1 Any
AnyHit
Hit 1
17
Scene representation
Context
Group
Acceleration
Transform
Transform
Acceleration
Geometry Group
Geometry Instance
Geometry
Intersection
Material
Closest
Hit 00
Closest
ClosestHit
Hit 0
Any
Hit 00
Any
AnyHit
Hit 0
Closest
Hit 11
Closest
ClosestHit
Hit 1
Any
Hit 11
Any
AnyHit
Hit 1
Geometry
Intersection
Geometry Instance
Material
Material
All objects can have multiple parents

(instancing)
Geometry data
Materials
Acceleration structures
Tree may have multiple roots

Separate objects for shadow rays
Not shown: Selector

Programmable (LOD, switch, etc).
Material
Material
Material
Closest
Hit 00
Closest
ClosestHit
Hit 0
Any
Hit 00
Any
AnyHit
Hit 0
Closest
Hit 11
Closest
ClosestHit
Hit 1
Any
Hit 11
Any
AnyHit
Hit 1
18
Object variable inheritance
Context
Group
Acceleration
Transform
Transform
Acceleration
Geometry Group
Geometry Instance
Geometry
Intersection
Geometry Instance
Material
Material
Material
Closest
Hit 00
Closest
ClosestHit
Hit 0
Any
Hit 00
Any
AnyHit
Hit 0
Closest
Hit 11
Closest
ClosestHit
Hit 1
Any
Hit 11
Any
AnyHit
Hit 1
Material parameters can be

attached to context (global),
geometry instance or material
Geometry
Intersection
Material
Material
Material
Closest
Hit 00
Closest
ClosestHit
Hit 0
Any
Hit 00
Any
AnyHit
Hit 0
Closest
Hit 11
Closest
ClosestHit
Hit 1
Any
Hit 11
Any
AnyHit
Hit 1
Variables are one of:

A small primitive type (float4, matrix, ...)
A small user defined type
A handle to a buffer (1D, 2D, 3D)
A texture
19
Lights
Context
Group
Acceleration
Transform
Transform
Acceleration
Geometry Group
Geometry Instance
Geometry
Intersection
Geometry Instance
Material
Material
Material
Closest
Hit 00
Closest
ClosestHit
Hit 0
Any
Hit 00
Any
AnyHit
Hit 0
Closest
Hit 11
Closest
ClosestHit
Hit 1
Any
Hit 11
Any
AnyHit
Hit 1

Geometry
Intersection
Lights
Material
Material
Material
Closest
Hit 00
Closest
ClosestHit
Hit 0
Any
Hit 00
Any
AnyHit
Hit 0
Closest
Hit 11
Closest
ClosestHit
Hit 1
Any
Hit 11
Any
AnyHit
Hit 1

A texture
19
Lights
Context
Group
Acceleration
Transform
Transform
Acceleration
Geometry Group
Geometry Instance
Geometry
Intersection
Geometry Instance
Material
Material
Material
Closest
Hit 00
Closest
ClosestHit
Hit 0
Any
Hit 00
Any
AnyHit
Hit 0
Closest
Hit 11
Closest
ClosestHit
Hit 1
Any
Hit 11
Any
AnyHit
Hit 1

Geometry
Intersection
Lights
Material
Material
Material
Closest
Hit 00
Closest
ClosestHit
Hit 0
Any
Hit 00
Any
AnyHit
Hit 0
Closest
Hit 11
Closest
ClosestHit
Hit 1
Any
Hit 11
Any
AnyHit
Hit 1

A texture
19
CUDA as a shading language
Shading language for OptiX is a restricted subset of the CUDA device code
Just C++
A few conventions for accessing runtime (object model, trace functions, etc.)
No data management functionality required from CUDA
Some CUDA functionality disallowed (shared memory, barriers, etc.)
Use CUDA compiler (nvcc)
Produces PTX
20
Intersection program
!"#"$%&'"()&*+,-"$")$./)-&#().')012"3#
4"#().5$&16#"()7'0$%.89)#":#6$")300$;&'.#"(<)
=(";)1/)380("(#)-&#)(-.;"$)*0$)(-.;&'>
4"8"3#(),-&3-)%.#"$&.8)#0)6("
=(";)*0$
?$0>$.%%.18")(6$*.3"(
@880,&'>).$1&#$.$/)#$&.'>8")16A"$)*0$%.#(
B#3C
21
rtDeclareVariable(float3, p0, );
rtDeclareVariable(float3, geometric_normal, attribute geometric_normal, );
rtDeclareVariable(float3, shading_normal, attribute shading_normal, );
rtDeclareVariable(optix::Ray, ray, rtCurrentRay, );
RT_PROGRAM void triangle_intersect(int)

{
// Intersect ray with triangle
float3 e0 = p1 - p0;
float3 n = cross( e0, e1 );
float v
float r
= dot( n, ray.direction );
= 1.0f / v;
if(t < ray.tmax && t > ray.tmin) {

float3 i
= cross( e2, ray.direction );
float v1
= dot( i, e1 );
float beta = r*v1;
if(beta >= 0.0f){
float v2 = dot( i, e0 );
float gamma = r*v2;
if( (v1+v2)*v <= v*v && gamma >= 0.0f ) {
if(
float3 e2 = p0 - ray.origin;
float va = dot( n, e2 );
float t
= r*va;
rtPotentialIntersection( t ) ) {
shading_normal = geometric_normal = -n;
rtReportIntersection( 0 );
}}}}}
22
rtBuffer<Vertex> vertex_buffer;
rtBuffer<uint3> index_buffer;
rtDeclareVariable(float3, geometric_normal, attribute geometric_normal, );
rtDeclareVariable(float3, shading_normal, attribute shading_normal, );
RT_PROGRAM void mesh_intersect(int primIdx)

{
uint3 v_idx = index_buffer[primIdx];
float3 p0 = vertex_buffer[v_idx].p0;
// Intersect ray with triangle
float3 n = cross( e0, e1 );
float v
float r
...
= dot( n, ray.direction );
= 1.0f / v;
23
Closest hit program

(traditional shader)
!"D'"(),-.#)-.EE"'(),-"').)$./)-&#().')012"3#
B:"36#";)*0$)'".$"(#)&'#"$("3F0')7380("(#)-&#<).80'>).)$./
@6#0%.F3.88/)E"$*0$%();"*"$$";)(-.;&'>)7(-.;")0'3")E"$)$./<
G.')$"36$(&H"8/)(-00#)%0$")$./(
4-.;0,(
I"J"3F0'(
@%1&"'#)03386(&0'
K0(#)30%%0'
24
Ray Payloads
G.');"D'").$1&#$.$/);.#.),&#-)#-")$./
40%"F%"()3.88";)#-")LE"$)$./);.#.M
!.#.)3.')1")E.((";);0,')0$)6E)#-")$./)#$"")70$)10#-<
N6(#).)6("$O;"D'";)(#$63#).33"((";)1/).88)(-.;"$)E$0>$.%(
P.$&"()E"$)$./)#/E"
Attenuation
Color
Depth
importance
Color
Depth
importance
25
struct PerRayData_radiance
{
float3 result;
};
rtDeclareVariable(PerRayData_radiance, prd_radiance, rtPayload,);
rtDeclareVariable(float3, shading_normal, attribute shading_normal,);
RT_PROGRAM void closest_hit_radiance()
{
float3 worldnormal = normalize(rtTransformNormal(RT_OBJECT_TO_WORLD,
shading_normal));
prd_radiance.result = worldnormal * 0.5f + 0.5f;
}
26
Normal shader
27
for(int i = 0; i < lights.size(); ++i) {

BasicLight light = lights[i];
float3 L = normalize(light.pos - hit_point);
float nDl = dot( ffnormal, L);
if( nDl > 0.0f ){
// cast shadow ray
PerRayData_shadow shadow_prd;
shadow_prd.attenuation = 1.0f;
float Ldist = length(light.pos - hit_point);
Ray shadow_ray = make_Ray( hit_point, L, shadow_ray_type, scene_epsilon, Ldist );
rtTrace(top_shadower, shadow_ray, shadow_prd);
float light_attenuation = shadow_prd.attenuation;
if( light_attenuation > 0.0f ){
float3 Lc = light.color * light_attenuation;
color += Kd * nDl * Lc;
}
}
}
28
rtTextureSampler<float4, 2> envmap;

rtDeclareVariable(PerRayData_Radiance, prd_radiance, ,);
RT_PROGRAM void envmap_miss()
{
float theta = atan2f( ray.direction.x, ray.direction.z );
float phi
= M_PIf * 0.5f - acosf( ray.direction.y );
float u
= (theta + M_PIf) * (0.5f * M_1_PIf);
float v
= 0.5f * ( 1.0f + sin(phi) );
prd_radiance.result = make_float3( tex2D(envmap, u, v) );
}
29
Environment Maps
30
rtDeclareVariable(uint2, launchIndex, rtLaunchIndex,);

