Update volume rendering workload: use AVX, remove reduce_equal() path.
Both of these changes gave a performance benefit!
This commit is contained in:
@@ -46,8 +46,8 @@ also included in the ``examples/`` directory.)
|
||||
- 4.05x
|
||||
- 15.53x
|
||||
* - `Volume Rendering`_
|
||||
- 3.11x
|
||||
- 15.80x
|
||||
- 3.60x
|
||||
- 17.53x
|
||||
|
||||
|
||||
.. _AOBench: https://github.com/ispc/ispc/tree/master/examples/aobench
|
||||
|
||||
@@ -8,10 +8,10 @@ TASK_OBJ=$(addprefix objs/, $(subst ../,, $(TASK_CXX:.cpp=.o)))
|
||||
CXX=g++
|
||||
CXXFLAGS=-Iobjs/ -O3 -Wall -m64
|
||||
ISPC=ispc
|
||||
ISPCFLAGS=-O2 --target=sse2,sse4-x2 --arch=x86-64
|
||||
ISPCFLAGS=-O2 --target=sse2,sse4-x2,avx --arch=x86-64
|
||||
|
||||
OBJS=objs/volume.o objs/volume_serial.o $(TASK_OBJ) objs/volume_ispc.o \
|
||||
objs/volume_ispc_sse2.o objs/volume_ispc_sse4.o
|
||||
objs/volume_ispc_sse2.o objs/volume_ispc_sse4.o objs/volume_ispc_avx.o
|
||||
|
||||
default: volume
|
||||
|
||||
@@ -34,5 +34,5 @@ objs/%.o: ../%.cpp
|
||||
|
||||
objs/volume.o: objs/volume_ispc.h
|
||||
|
||||
objs/%_ispc.h objs/%_ispc.o objs/%_ispc_sse2.o objs/%_ispc_sse4.o: %.ispc
|
||||
objs/%_ispc.h objs/%_ispc.o objs/%_ispc_sse2.o objs/%_ispc_sse4.o objs/%_ispc_avx.o: %.ispc
|
||||
$(ISPC) $(ISPCFLAGS) $< -o objs/$*_ispc.o -h objs/$*_ispc.h
|
||||
|
||||
@@ -124,24 +124,13 @@ static inline float D(int x, int y, int z, uniform int nVoxels[3],
|
||||
}
|
||||
|
||||
|
||||
static inline float Du(uniform int x, uniform int y, uniform int z,
|
||||
uniform int nVoxels[3], uniform float density[]) {
|
||||
x = clamp(x, 0, nVoxels[0]-1);
|
||||
y = clamp(y, 0, nVoxels[1]-1);
|
||||
z = clamp(z, 0, nVoxels[2]-1);
|
||||
|
||||
return density[z*nVoxels[0]*nVoxels[1] + y*nVoxels[0] + x];
|
||||
}
|
||||
|
||||
|
||||
static inline float3 Offset(float3 p, float3 pMin, float3 pMax) {
|
||||
return (p - pMin) / (pMax - pMin);
|
||||
}
|
||||
|
||||
|
||||
static inline float Density(float3 Pobj, float3 pMin, float3 pMax,
|
||||
uniform float density[], uniform int nVoxels[3],
|
||||
uniform bool &checkForSameVoxel) {
|
||||
uniform float density[], uniform int nVoxels[3]) {
|
||||
if (!Inside(Pobj, pMin, pMax))
|
||||
return 0;
|
||||
// Compute voxel coordinates and offsets for _Pobj_
|
||||
@@ -153,39 +142,14 @@ static inline float Density(float3 Pobj, float3 pMin, float3 pMax,
|
||||
float dx = vox.x - vx, dy = vox.y - vy, dz = vox.z - vz;
|
||||
|
||||
// Trilinearly interpolate density values to compute local density
|
||||
float d00, d10, d01, d11;
|
||||
uniform int uvx, uvy, uvz;
|
||||
if (checkForSameVoxel && reduce_equal(vx, &uvx) && reduce_equal(vy, &uvy) &&
|
||||
reduce_equal(vz, &uvz)) {
|
||||
// If all of the program instances are inside the same voxel, then
|
||||
// we'll call the 'uniform' variant of the voxel density lookup
|
||||
// function, thus doing a single load for each value rather than a
|
||||
// gather.
