This is going to be a C++ program, that computes the fancy Popcorn 2D fractal by Clifford Pickover.
INSTALL
mkdir release
cd release
cmake ..
make
./popcorn2d
mkdir debug
cd debug
cmake -DCMAKE_BUILD_TYPE=Debug ..
make
./popcorn2d
- computeImage kernel:
1.2 ms- 128 Threads, 256 Blocks
- colorImage kernel:
2.4 ms- memory resides on device
- uses shared memory, but seems to be inefficient compared to simple CUDA kernel below
nvprof --print-gpu-trace ./popcorn2d_kepler test 0 0 1 256 256 1
[...]
Start Duration Grid Size Block Size Regs* SSMem* DSMem* Size Throughput Device Context Stream Name
[...warmup...]
461.73ms 72.703us - - - - - 768.00KB 10.074GB/s Tesla K80 (0) 1 13 [CUDA memcpy HtoD]
461.83ms 1.2267ms (256 1 1) (128 1 1) 45 0B 0B - - Tesla K80 (0) 1 13 void computeImage<float>(float*, float, unsigned int, unsigned int)_105_gpu [40]
463.07ms 2.3851ms (256 1 1) (128 1 1) 43 0B 12B - - Tesla K80 (0) 1 13 void colorImage<float>(float*, unsigned int, unsigned int, unsigned int)_55_gpu [42]
465.48ms 72.032us - - - - - 768.00KB 10.168GB/s Tesla K80 (0) 1 13 [CUDA memcpy DtoH]
- 1D grid-striding for computeImage kernel:
1.05 ms- 128 Threads, 32*13 Blocks
- 1D grid-striding for colorImage kernel:
0.039 ms
$ nvprof --print-gpu-trace popcorn2d_cuda testc 0 0 1 256 256 1
[...]
Start Duration Grid Size Block Size Regs* SSMem* DSMem* Size Throughput Device Context Stream Name
[...warmup...]
444.67ms 11.872us - - - - - 768.00KB 61.693GB/s Tesla K80 (0) 1 7 [CUDA memset]
444.69ms 1.0457ms (416 1 1) (128 1 1) 45 0B 0B - - Tesla K80 (0) 1 7 void d_computeImage_1D<float>(float*, Parameters<float>) [124]
445.76ms 39.295us (416 1 1) (128 1 1) 42 0B 0B - - Tesla K80 (0) 1 7 void d_colorImage<float>(float*, Parameters<float>) [130]
- 2D grid-striding for computeImage kernel:
2.9 ms- 16x16 Threads, 32*13 Blocks
- 1D grid-striding for colorImage kernel:
0.039 ms
$ nvprof --print-gpu-trace popcorn2d_cuda testc 0 0 1 256 256 1
[...]
Start Duration Grid Size Block Size Regs* SSMem* DSMem* Size Throughput Device Context Stream Name
[...warmup...]
428.16ms 11.904us - - - - - 768.00KB 61.527GB/s Tesla K80 (0) 1 7 [CUDA memset]
428.17ms 2.9001ms (416 1 1) (16 16 1) 50 0B 0B - - Tesla K80 (0) 1 7 void d_computeImage<float>(float*, Parameters<float>) [124]
431.09ms 39.168us (416 1 1) (128 1 1) 42 0B 0B - - Tesla K80 (0) 1 7 void d_colorImage<float>(float*, Parameters<float>) [130]
- computeImage kernel:
20.3 ms
$ nvprof --print-gpu-trace ./popcorn2d_kepler test 0.1 0 1 1024 1024 1
[...]
Start Duration Grid Size Block Size Regs* SSMem* DSMem* Size Throughput Device Context Stream Name
[...warmup...]
489.29ms 1.0841ms - - - - - 12.000MB 10.810GB/s Tesla K80 (0) 1 13 [CUDA memcpy HtoD]
490.40ms 20.323ms (1024 1 1) (128 1 1) 45 0B 0B - - Tesla K80 (0) 1 13 void computeImage<float>(float*, float, unsigned int, unsigned int)_105_gpu [42]
510.75ms 25.289ms (1024 1 1) (128 1 1) 43 0B 12B - - Tesla K80 (0) 1 13 void colorImage<float>(float*, unsigned int, unsigned int, unsigned int)_55_gpu [44]
536.06ms 1.0977ms - - - - - 12.000MB 10.676GB/s Tesla K80 (0) 1 13 [CUDA memcpy DtoH]
- 1D grid-striding computeImage kernel:
16.3 ms
$ nvprof --print-gpu-trace ../release_cuda/popcorn2d_cuda testc 0.1 0 1 1024 1024 1
[...]
Start Duration Grid Size Block Size Regs* SSMem* DSMem* Size Throughput Device Context Stream Name
[...warmup...]
438.40ms 68.671us - - - - - 12.000MB 170.65GB/s Tesla K80 (0) 1 7 [CUDA memset]
438.47ms 16.297ms (416 1 1) (128 1 1) 45 0B 0B - - Tesla K80 (0) 1 7 void d_computeImage_1D<float>(float*, Parameters<float>) [124]
454.79ms 602.75us (416 1 1) (128 1 1) 42 0B 0B - - Tesla K80 (0) 1 7 void d_colorImage<float>(float*, Parameters<float>) [130]