Team Sonic
Contents
Team Sonic
Members
About
Required Files
Progress
Assignment 1
Daniel
Topic
I'm interested in profiling and parallelizing an algorithm that finds the average value of an array of integral values. Although it might not sound exciting to some, I think it's a great starting point.
Updates
Feb 6: I created a program that does what I specified. I profiled it, and the profile showed that my application spends most of its time in the init_and_avg() function. That function has a loop that can be parallelized because each iteration is independent.
Prasanth
Topic
Find open source tomography algorithm, profile it, then attempt to convert parts of source code to take advantage of GPU using CUDA.
Update
Project
I have found an open source project called RabbitCT which aims to support the creation of efficient algorithms to process voxel data. The ultimate goal is to improve the performance of 3-D cone beam reconstruction, which is used extensively in the medical field to processing cone bean computed tomography. "Cone bean computed tomography (CT) is an emerging imaging technology. It provides all projections needed for three-dimensional (3D) reconstruction in a single spin of the X- ray source-detector pair. This facilitates fast, low-dose data acquisition, which is required for the imaging of rapidly moving objects, such as the human heart, as well as for intra-operative CT applications." [Klaus Mueller]
Profiling
After many many(!) issues, I was able to profile the reconstruction algorithm on my laptop.
RabbitCT provides three files:
- RabbitCTRunner - the application that accepts the data-set and algorithm module and gives some basic metrics of algorithm performance (i.e. elapsed time)
- RabbitCT Dataset V2 - platform independent data-set of actual C-arm scan of rabbit (~2.5Gb for "small" ver)
- LolaBunny - a basic non-optimized reference algorithm module (shared library)
Since the main processing is offloaded to the shared library (or dll on windows) I can't simply compile and run gprof on the main application (RabbitCTRunner), I had to set some special settings in shell to profile a shared library and use sprof; gprof can not be used for profiling shared libraries (found this out the hard long way). After many failed attempts these following steps allowed me to get a flat profile of the module.
Steps to reproduce:
- run cmake and create executables RabbitCTRunner and the share library LolaBunny
- $LD_PROFILE=libLolaBunny.so ./RabbitCTRunner [LOCATION OF MODULE] [LOCATION OF DATA-SET] [REPORT FILE] [VOLUMLE SIZE]
Example: LD_PROFILE=libLolaBunny.so ./RabbitCTRunner ../modules/LolaBunny/libLolaBunny.so ~/datasets/rabbitct_512-v2.rctd ./resultFile 128 - $ sprof -p [LOCATION OF MODULE] [LOCATION OF PROFILE FILE] > log
Example: sprof -p ../modules/LolaBunny/libLolaBunny.so /var/tmp/libLolaBunny.so.profile > log
Note. /var/tmp/ was my default location for profiles. Man ld.so for LD_PROFILE_EXPORT
You should now have a log file in your current dir, which has a flat-profile of the module: (Note. This was run with the smallest possible volume size, running it at 512 or higher takes much much longer)
[prasanth@localhost RabbitCTRunner]$ cat log Flat profile: Each sample counts as 0.01 seconds. % cumulative self self total time seconds seconds calls us/call us/call name 100.00 112.58 112.58 0 0.00 RCTAlgorithmBackprojection
This is the output from RabbitCTRunner:
[prasanth@localhost RabbitCTRunner]$ LD_PROFILE=libLolaBunny.so ./RabbitCTRunner ../modules/LolaBunny/libLolaBunny.so ~/datasets/rabbitct_512-v2.rctd ./resultFile 128 RabbitCT runner http://www.rabbitct.com/ Info: using 4 buffer subsets with 240 projections each. Running ... this may take some time. (\_/) (='.'=) (")_(") -------------------------------------------------------------- Quality of reconstructed volume: Root Mean Squared Error: 38914.3 HU Mean Squared Error: 1.51433e+09 HU^2 Max. Absolute Error: 65535 HU PSNR: -19.5571 dB -------------------------------------------------------------- Runtime statistics: Total: 112.627 s Average: 117.32 ms
Summary
With the flat-profile data, I can say with some certainty that 100% of the time is spent on the 'RCTAlgorithmBackprojection' method. Digging in to the source code this is the actual code of this method:
LolaBunny.cpp
FNCSIGN bool RCTAlgorithmBackprojection(RabbitCtGlobalData* r) { unsigned int L = r->L; float O_L = r->O_L; float R_L = r->R_L; double* A_n = r->A_n; float* I_n = r->I_n; float* f_L = r->f_L; s_rcgd = r; for (unsigned int k=0; k<L; k++) { double z = O_L + (double)k * R_L; for (unsigned int j=0; j<L; j++) { double y = O_L + (double)j * R_L; for (unsigned int i=0; i<L; i++) { double x = O_L + (double)i * R_L; double w_n = A_n[2] * x + A_n[5] * y + A_n[8] * z + A_n[11]; double u_n = (A_n[0] * x + A_n[3] * y + A_n[6] * z + A_n[9] ) / w_n; double v_n = (A_n[1] * x + A_n[4] * y + A_n[7] * z + A_n[10]) / w_n; f_L[k * L * L + j * L + i] += (float)(1.0 / (w_n * w_n) * p_hat_n(u_n, v_n)); } } } return true; }
We can see there is a nested for loop, containing three for loops. In Big-O notation the order of growth for this method would be O(N3). It is also using double precision and matrix multiplications, therefore I think this code can be optimized using CUDA. These results here were calculated on a:
Lenovo T400 laptop Intel® Core™2 Duo CPU P8600 @ 2.40GHz × 2 4GB 1066 MHz Memory Fedora Release 18 (Spherical Cow) 64-bit Kernel: 3.6.10-4.fc18.x86_64
I'm sure running this on our GTX 480 can yield better results. (hopefully).