RT_PROGRAM void accumulation_camera()
{
...
float4 acc_val = accum_buffer[launch_index];
if( frame > 0 )
acc_val += make_float4(prd.result, 0.f);
else
acc_val = make_float4(prd.result, 0.f);
output_buffer[launch_index] = make_color( make_float3(acc_val) * 1.f/(float(frame+1)) );
accum_buffer[launch_index] = acc_val;
}
31
Accumulation Camera
32
Under the hood

.global .align 16 .b8 ray[36];
.global .align 4 .b8 _ZN21rti_internal_typeinfo3rayE[8] = {82,97,121,0,36,0,0,0};
.global .align 1 .b8 _ZN21rti_internal_typename3rayE[11] = {0x6f,0x70,0x74,0x69,0x78,0x3a,
0x3a,0x52,0x61,0x79,0x0};
.global .align 1 .b8 _ZN21rti_internal_semantic3rayE[13] =
{0x72,0x74,0x43,0x75,0x72,0x72,0x65,0x6e,0x74,0x52,0x61,0x79,0x0};
.global .align 1 .b8 _ZN23rti_internal_annotation3rayE[1] = {0x0};
...
ld.global.v4.f32
ld.global.v2.f32
{%f30,%f31,%f32,%f33}, [ray+0];
{%f38,%f39}, [ray+16];
mov %f30, ray_ox;

mov %f31, ray_oy;
mov %f33, ray_oz;
...
33
Under the hood

int3 vertex = vindex_buffer[primIdx];
ld.param.s32
%r1, [__cudaparm__Z14mesh_intersecti_primIdx];
cvt.s64.s32
%rd1, %r1;
mov.u64
%rd2, vindex_buffer;
mov.u32
%r2, 1;
mov.u32
%r4, 12;
mov.u64
%rd5, 0;
mov.u64
%rd7, 0;
mov.u64
%rd9, 0;
call (%rd11), _rt_buffer_get_64, (%rd2, %r2, %r4, %rd1, %rd5, %rd7, %rd9);
ld.global.s32
%r6, [%rd12+0];
ld.global.s32
%r7, [%rd12+4];
ld.global.s32
%r12, [%rd12+8];
34
Under the hood