|
||||
d00 = Lerp(dx, Du(uvx, uvy, uvz, nVoxels, density),
|
||||
Du(uvx+1, uvy, uvz, nVoxels, density));
|
||||
d10 = Lerp(dx, Du(uvx, uvy+1, uvz, nVoxels, density),
|
||||
Du(uvx+1, uvy+1, uvz, nVoxels, density));
|
||||
d01 = Lerp(dx, Du(uvx, uvy, uvz+1, nVoxels, density),
|
||||
Du(uvx+1, uvy, uvz+1, nVoxels, density));
|
||||
d11 = Lerp(dx, Du(uvx, uvy+1, uvz+1, nVoxels, density),
|
||||
Du(uvx+1, uvy+1, uvz+1, nVoxels, density));
|
||||
}
|
||||
else {
|
||||
// Otherwise, we have to do an actual gather in the more general
|
||||
// D() function. Once the reduce_equal tests above fail, we stop
|
||||
// checking in subsequent steps, since it's unlikely that this will
|
||||
// be true in the future once they've diverged into different
|
||||
// voxels.
|
||||
checkForSameVoxel = false;
|
||||
d00 = Lerp(dx, D(vx, vy, vz, nVoxels, density),
|
||||
D(vx+1, vy, vz, nVoxels, density));
|
||||
d10 = Lerp(dx, D(vx, vy+1, vz, nVoxels, density),
|
||||
D(vx+1, vy+1, vz, nVoxels, density));
|
||||
d01 = Lerp(dx, D(vx, vy, vz+1, nVoxels, density),
|
||||
D(vx+1, vy, vz+1, nVoxels, density));
|
||||
d11 = Lerp(dx, D(vx, vy+1, vz+1, nVoxels, density),
|
||||
D(vx+1, vy+1, vz+1, nVoxels, density));
|
||||
}
|
||||
float d00 = Lerp(dx, D(vx, vy, vz, nVoxels, density),
|
||||
D(vx+1, vy, vz, nVoxels, density));
|
||||
float d10 = Lerp(dx, D(vx, vy+1, vz, nVoxels, density),
|
||||
D(vx+1, vy+1, vz, nVoxels, density));
|
||||
float d01 = Lerp(dx, D(vx, vy, vz+1, nVoxels, density),
|
||||
D(vx+1, vy, vz+1, nVoxels, density));
|
||||
float d11 = Lerp(dx, D(vx, vy+1, vz+1, nVoxels, density),
|
||||
D(vx+1, vy+1, vz+1, nVoxels, density));
|
||||
float d0 = Lerp(dy, d00, d10);
|
||||
float d1 = Lerp(dy, d01, d11);
|
||||
return Lerp(dz, d0, d1);
|
||||
@@ -221,10 +185,8 @@ transmittance(uniform float3 p0, float3 p1, uniform float3 pMin,
|
||||
float t = rayT0;
|
||||
float3 pos = ray.origin + ray.dir * rayT0;
|
||||
float3 dirStep = ray.dir * stepT;
|
||||
uniform bool checkForSameVoxel = true;
|
||||
while (t < rayT1) {
|
||||
tau += stepDist * sigma_t * Density(pos, pMin, pMax, density, nVoxels,
|
||||
checkForSameVoxel);
|
||||
tau += stepDist * sigma_t * Density(pos, pMin, pMax, density, nVoxels);
|
||||
pos = pos + dirStep;
|
||||
t += stepT;
|
||||
}
|
||||
@@ -268,9 +230,8 @@ raymarch(uniform float density[], uniform int nVoxels[3], Ray ray) {
|
||||
float t = rayT0;
|
||||
float3 pos = ray.origin + ray.dir * rayT0;
|
||||
float3 dirStep = ray.dir * stepT;
|
||||
uniform bool checkForSameVoxel = true;
|
||||
cwhile (t < rayT1) {
|
||||
float d = Density(pos, pMin, pMax, density, nVoxels, checkForSameVoxel);
|
||||
float d = Density(pos, pMin, pMax, density, nVoxels);
|
||||
|
||||
// terminate once attenuation is high
|
||||
float atten = exp(-tau);
|
||||
|
||||
@@ -156,18 +156,18 @@
|
||||
<ItemGroup>
|
||||
<CustomBuild Include="volume.ispc">
|
||||
<FileType>Document</FileType>
|
||||
<Command Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">ispc -O2 %(Filename).ispc -o $(TargetDir)%(Filename).obj -h $(TargetDir)%(Filename)_ispc.h --arch=x86 --target=sse2,sse4-x2
|
||||
<Command Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">ispc -O2 %(Filename).ispc -o $(TargetDir)%(Filename).obj -h $(TargetDir)%(Filename)_ispc.h --arch=x86 --target=sse2,sse4-x2,avx
