Download as PDF, PPTX
![Af_Array An array of numbers is stored as an Af_Array object. This array is stored on GPU. [1] pry(main)> a = ArrayFire::Af_Array.new 2, [2,2],[1,2,3,4] No Name Array [2 2 1 1] Offsets: [0 0 0 0] Strides: [1 2 4 4] 1.0000 3.0000 2.0000 4.0000 => #<ArrayFire::Af_Array:0x000000020aeab8>](https://image.slidesharecdn.com/highperformancegpucomputingwithruby2-171101051115/75/High-performance-GPU-computing-with-Ruby-6-2048.jpg)
![[2] pry(main)> b = a + a No Name Array [2 2 1 1] Offsets: [0 0 0 0] Strides: [1 2 4 4] 2.0000 6.0000 4.0000 8.0000 => #<ArrayFire::Af_Array:0x000000020625c8>](https://image.slidesharecdn.com/highperformancegpucomputingwithruby2-171101051115/75/High-performance-GPU-computing-with-Ruby-7-2048.jpg)
![[3] pry(main)> b = a * a No Name Array [2 2 1 1] Offsets: [0 0 0 0] Strides: [1 2 4 4] 1.0000 9.0000 4.0000 16.0000 => #<ArrayFire::Af_Array:0x00000001fe6f90>](https://image.slidesharecdn.com/highperformancegpucomputingwithruby2-171101051115/75/High-performance-GPU-computing-with-Ruby-8-2048.jpg)
![VALUE arf_init(int argc, VALUE* argv, VALUE self) { afstruct* afarray; Data_Get_Struct(self, afstruct, afarray); dim_t ndims = (dim_t)NUM2LONG(argv[0]); dim_t* dimensions = (dim_t*)malloc(ndims * sizeof(dim_t)); dim_t count = 1; for (size_t index = 0; index < ndims; index++) { dimensions[index] = (dim_t)NUM2LONG(RARRAY_AREF(argv[1], index)); count *= dimensions[index]; } double* host_array = (double*)malloc(count * sizeof(double)); for (size_t index = 0; index < count; index++) { host_array[index] = (double)NUM2DBL(RARRAY_AREF(argv[2], index)); } af_create_array(&afarray->carray, host_array, ndims, dimensions, f64) return self; }](https://image.slidesharecdn.com/highperformancegpucomputingwithruby2-171101051115/75/High-performance-GPU-computing-with-Ruby-9-2048.jpg)
![=> #<ArrayFire::Af_Array:0x00000001591db0> [3] pry(main)> result = ArrayFire::BLAS.matmul(left, right, :AF_MAT_NONE, :AF_MAT_NONE) No Name Array [3 2 1 1] -39.0000 -74.0000 68.0000 -17.0000 86.0000 118.0000](https://image.slidesharecdn.com/highperformancegpucomputingwithruby2-171101051115/75/High-performance-GPU-computing-with-Ruby-12-2048.jpg)
![vadd_kernel_src = <<-EOS extern "C" { __global__ void matSum(int *a, int *b, int *c) { int tid = blockIdx.x; if (tid < 100) c[tid] = a[tid] + b[tid]; } } EOS f = compile(vadd_kernel_src) puts f.path](https://image.slidesharecdn.com/highperformancegpucomputingwithruby2-171101051115/75/High-performance-GPU-computing-with-Ruby-26-2048.jpg)
Uploaded byPrasun Anand
High performance GPU computing with Ruby
Prasun Anand presented on using Ruby for high performance GPU computing. He discussed SciRuby's efforts to push Ruby for scientific computing and popular Ruby gems like NMatrix. He explained CUDA and OpenCL for parallel computing on GPUs and demonstrated ArrayFire's Af_Array for storing arrays on the GPU. He also covered BLAS, LAPACK, statistics, random number generation, and benchmarks comparing GPU performance to CPU. Finally, he discussed RbCUDA for running custom kernels, GPU arrays, and interfaces to CuBLAS, CuSolver and CuRand.
High performance GPU computing with Ruby
- 1. High Performance GPUComputing with Ruby Prasun Anand
- 2. About me ● SciRubyContributor ● Google Summer of Code 2016, 2017 ● Genenetwork project ● Ruby Grant 2017 ● Projects: ○ JRuby port of NMatrix ○ ArrayFire gem ○ RbCUDA
- 3. SciRuby has beentrying to push Ruby for scientific computing. Popular Rubygems: ● NMatrix ● Daru ● Mixed_models ● Nyaplot ● Ipython Notebook SciRuby
- 4. CUDA and OpenCL CUDAis a parallel computing platform and programming model limited to NVIDIA hardware. OpenCL (Open Computing Language) is supported across all GPU hardware.