int3 vertex = vindex_buffer[primIdx];
ld.param.s32
%r1, [__cudaparm__Z14mesh_intersecti_primIdx];
cvt.s64.s32
%rd1, %r1;
mov.u64
%rd2, vindex_buffer;
mov.u32
%r2, 1;
mov.u32
%r4, 12;
mov.u64
%rd5, 0;
mov.u64
%rd7, 0;
mov.u64
%rd9, 0;
call (%rd11), _rt_buffer_get_64, (%rd2, %r2, %r4, %rd1, %rd5, %rd7, %rd9);
ld.global.s32
%r6, [%rd12+0];
ld.global.s32
%r7, [%rd12+4];
ld.global.s32
%r12, [%rd12+8];
!"#$%&'$()*#+&)%#%&,%-*.-
34
Execution models
QEFR)E&E"8&'");0"()'0#)30%E8"#"8/)(E"3&*/).')":"36F0')%0;"8
4"S6"'F.8)30'(&(#"'3/),&#-&').)#-$".;)76(6.88/)E&:"8+(.%E8"<
T0)30'(&(#"'3/)1"#,""')#-$".;(
T0)0$;"$&'>)>6.$.'#""()*0$)&'#"$("3F0')#"(#(
4&;"O"A"3#()8&%&#";)#0)06#E6#)16A"$().';)$./)E./80.;(
K.'/)E0((&18")":"36F0')%0;"8(
/00123&41$&56786&*..19%:1.
35
Execution models
QEFR)E&E"8&'");0"()'0#)30%E8"#"8/)(E"3&*/).')":"36F0')%0;"8
4"S6"'F.8)30'(&(#"'3/),&#-&').)#-$".;)76(6.88/)E&:"8+(.%E8"<
T0)30'(&(#"'3/)1"#,""')#-$".;(
T0)0$;"$&'>)>6.$.'#""()*0$)&'#"$("3F0')#"(#(
4&;"O"A"3#()8&%&#";)#0)06#E6#)16A"$().';)$./)E./80.;(
K.'/)E0((&18")":"36F0')%0;"8(
/00123&41$&56786&*..19%:1.
;<&=%,&0%>.'?@.)("&AB&CDE/&#?$(%)
F<&8(G>(.:%0&("('>:1.&2*#?*.&(%'?&#?$(%)
35
Execution models
QEFR)E&E"8&'");0"()'0#)30%E8"#"8/)(E"3&*/).')":"36F0')%0;"8
4"S6"'F.8)30'(&(#"'3/),&#-&').)#-$".;)76(6.88/)E&:"8+(.%E8"<
T0)30'(&(#"'3/)1"#,""')#-$".;(
T0)0$;"$&'>)>6.$.'#""()*0$)&'#"$("3F0')#"(#(
4&;"O"A"3#()8&%&#";)#0)06#E6#)16A"$().';)$./)E./80.;(
K.'/)E0((&18")":"36F0')%0;"8(
/00123&41$&56786&*..19%:1.
;<&=%,&0%>.'?@.)("&AB&CDE/&#?$(%)
;<&=%,&0%>.'?@.)("&AB&CDE/&#?$(%)
F<&8(G>(.:%0&("('>:1.&2*#?*.&(%'?&#?$(%)
F<&!",0*'*#0H&I%.%-(&)*9($-(.'(&*.&%&=@=E&3#%#(&
I%'?*.(
35
Execution models
QEFR)E&E"8&'");0"()'0#)30%E8"#"8/)(E"3&*/).')":"36F0')%0;"8
4"S6"'F.8)30'(&(#"'3/),&#-&').)#-$".;)76(6.88/)E&:"8+(.%E8"<
T0)30'(&(#"'3/)1"#,""')#-$".;(
T0)0$;"$&'>)>6.$.'#""()*0$)&'#"$("3F0')#"(#(
4&;"O"A"3#()8&%&#";)#0)06#E6#)16A"$().';)$./)E./80.;(
K.'/)E0((&18")":"36F0')%0;"8(
/00123&41$&56786&*..19%:1.
;<&=%,&0%>.'?@.)("&AB&CDE/&#?$(%)
;<&=%,&0%>.'?@.)("&AB&CDE/&#?$(%)
;<&=%,&0%>.'?@.)("&AB&CDE/&#?$(%)
F<&8(G>(.:%0&("('>:1.&2*#?*.&(%'?&#?$(%)
F<&!",0*'*#0H&I%.%-(&)*9($-(.'(&*.&%&=@=E&3#%#(&
F<&!",0*'*#0H&I%.%-(&)*9($-(.'(&*.&%&=@=E&3#%#(&
I%'?*.(
I%'?*.(
J<&E(4($&21$K&2?(.&)%#%&.1#&,$(3(.#
35
Execution models
QEFR)E&E"8&'");0"()'0#)30%E8"#"8/)(E"3&*/).')":"36F0')%0;"8
4"S6"'F.8)30'(&(#"'3/),&#-&').)#-$".;)76(6.88/)E&:"8+(.%E8"<
T0)30'(&(#"'3/)1"#,""')#-$".;(
T0)0$;"$&'>)>6.$.'#""()*0$)&'#"$("3F0')#"(#(
4&;"O"A"3#()8&%&#";)#0)06#E6#)16A"$().';)$./)E./80.;(
K.'/)E0((&18")":"36F0')%0;"8(
/00123&41$&56786&*..19%:1.
;<&=%,&0%>.'?@.)("&AB&CDE/&#?$(%)
;<&=%,&0%>.'?@.)("&AB&CDE/&#?$(%)
;<&=%,&0%>.'?@.)("&AB&CDE/&#?$(%)
F<&8(G>(.:%0&("('>:1.&2*#?*.&(%'?&#?$(%)
;<@.*:%#(&$%H3&*.&0%$-(A*3?&L%#'?(3
F<&!",0*'*#0H&I%.%-(&)*9($-(.'(&*.&%&=@=E&3#%#(&
F<&!",0*'*#0H&I%.%-(&)*9($-(.'(&*.&%&=@=E&3#%#(&
F<81$#&1$&3'%.&41$&3#%#(&%.)&)%#%&'1?($(.'(
I%'?*.(
I%'?*.(
J<M$1'(33&L%#'?(3&>.:0&)1.(
J<&E(4($&21$K&2?(.&)%#%&.1#&,$(3(.#
35
Execution models
QEFR)E&E"8&'");0"()'0#)30%E8"#"8/)(E"3&*/).')":"36F0')%0;"8
4"S6"'F.8)30'(&(#"'3/),&#-&').)#-$".;)76(6.88/)E&:"8+(.%E8"<
T0)30'(&(#"'3/)1"#,""')#-$".;(
T0)0$;"$&'>)>6.$.'#""()*0$)&'#"$("3F0')#"(#(
4&;"O"A"3#()8&%&#";)#0)06#E6#)16A"$().';)$./)E./80.;(
K.'/)E0((&18")":"36F0')%0;"8(
/00123&41$&56786&*..19%:1.
;<&=%,&0%>.'?@.)("&AB&CDE/&#?$(%)
;<&=%,&0%>.'?@.)("&AB&CDE/&#?$(%)
;<&=%,&0%>.'?@.)("&AB&CDE/&#?$(%)
F<&8(G>(.:%0&("('>:1.&2*#?*.&(%'?&#?$(%)
;<@.*:%#(&$%H3&*.&0%$-(A*3?&L%#'?(3
F<&!",0*'*#0H&I%.%-(&)*9($-(.'(&*.&%&=@=E&3#%#(&
;<@.*:%#(&$%H3&#1&N00&2%$,
F<&!",0*'*#0H&I%.%-(&)*9($-(.'(&*.&%&=@=E&3#%#(&
F<81$#&1$&3'%.&41$&3#%#(&%.)&)%#%&'1?($(.'(
I%'?*.(
F<=%.%-(&1.A'?*,&G>(>(3&#1&,$1'(33&(%'?&3#%#(
I%'?*.(
J<M$1'(33&L%#'?(3&>.:0&)1.(
J<&E(4($&21$K&2?(.&)%#%&.1#&,$(3(.#
35
Compilation to state machine

{

{
Ray ray = make_ray();
save prd, index;
restore prd, index;
output_buffer[index] =

}
Optix Just-in Time Compiler

Inserts continuations
Transforms to state
machine
Rewrites variable load/
store for object model
Inlines intrinsic functions
36
Compilation to state machine

{

{
Optix Just-in Time Compiler
State 1
save prd, index;
restore prd, index;

}
Inserts continuations
Transforms to state
Statemachine
2
Rewrites variable load/
store for object model
Inlines intrinsic functions
36
Step 1: Compile to PTX

for( int i = 0; i < 5; ++i ) {
Ray ray = make_Ray( make_float3( i, 0, 0 ),
make_float3( 0, 0, 1 ),
0, 1e-4f, 1e20f );
UserPayloadStruct payload;
rtTrace( top_object, ray, payload );
}
37
Step 1: Compile to PTX

for( int i = 0; i < 5; ++i ) {
Ray ray = make_Ray( make_float3( i, 0, 0 ),
make_float3( 0, 0, 1 ),
0, 1e-4f, 1e20f );
UserPayloadStruct payload;
rtTrace( top_object, ray, payload );
}
ld.global.u32
%node, [top_object+0];
mov.s32
%i, 0;
loop:
call _rt_trace, ( %node, %i, 0, 0, 0, 0, 1,
0, 1e-4f, 1e20f,
payload );
%i, %i, 1;
nvcc add.s32
mov.u32
%iend, 5;
setp.ne.s32
%predicate, %i, %iend;
@%predicate bra loop;
37
Step 2: Insert continuations

ld.global.u32
mov.s32
%i, 0;
loop:
0, 1e-4f, 1e20f,
payload );
add.s32
%i, %i, 1;
mov.u32
%iend, 5;
setp.ne.s32
38
Step 2: Insert continuations

