Tech-Demo: SVO Voxel Raytracing with Image Warping (4x Speedup)

[spacerat]

Summary The tech-demo consists of a sparse voxel octree raytracer that uses frame-to-frame coherency to speed up the raytracing. Pixels from one frame are projected into the next and only holes are raycasted, which leads to a speed up of up to 4x.

Download Demo/Zip (Note: you might need VS2012 redistributables )

Controls Mouse: look around, wasdqe: move around, space: just show reprojected pixels

Details The octree is a mixed normal / compact / linear node octree. The warping method uses 4 buffers to cache the most recent pixels on the screen. I have developed this algorithm already a year ago and finally found time to make a small demo for everyone to try. It uses OGL and OCL. So far it works well with NVIDIA, but I havent tried it on ATI or IntelHD yet.
Original Article : Blog Link / Further reading

 

[Jawed]

read successfully ../data/palette.bmp ; 16x16x32 Bit
EXE_DIR
>> 1 OpenCL platform(s) found:
  > 0
  PROFILE = FULL_PROFILE
  VERSION = OpenCL 2.0 AMD-APP (1642.5)
  NAME = AMD Accelerated Parallel Processing
  VENDOR = Advanced Micro Devices, Inc.
  EXTENSIONS = cl_khr_icd cl_khr_d3d10_sharing cl_khr_d3d11_sharing cl_khr_dx9_media_sharing cl_amd_event_callback cl_amd_offline_devices
> 0
>> Try platform 0 ( 2 device(s) )
  > 0
  DEVICE_NAME = Capeverde
  DEVICE_VENDOR = Advanced Micro Devices, Inc.
  DEVICE_VERSION = OpenCL 1.2 AMD-APP (1642.5)
  DRIVER_VERSION = 1642.5 (VM)
  DEVICE_MAX_COMPUTE_UNITS = 10
  DEVICE_MAX_CLOCK_FREQUENCY = 1000
  DEVICE_GLOBAL_MEM_SIZE = 1073741824
  CL_DEVICE_LOCAL_MEM_SIZE = 32768
  > 1
  DEVICE_NAME = Tahiti
  DEVICE_VENDOR = Advanced Micro Devices, Inc.
  DEVICE_VERSION = OpenCL 1.2 AMD-APP (1642.5)
  DRIVER_VERSION = 1642.5 (VM)
  DEVICE_MAX_COMPUTE_UNITS = 32
  DEVICE_MAX_CLOCK_FREQUENCY = 1000
  DEVICE_GLOBAL_MEM_SIZE = 3221225472
  CL_DEVICE_LOCAL_MEM_SIZE = 32768

Selected : platform 0 device -1

No compatible OpenCL Device found.

Callstack:
ocl_init [c:\code\old\game_voxelmaster_repro\src\ocl.h:143]

 

[fellix]

read successfully ../data/palette.bmp ; 16x16x32 Bit
EXE_DIR D:\User Folders\Desktop\SVO-Reprojection-TechDemo\bin32
>> 1 OpenCL platform(s) found:
  > 0
  PROFILE = FULL_PROFILE
  VERSION = OpenCL 1.2 CUDA 7.5.8
  NAME = NVIDIA CUDA
  VENDOR = NVIDIA Corporation
  EXTENSIONS = cl_khr_byte_addressable_store cl_khr_icd cl_khr_gl_sharing cl_nv_
compiler_options cl_nv_device_attribute_query cl_nv_pragma_unroll cl_nv_d3d9_sha
ring cl_nv_d3d10_sharing cl_khr_d3d10_sharing cl_nv_d3d11_sharing cl_nv_copy_opt
s
> 0
>> Try platform 0 ( 1 device(s) )
  > 0
  DEVICE_NAME = GeForce GTX 780 Ti
  DEVICE_VENDOR = NVIDIA Corporation
  DEVICE_VERSION = OpenCL 1.2 CUDA
  DRIVER_VERSION = 352.86
  DEVICE_MAX_COMPUTE_UNITS = 15
  DEVICE_MAX_CLOCK_FREQUENCY = 1084
  DEVICE_GLOBAL_MEM_SIZE = 3221225472
  CL_DEVICE_LOCAL_MEM_SIZE = 49152

Selected : platform 0 device 0

CL Compilation failed:
<kernel>:172:2: warning: backslash and newline separated by space
\
^
<kernel>:498:84: warning: use of logical '&&' with constant operand
                a_screenBuffer[ofs]=( (col1&0xfcfcfcfc)+(col2&0xfcfcfcfc)+(col3&
0xfcfcfcfc)+(col4&&0xfcfcfcfc))>>2;

                 ^ ~~~~~~~~~~
<kernel>:498:84: note: use '&' for a bitwise operation
                a_screenBuffer[ofs]=( (col1&0xfcfcfcfc)+(col2&0xfcfcfcfc)+(col3&
0xfcfcfcfc)+(col4&&0xfcfcfcfc))>>2;

                 ^~~~~~~~~~~~

                 &
<kernel>:498:84: note: remove constant to silence this warning
                a_screenBuffer[ofs]=( (col1&0xfcfcfcfc)+(col2&0xfcfcfcfc)+(col3&
0xfcfcfcfc)+(col4&&0xfcfcfcfc))>>2;

                 ^~~~~~~~~~~~
<kernel>:729:26: error: can't convert between vector values of different size ('
float3' and 'double')
        float3 raypos = a_m0.xyz*16.0;//+normalize(delta)*distance)*16.0;
                        ~~~~~~~~^~~~~
<kernel>:749:14: error: can't convert between vector values of different size ('
float3' and 'double')
                phit=raypos/16.0;
                     ~~~~~~^~~~~
<kernel>:845:26: error: can't convert between vector values of different size ('
float3' and 'double')
        float3 raypos = a_m0.xyz*16.0;
                        ~~~~~~~~^~~~~
<kernel>:866:14: error: can't convert between vector values of different size ('
float3' and 'double')
                phit=raypos/16.0;
                     ~~~~~~^~~~~
<kernel>:975:26: error: can't convert between vector values of different size ('
float3' and 'double')
        float3 raypos = a_m0.xyz*16.0;
                        ~~~~~~~~^~~~~
<kernel>:995:14: error: can't convert between vector values of different size ('
float3' and 'double')
                phit=raypos/16.0;
                     ~~~~~~^~~~~


Callstack:
ocl_compile [c:\code\old\game_voxelmaster_repro\src\ocl.h:50]

 

[lanek]

will try on my system, set for luxrender raytracing openCL 2.1

Edit: seems to compil normally.

2x HD7970 driver 15.4 beta, but i have the AMD APP SDK installed separetely of the driver.

 

[Kaarlisk]

On Haswell, it crashed... the Intel driver.
10.18.15.4204 on Win10.

 

[Arwin]

Runs fine on my GTX 970, can't get it below 160fps on the Imrod+Lucy one and wild movements can bring it down to about 100fps on the Trees demo.

 

[homerdog]

I'm away from the PC so not gonna try it on my laptop with Intel 3000 grafix .

Obviously this method relies on slow camera/object movements so it wouldn't suit a fast paced fps or anything like that. Still it could be useful for some things. Thanks for sharing

Question: In a static scene with no camera movement does the framerate→∞ ?