- 6. Af_Array An array ofnumbers is stored as an Af_Array object. This array is stored on GPU. [1] pry(main)> a = ArrayFire::Af_Array.new 2, [2,2],[1,2,3,4] No Name Array [2 2 1 1] Offsets: [0 0 0 0] Strides: [1 2 4 4] 1.0000 3.0000 2.0000 4.0000 => #<ArrayFire::Af_Array:0x000000020aeab8>
- 7. [2] pry(main)> b= a + a No Name Array [2 2 1 1] Offsets: [0 0 0 0] Strides: [1 2 4 4] 2.0000 6.0000 4.0000 8.0000 => #<ArrayFire::Af_Array:0x000000020625c8>
- 8. [3] pry(main)> b= a * a No Name Array [2 2 1 1] Offsets: [0 0 0 0] Strides: [1 2 4 4] 1.0000 9.0000 4.0000 16.0000 => #<ArrayFire::Af_Array:0x00000001fe6f90>
- 9. VALUE arf_init(int argc,VALUE* argv, VALUE self) { afstruct* afarray; Data_Get_Struct(self, afstruct, afarray); dim_t ndims = (dim_t)NUM2LONG(argv[0]); dim_t* dimensions = (dim_t*)malloc(ndims * sizeof(dim_t)); dim_t count = 1; for (size_t index = 0; index < ndims; index++) { dimensions[index] = (dim_t)NUM2LONG(RARRAY_AREF(argv[1], index)); count *= dimensions[index]; } double* host_array = (double*)malloc(count * sizeof(double)); for (size_t index = 0; index < count; index++) { host_array[index] = (double)NUM2DBL(RARRAY_AREF(argv[2], index)); } af_create_array(&afarray->carray, host_array, ndims, dimensions, f64) return self; }
- 10. static VALUE arf_ew_add(VALUEleft_val, VALUE right_val) { afstruct* left; afstruct* right; afstruct* result = ALLOC(afstruct); Data_Get_Struct(left_val, afstruct, left); Data_Get_Struct(right_val, afstruct, right); af_add(&result->carray, left->carray, right->carray, true); return Data_Wrap_Struct(CLASS_OF(left_val), NULL, arf_free, result); }
- 11. BLAS and LAPACK BLASfunctionalities ● matmul ● Transpose LAPACK functionalities ● det ● inverse ● norm ● qr ● cholesky ● svd ● lu
- 12. => #<ArrayFire::Af_Array:0x00000001591db0> [3] pry(main)>result = ArrayFire::BLAS.matmul(left, right, :AF_MAT_NONE, :AF_MAT_NONE) No Name Array [3 2 1 1] -39.0000 -74.0000 68.0000 -17.0000 86.0000 118.0000
- 13. Statistics ● Mean ● Median ●Variance
- 14. Random Engine Random numbergenerators Types of Engines: ● :AF_RANDOM_ENGINE_PHILOX_4X32_10 ● :AF_RANDOM_ENGINE_THREEFRY_2X32_16 ● :AF_RANDOM_ENGINE_MERSENNE_GP11213
- 15. Benchmarks ● AMD FX8350 octacore processor ● Nvidia GTX 750Ti GPU ● CUDA backend ● Double dtype
- 17. 10,000 X Faster thanNMatrix-Ruby
- 20. 100,000 X Faster thanNMatrix-Ruby-BLAS
- 22. 10 X Faster thanNMatrix-Ruby-Lapack
- 24.
- 25. Custom Kernels Scientific softwarerequire custom kernel code that suites to its needs. Run kernel from a Ruby file, on the fly. The kernel code is dynamically compiled Run on the GPU hardware to manipulate the array pointers.
- 26. vadd_kernel_src = <<-EOS extern"C" { __global__ void matSum(int *a, int *b, int *c) { int tid = blockIdx.x; if (tid < 100) c[tid] = a[tid] + b[tid]; } } EOS f = compile(vadd_kernel_src) puts f.path
- 27. GPU Array Generic pointerused to handle an array of elements on the GPU. Memory copying from CPU to GPU and vice-versa. Interfaced with NMatrix Interface with NArray
- 28. CuBLAS, CuSolver andCuRand ● BLAS routines ● Matrix Decomposition Routines ● Random Number Generators
- 29. Benchmarks ● AMD FX8350 octacore processor ● Nvidia GTX 750Ti GPU ● Double dtype
- 31. 1,000, 000 X Fasterthan NMatrix-Ruby-BLAS
- 32. Future Work ● ImageProcessing APIs and Indexers ● Multiple dtypes ● RbCUDA is under active development. ● Project RbCUDA is being funded by Ruby Association ● (Ruby Grant 2017)
- 33.
- 34. Acknowledgements 1. Pjotr Prins 2.Pradeep Garigipati
- 36. Thank You Github: prasunanand Twitter:@prasun_anand Blog: prasunanand.com