ld.global.u32
mov.s32
%i, 0;
loop:
0, 1e-4f, 1e20f,
payload );
add.s32
%i, %i, 1;
mov.u32
%iend, 5;
setp.ne.s32
ld.global.u32
mov.s32
%i, 0;
loop:
mov payload, %stack;
save %i, %iend, %node;
1e-4f, 1e20f,payload );
OptiX restore %i, %iend,0,%node;
add.s32
%i, %i, 1;
mov.u32
%iend, 5;
setp.ne.s32
38
Step 3: Apply optimizations

ld.global.u32
mov.s32
%i, 0;
loop:
0, 1e-4f, 1e20f,payload );
restore %i, %iend, %node;
add.s32
%i, %i, 1;
mov.u32
%iend, 5;
setp.ne.s32
39
Step 3: Apply optimizations

ld.global.u32
ld.const.u32
mov.s32
%i, 0;
mov.s32
%i, 0;
loop:
loop:
save %i;
0, 1e-4f, 1e20f,payload );OptiX
restore %i, %iend, %node;
restore %i;
add.s32
%i, %i, 1;
rematerialize %iend, %node;
mov.u32
%iend, 5;
add.s32
%i, %i, 1;
setp.ne.s32
mov.u32
%iend, 5;
setp.ne.s32
39
Step 4: Transform to state machine

ld.const.u32
mov.s32
%i, 0;
loop:
save %i;
restore %i;
add.s32
%i, %i, 1;
mov.u32
%iend, 5;
setp.ne.s32
40
Step 4: Transform to state machine

ld.const.u32
state 1:
mov.s32
%i, 0;
ld.const.u32
loop:
mov.s32
%i, 0;
loop:
save %i;
save %i;
0, 1e-4f, 1e20f,payload );OptiX bra mainloop;
restore %i;
state 2:
add.s32
%i, %i, 1;
restore %i;
mov.u32
%iend, 5;
setp.ne.s32
add.s32
%i, %i, 1;
mov.u32
%iend, 5;
setp.ne.s32
bra mainloop;
40
Step 5: Restore structured control flow

state 1:
ld.const.u32
mov.s32
%i, 0;
loop:
save %i;
bra mainloop;
state 2:
restore %i;
add.s32
%i, %i, 1;
mov.u32
%iend, 5;
setp.ne.s32
bra mainloop;
41
Step 5: Restore structured control flow

state 1:
ld.const.u32
mov.s32
%i, 0;
loop:
save %i;
bra mainloop;
state 2:
restore %i;
add.s32
%i, %i, 1;
mov.u32
%iend, 5;
setp.ne.s32
bra mainloop;
state 1:
ld.const.u32
mov.s32
%i, 0;
loop:
save %i;
OptiX bra mainloop;
loop_copy:
save %i;
bra mainloop;
state 2:
restore %i;
add.s32
%i, %i, 1;
mov.u32
%iend, 5;
setp.ne.s32
@%predicate bra loop_copy;
41
Graphical View
Initial
Transformed
State
Machine
begin
begin
State
Machine
Restored
begin
loop part 1
loop
loop part 1
loop part 2
loop part 2
loop part 1
42
Other optimizations
JIT Compiler gives opportunity for data-dependent optimizations
Elide unused transforms: up to 7%

Eliminate printing/exception code when not enabled: arbitrary
Reduce continuation size with rematerialization: arbitrary
Specialize traversal based on tree characteristics: 10-15%
Move small read-only data to textures or constant memory: up to 29%
Traditional compiler optimizations
Loop hoisting, dead-code elimination, copy propagation, etc.
Different load-balancing, scheduling, traversal
Architecture-dependent optimizations
43
Execution
Goal: execute efficiently on a GPU

Must manage execution coherence within a warp (32 threads)
Must manage data coherence
Minimize local state
Minimize context switch overhead
Strategy: compile to a megakernal

All programs (traversal, shading, intersection)
Utilize dynamic load-balancing
Some divergence unavoidable
Key observation
Although threads may temporarily diverge, they return to frequently used states
44
Fine-grained scheduling
Naive SIMD Schedule
Prioritized SIMD Schedule
13 iterations, 88% occupancy
A12121
B12123121212
B1212
C1212121231212
D121212121212
Time
2
3
1
1
2
2
1
1
C12121
D12121
2
Time
2
2
12
3
A1212
12
2
1
1
2
2
Thread
A1212121212
Thread
Thread
MIMD Schedule
1212
12
B121231212
12
C1212
121231212
D1 2 1 2
1212
1212
Time
45
ilt
in
fu
n
ct
io
pi
n
r
bu etu
nh il rn
_
o ti
pi le_ n e 0
n h c xi t
_
o a
pi le_ me 1
n h c ra
ol am _6
e_ e
ca ra _
m 7
er
m a_8
i n i ss
te
rs _ 1
in ec 0
te t_
r
1
in sec 2
te t_
rs 1
in ec 3
te t_
rs 1
in ec 4
te t_
rs 1
in ec 5
te t_
rs 1
in ec 6
te t_
rs 1
in ec 7
te t_
rs 1
in ec 9
te t_
rs 2
in ec 0
cl
os
te t_
rs 2
e
cl st_ int ec 1
os h
t
e
rs _ 2
e it
cl st_ _ra ec 3
os h di t_
2
a
e it
cl st_ _ra nce 4
os h di
_
a
2
e it
cl st_ _ra nce 5
os h di
_
i
es t_ an 26
t_ rad ce
h
_
an it_r ian 27
c
cl
o s y_ h a d i e _
2
e
i a
cl st_ t_s nce 8
os h ha _
i
e t
d 2
cl st_ _ra ow 9
os h di
a _3
e it
cl st_ _ra nce 0
os h di
a _3
e it
cl st_ _ra nce 1
os h di
_
es it_ an 32
t_ rad ce
hi
_3
a t ia
cl ny _ra nce 3
os _h di _
e
i a 3
cl st_ t_s nce 4
os h ha _
e it
d 3
cl st_ _ra ow 5
os h di
_
es it_ an 36
r
c
t
a
tra _h di e_
ve it_r an 37
c
tra rse adi e_
ve _ g a n 3 8
rs eo ce
e_ m _3
ge etr 9
om y_
et 43
ry
_4
4
bu
3.5E+06
3.0E+06
!"#$%&'()*#+(+,-%
.#/+#/0123%
2.5E+06
2.0E+06
1.5E+06
1.0E+06
5.0E+05
0.0E+00
46
ilt
in
fu
n
ct
io
pi
n
r
bu etu
nh il rn
_
o ti
pi le_ n e 0
n h c xi t
_
o a
pi le_ me 1
n h c ra
ol am _6
e_ e
ca ra _
m 7
er
m a_8
i n i ss
te
rs _ 1
in ec 0
te t_
r
1
in sec 2
te t_
rs 1
in ec 3
te t_
rs 1
in ec 4
te t_
rs 1
in ec 5
te t_
rs 1
in ec 6
te t_
rs 1
in ec 7
te t_
rs 1
in ec 9
te t_
rs 2
in ec 0
cl
os
te t_
rs 2
e
cl st_ int ec 1
os h
t
e
rs _ 2
e it
cl st_ _ra ec 3
os h di t_
2
a
e it
cl st_ _ra nce 4
os h di
_
a
2
e it
cl st_ _ra nce 5
os h di
_
i
es t_ an 26
t_ rad ce
h
_
an it_r ian 27
c
cl
o s y_ h a d i e _
2
e
i a
cl st_ t_s nce 8
os h ha _
i
e t
d 2
cl st_ _ra ow 9
os h di
a _3
e it
cl st_ _ra nce 0
os h di
a _3
e it
cl st_ _ra nce 1
os h di
_
es it_ an 32
t_ rad ce
hi
_3
a t ia
cl ny _ra nce 3
os _h di _
e
i a 3
cl st_ t_s nce 4
os h ha _
e it
d 3
cl st_ _ra ow 5
os h di
_
es it_ an 36
r
c
t
a
tra _h di e_
ve it_r an 37
c
tra rse adi e_
ve _ g a n 3 8
rs eo ce
e_ m _3
ge etr 9
om y_
et 43
ry
_4
4
bu
3.5E+06
3.0E+06
!"#$%&'()*#+(+,-%
5.0E+05
Divergence Before and After
.#/+#/0123%
2.5E+06
2.0E+06
1.5E+06
1.0E+06
Time Before and After
0.0E+00
46
Raw Traversal Performance

