Created
August 29, 2017 03:14
-
-
Save ASvyatkovskiy/91a2652c3de35a96499266b69f0e8140 to your computer and use it in GitHub Desktop.
Test CUDA-aware mpi4py with pycuda gpuarrays
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| #!/usr/bin/env python | |
| from mpi4py import MPI | |
| import numpy as np | |
| import time | |
| import pycuda.autoinit | |
| from pycuda import gpuarray | |
| comm = MPI.COMM_WORLD | |
| comm_rank = comm.Get_rank() | |
| comm_size = comm.Get_size() | |
| sizes = 16000 | |
| print ("MPI comm_size {}".format(comm_size)) | |
| #define a float16 mpi datatype | |
| mpi_float16 = MPI.BYTE.Create_contiguous(2).Commit() | |
| MPI._typedict['e'] = mpi_float16 | |
| def sum_f16_cb(buffer_a, buffer_b, t): | |
| assert t == mpi_float16 | |
| array_a = np.frombuffer(buffer_a, dtype='float16') | |
| array_b = np.frombuffer(buffer_b, dtype='float16') | |
| array_b += array_a | |
| #create new OP | |
| mpi_sum_f16 = MPI.Op.Create(sum_f16_cb, commute=True) | |
| data = np.array([comm_rank] * sizes,dtype=np.float32) | |
| data_gpu = gpuarray.to_gpu(data) | |
| data_buf = data_gpu.gpudata.as_buffer(data_gpu.nbytes) | |
| result = np.empty_like(data) | |
| result_gpu = gpuarray.empty(data.shape, np.float32) | |
| result_buf = result_gpu.gpudata.as_buffer(result_gpu.nbytes) | |
| t1 = time.time() | |
| comm.Allreduce([data_buf,MPI.FLOAT], [result_buf,MPI.FLOAT], op=MPI.SUM) #mpi_sum_f16) | |
| #result_buf = comm.allreduce([data_buf,MPI.FLOAT],op=MPI.SUM) #mpi_sum_f16) | |
| t1 = time.time() - t1 | |
| final_data = np.array([data] * sizes,dtype=np.float32) | |
| final_data_gpu = gpuarray.to_gpu(final_data) | |
| final_data_buf = final_data_gpu.gpudata.as_buffer(final_data_gpu.nbytes) | |
| final_result = np.empty_like(final_data) | |
| final_result_gpu = gpuarray.empty(final_data.shape, np.float32) | |
| final_result_buf = final_result_gpu.gpudata.as_buffer(final_result_gpu.nbytes) | |
| t2 = time.time() | |
| comm.Allreduce([final_data_buf,MPI.FLOAT], [final_result_buf,MPI.FLOAT], op=MPI.SUM) #mpi_sum_f16) | |
| #final_result_buf = comm.allreduce([final_data_buf,MPI.SUM],op=MPI.SUM) #mpi_sum_f16) | |
| t2 = time.time() - t2 | |
| #result_gpu.get(result, result_gpu.gpudata) | |
| #print(result) | |
| #final_result_gpu.get(final_result, final_result_gpu.gpudata) | |
| #print(final_result) | |
| if comm_rank == 0: | |
| print ("Elapsed time {}".format(t1+t2)) |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment