[Solved] CAD2122 Project 1

$25

File Name: CAD2122_Project_1.zip
File Size: 160.14 KB

SKU: [Solved] CAD2122 Project 1 Category: Tag:
5/5 - (1 vote)

Image filters, like the blur filter used in lab 3, are examples of matrix processing that can be taken to the GPU. Consider two classes of filters: area filters, that considers an area around each pixel to define the final value of that pixel; and the point filters, that applies some function on the pixel value. For this project you need to implement a sequence of two filters with some parameters, that could be applied to any given image.

For the area filter, considering just the direct neighbors, we get a 33 grid of pixels with the original pixel at its center.

The final value for that pixel will be the weighted average defined by a 33 matrix of coefficients. The following examples of coefficients allows the implementation of several types of filters:

For the point filter, we pretend to gray the image using the following expression:

(r,g,b) = alpha*(c,c,c) + (1-alpha)(r,g,b), where c = 0.3 r + 0.59 g + 0.11 b Alfa is a value in [0 .. 1] that defines the color shift to grayscale.

A sequential code example is provided for reference. The main objective is to achieve the best performance, particularly for big images. For convenience, this work is presented in several stages (number 5 is mandatory):

  1. (30%) Implement a CUDA or OpenCL solution that
    1. Parallelizes both filter operations

Suggestion: experiment with different parallel strategies. Examples: just one kernel with both filters vs. one kernel per filter.

  1. Experiment with different grid layouts and block sizes.

Study if using a local shared area can improve your kernel(s) performance (nvprof and section 8 of CUDA Best Practices Guide[1] may help you evaluate any improvement).

  1. (20%) Evaluate these solutions against
    1. The sequential version
    2. Using different grid arrangements and thread block sizes
    3. Using/not using shared memory

Note: ignore file I/O times but include in your timings memcopy to/from device.

  1. (15%) Complement your solution with the ability to overlap computation with communication by partitioning the image space to process into a given number of partitions (NP). Assigning each of such partitions to a dedicated stream. Consider the following example with NP=3.

While kernel(s) is(are) being applied over Partition 2, Partition 3 can be simultaneously uploaded to the GPU and, possibly, the result of processing Partition 1 can be simultaneously downloaded from the GPU.

  1. (15%) Evaluate this solution against the others.
  1. (20%) Write a report (max of 5 pages A4 11pt font) that presents
    1. Tested approaches and final solution
    2. Relevant implementation details
    3. Your evaluation results (include times and/or graphs to compare and justify your solution)
    4. An analysis and interpretation of these results

Other relevant optimizations may also be accounted in the final grade.

[1] https://docs.nvidia.com/cuda/cuda-c-best-practices-guide/index.html#performance-metrics

Reviews

There are no reviews yet.

Only logged in customers who have purchased this product may leave a review.

Shopping Cart
[Solved] CAD2122 Project 1
$25