Aila-Laine
OptiX
Aila-Laine
Conference (Primary)
Conference (Primary)
Conference (AO)
Conference (AO)
Fairy Forest (Primary)
Fairy Forest (Primary)
Fairy Forest (AO)
Fairy Forest (AO)
Sibenik (Primary)
Sibenik (Primary)
Sibenik (AO)
Sibenik (AO)
50
100
150
GTX285
200
250
50
100
OptiX
150
200
250
GTX480
47
System performance (ISPM)

Frog Performance
150
OpenGL local
Trace
OpenGL Gather
I/O
Frame ms
120
90
60
30
OptiX
GTX480
McGuire-Luebke
Quad Core2 DUO
3.5X ray tracing speedup

2.5X net speedup
48
ISPM
ISPM
ISPM
Caustics
Diffuse Interreflection
2.8X
Color Bleeding
3.0X
OptiX Examples
49
Cook
Photon Map
Path tracer
Motion Blur and Depth of Field

14 programs
30 states
4648 bytes constant data
Progressive Photon Mapping

14 programs
30 states
Forward Path Tracing

11 programs
21 states
OptiX SDK Samples

50
Sample6
Julia
Tutorial
Ambient occlusion and .obj loading

10 programs
20 states
Fractals using custom intersection

20 programs
37 states
Programmable shading and intersection

17 programs
41 states
OptiX SDK Samples

51
Whirligig
Whitted
Collision
Animated scenes
14 programs
41 states
Interactive reconstruction of 1980 paper

20 programs
57 states
Line of sight and collision detection with

ray casting
13 programs
23 states
OptiX SDK Samples

52
Mandelbulb
Path tracer
Design Garage
Fractal intersection
Reconstruction of 1986 Kajiya paper
Progressive path tracing

28 programs
66 states
OptiX Examples
53
54
Questions?
55

Compilation For GPU Accelerated Ray Tracing in OptiX PDF

Hochgeladen von

Dokumentinformationen

Originalbeschreibung:

Originaltitel

Copyright

Verfügbare Formate

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Verfügbare Formate

Compilation For GPU Accelerated Ray Tracing in OptiX PDF

Hochgeladen von

Copyright:

Verfügbare Formate

Compilation for GPU

Accelerated Ray Tracing in OptiX

Tuesday, August 16, 2011

Lessons from history

Tuesday, August 16, 2011

RTSL: a Ray Tracing Shading

Tuesday, August 16, 2011

Convenient element access syntax:

Graphics oriented operators, functions

Tuesday, August 16, 2011

Tuesday, August 16, 2011

Deviation from GLSL

First class color type (not always 3/4 components)

Tuesday, August 16, 2011

Tuesday, August 16, 2011

Make GPU ray tracing simpler

User defined payloads

Dont require a new language

Leverage CUDA architecture and existing compiler infrastructure

Tuesday, August 16, 2011

Tuesday, August 16, 2011

RT_PROGRAM void pinhole_camera()

Tuesday, August 16, 2011

RT_PROGRAM void intersect_sphere()

RT_PROGRAM void closest_hit_radiance3()

Program objects (shaders)

Interconnection of programs defines the outcome

optix::Ray ray = optix::make_Ray

Input language is based on CUDA C/C++

Data associated with ray is programmable

Can also take raw PTX as input

Caveat: still need to use it responsibly to get performance

Tuesday, August 16, 2011

Abstract Execution Model

Tuesday, August 16, 2011

Tuesday, August 16, 2011

Tuesday, August 16, 2011

Tuesday, August 16, 2011

Tuesday, August 16, 2011

Geometry instance binds

Closest Hit/Any Hit pairs per raytype (user defined)

Tuesday, August 16, 2011

All objects can have multiple parents

Tree may have multiple roots

Not shown: Selector

Object variable inheritance

Tuesday, August 16, 2011

Material parameters can be

Variables are one of:

Object variable inheritance

Tuesday, August 16, 2011

Material parameters can be

Variables are one of:

Object variable inheritance

Tuesday, August 16, 2011

Material parameters can be

Variables are one of:

CUDA as a shading language

Use CUDA compiler (nvcc)

Tuesday, August 16, 2011

Tuesday, August 16, 2011

RT_PROGRAM void triangle_intersect(int)