PC Lab 6: Diferență între versiuni

De la WikiLabs
Linia 30: Linia 30:
  
 
Please read chapters 1,2, 4,5 (skip Ch3 which is very particular to CUDA)
 
Please read chapters 1,2, 4,5 (skip Ch3 which is very particular to CUDA)
'''GP-GPU Programming guide''' https://developer.download.nvidia.com/compute/DevZone/docs/html/C/doc/CUDA_C_Programming_Guide.pdf
+
['''GP-GPU Programming guide''' [https://developer.download.nvidia.com/compute/DevZone/docs/html/C/doc/CUDA_C_Programming_Guide.pdf]]
  
  

Versiunea de la data 26 aprilie 2018 16:31

Session 6

Task: run matrix-column normalization using OpenCL (https://www.khronos.org/opencl)

Matrix-column normalization means that, at the end of the process, every sum of squared elements on the same column is 1.

Example: Assuming matrix is

[ 1, 2 ] 
[ 3, 4 ]

the result of normalization is :

[ 0.3162     0.4472 ]
[ 0.9487     0.8944 ]

That is: 0.3162 * 0.3162 + 0.9487 * 0.9487 = 1 and of course, 0.3162 / 0.9487 is kept as 1 / 3 ratio That is: 0.4472 * 0.4472 + 0.8944 * 0.8944 = 1 and of course, 0.4472 / 0.8944 is kept as 2 / 4 ratio

  1. Install opencl drivers for your platform
  2. Check what opencl-capable devices with command clinfo
  3. Run the VectorAddOpenCL app [[1]] to see that all works ok
  4. Implement the normalization operation on a CPU, for reference.
  5. Implement the normalization operation across 1 OpenCL thread of a single device. Check the result.
  6. Implement the normalization operation across multiple OpenCL threads of the same device. Check the result.
  7. How much faster is the OpenCL op performed on all threads vs. 1 thread on the same Open CL device ?

Note In order to use the ACS GPGPU Cluster see Using ACS Cluster

Please read chapters 1,2, 4,5 (skip Ch3 which is very particular to CUDA) [GP-GPU Programming guide [2]]


Points (out of 10) vs. expected performance:


[[]]