Assignment 2
Background
In assignment 1 we identified the 'RCTAlgorithmBackprojection' function located in the dll/shared library was taking up virtually 100% of the execution time. Our goal for this assignment was to off-load that logic to the GPU using cuda api, leaving everything else intact.
Process
The work we did can be separated in to three basic steps:
- Compile base software from source (RabbitCTRunner.exe and dll)
- Write cuda kernel to replace 'RCTAlgorithmBackprojection' in dll
- Integrate cuda kernel in to RabbitCTRunner
Using preprocessor directives the RabbitCTRunner was fitted with a token named GPU. If GPU was defined, our kernel along with all the necessary code will be compiled and used by RabbitCTRunner. This allowed us to have a small footprint on the changes applied and also allows us to use the flag to compile back the original RabbitCTRunner.
Kernel
__global__ void Backprojection(float * gpu_matrix_result, RabbitCtGlobalData * r) // problem_size is r.L { unsigned int L = r->L; float O_L = r->O_L; float R_L = r->R_L; double* A_n = r->A_n; //float* I_n = r->I_n; //float* f_L = r->f_L; // for optimization, put these ^ in shared memory. //s_rcgd = r; unsigned int col = blockDim.x * blockIdx.x + threadIdx.x; // this is like the original "i" iterator unsigned int row = blockDim.y * blockIdx.y + threadIdx.y; // this is like the original "j" iterator unsigned int depth = blockDim.z * blockIdx.z + threadIdx.z; // this is like the original "k" iterator if (row < L && col < L && depth < L) { unsigned int final_index = L * L * depth + L * row + col; double z = O_L + (double)depth * R_L; double y = O_L + (double)row * R_L; double x = O_L + (double)col * R_L; double w_n = A_n[2] * x + A_n[5] * y + A_n[8] * z + A_n[11]; double u_n = (A_n[0] * x + A_n[3] * y + A_n[6] * z + A_n[9] ) / w_n; // inline this double v_n = (A_n[1] * x + A_n[4] * y + A_n[7] * z + A_n[10]) / w_n; // inline this // p_hat_n inlined: int i = (int)u_n; int j = (int)v_n; double alpha = u_n - (int)u_n; double beta = v_n - (int)v_n; double p_hat_n_result = (1.0 - alpha) * (1.0 - beta) * p_n(i , j ) + alpha * (1.0 - beta) * p_n(i+1, j ) + (1.0 - alpha) * beta * p_n(i , j+1) + alpha * beta * p_n(i+1, j+1); /////// gpu_matrix_result[final_index] += (float)(1.0 / (w_n * w_n) * p_hat_n_result); // this call calculates the index value twice because of += } }
Memory Allocation
#ifdef GPU float * gpu_matrix; cudaMalloc((void**)&gpu_matrix, numel_vol * sizeof(float)); cudaMemcpy(gpu_matrix, rctgdata.f_L, numel_vol * sizeof(float), cudaMemcpyHostToDevice); const int threads_per_block = 16; const int num_of_blocks = (problem_size + threads_per_block - 1) / threads_per_block; dim3 block(threads_per_block, threads_per_block, threads_per_block); dim3 grid(num_of_blocks, num_of_blocks, num_of_blocks); #endif
Execution Config
#ifdef GPU Backprojection<<<grid, block>>>(gpu_matrix, &rctgdata); #else
Metrics
Original RabbitCTRunner with volume size 128:
C:\Users\pvaaheeswaran\Desktop\cuda rabbit>Rabbit.exe LolaBunny.dll rabbitct_512-v2.rctd original 128 RabbitCT runner http://www.rabbitct.com/ Info: using 4 buffer subsets with 240 projections each. Running ... this may take some time. (\_/) (='.'=) (")_(") -------------------------------------------------------------- Quality of reconstructed volume: Root Mean Squared Error: 38914.3 HU Mean Squared Error: 1.51433e+009 HU^2 Max. Absolute Error: 65535 HU PSNR: -19.5571 dB -------------------------------------------------------------- Runtime statistics: Total: 339.759 s Average: 353.915 ms FULL RUNTIME: 340.507 secs C:\Users\pvaaheeswaran\Desktop\cuda rabbit>
CudaRabbit, volume size 128.
C:\Users\pvaaheeswaran\Desktop\cuda rabbit>CudaRabbit.exe LolaBunny.dll rabbitct_512-v2.rctd result 128/ RabbitCT runner http://www.rabbitct.com/ Info: using 4 buffer subsets with 240 projections each. Running ... this may take some time. (\_/) (='.'=) (")_(") -------------------------------------------------------------- Quality of reconstructed volume: Root Mean Squared Error: 38914.3 HU Mean Squared Error: 1.51433e+009 HU^2 Max. Absolute Error: 65535 HU PSNR: -19.5571 dB -------------------------------------------------------------- Runtime statistics: Total: 0.004881 s Average: 0.00508437 ms FULL RUNTIME: 0.512 secs C:\Users\pvaaheeswaran\Desktop\cuda rabbit>
Graph
This graph represents for only 128 volume size.
Summary
As you can see, using cuda has improved the performance by ~99.85%, we went from over 5 minutes to under a second. However, this code still needs to be optimized and run on bigger volume sizes. This will be our next goal.