|
||||
</Command>
|
||||
<Command Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">ispc -O2 %(Filename).ispc -o $(TargetDir)%(Filename).obj -h $(TargetDir)%(Filename)_ispc.h --target=sse2,sse4-x2
|
||||
<Command Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">ispc -O2 %(Filename).ispc -o $(TargetDir)%(Filename).obj -h $(TargetDir)%(Filename)_ispc.h --target=sse2,sse4-x2,avx
|
||||
</Command>
|
||||
<Outputs Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">$(TargetDir)%(Filename).obj;$(TargetDir)%(Filename)_sse2.obj;$(TargetDir)%(Filename)_sse4.obj;$(TargetDir)%(Filename)_ispc.h</Outputs>
|
||||
<Outputs Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">$(TargetDir)%(Filename).obj;$(TargetDir)%(Filename)_sse2.obj;$(TargetDir)%(Filename)_sse4.obj;$(TargetDir)%(Filename)_ispc.h</Outputs>
|
||||
<Command Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">ispc -O2 %(Filename).ispc -o $(TargetDir)%(Filename).obj -h $(TargetDir)%(Filename)_ispc.h --arch=x86 --target=sse2,sse4-x2
|
||||
<Outputs Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">$(TargetDir)%(Filename).obj;$(TargetDir)%(Filename)_sse2.obj;$(TargetDir)%(Filename)_sse4.obj;$(TargetDir)%(Filename)_avx.obj;$(TargetDir)%(Filename)_ispc.h</Outputs>
|
||||
<Outputs Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">$(TargetDir)%(Filename).obj;$(TargetDir)%(Filename)_sse2.obj;$(TargetDir)%(Filename)_sse4.obj;$(TargetDir)%(Filename)_avx.obj;$(TargetDir)%(Filename)_ispc.h</Outputs>
|
||||
<Command Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">ispc -O2 %(Filename).ispc -o $(TargetDir)%(Filename).obj -h $(TargetDir)%(Filename)_ispc.h --arch=x86 --target=sse2,sse4-x2,avx
|
||||
</Command>
|
||||
<Command Condition="'$(Configuration)|$(Platform)'=='Release|x64'">ispc -O2 %(Filename).ispc -o $(TargetDir)%(Filename).obj -h $(TargetDir)%(Filename)_ispc.h --target=sse2,sse4-x2
|
||||
<Command Condition="'$(Configuration)|$(Platform)'=='Release|x64'">ispc -O2 %(Filename).ispc -o $(TargetDir)%(Filename).obj -h $(TargetDir)%(Filename)_ispc.h --target=sse2,sse4-x2,avx
|
||||
</Command>
|
||||
<Outputs Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">$(TargetDir)%(Filename).obj;$(TargetDir)%(Filename)_sse2.obj;$(TargetDir)%(Filename)_sse4.obj;$(TargetDir)%(Filename)_ispc.h</Outputs>
|
||||
<Outputs Condition="'$(Configuration)|$(Platform)'=='Release|x64'">$(TargetDir)%(Filename).obj;$(TargetDir)%(Filename)_sse2.obj;$(TargetDir)%(Filename)_sse4.obj;$(TargetDir)%(Filename)_ispc.h</Outputs>
|
||||
<Outputs Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">$(TargetDir)%(Filename).obj;$(TargetDir)%(Filename)_sse2.obj;$(TargetDir)%(Filename)_sse4.obj;$(TargetDir)%(Filename)_avx.obj;$(TargetDir)%(Filename)_ispc.h</Outputs>
|
||||
<Outputs Condition="'$(Configuration)|$(Platform)'=='Release|x64'">$(TargetDir)%(Filename).obj;$(TargetDir)%(Filename)_sse2.obj;$(TargetDir)%(Filename)_sse4.obj;$(TargetDir)%(Filename)_avx.obj;$(TargetDir)%(Filename)_ispc.h</Outputs>
|
||||
</CustomBuild>
|
||||
</ItemGroup>
|
||||
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
|
||||
|
||||
Reference in New Issue
Block a user