1 Objective
The purpose of this lab is to implement a tiled dense matrix multiplication routine using shared memory.
2 Instructions
The code template in template.cu provides a starting point and handles the import and export as well as the checking of the solution. Students are expected to insert their code is demarcated with //@@.
Students are expected to leave the other code unchanged. Edit the skeleton code to perform the following:
• Allocate device memory
• Copy host memory to device
• Initialize thread block and grid dimensions
• Invoke CUDA kernel
• Copy results from device to host
• Free device memory
• Write the CUDA kernel
Compile the template with the provided Makefile. The executable generated as a result of compilation can be run using the following code:
./TiledGEMM Template -e <expected.raw> -i <input1.raw>,<input2.raw>
-o <output.raw> -t matrix
where <expected.raw> is the expected output, <input0.raw>,<input1.raw> is the input dataset,
and <output.raw> is an optional path to store the results.
README.md has details on how to build libgputk, template.cpp and the dataset generator.
3 What to Turn in
Submit a report that includes the following:
1. How many floating operations are being performed by your kernel?
2. How many global memory reads are being performed by your kernel?
3. How many global memory writes are being performed by your kernel?
4. Describe what further optimizations can be implemented to your kernel to achieve a performancespeedup.
5. Your version of template.cu.
6. Execution times of the kernel with the input data generated by the dataset generator (in a tableor graph). Please include the system information where you performed your evaluation. For time measurement, use gpuTKTime start and gpuTKTime stop functions (You can find details in libgputk/README.md).
7. Execution times of the kernel for 4096*8000 and 8000*512 input matrices with different tile widths(2, 4, 8, 12, 16, 24, 32). Please include the system information where you performed your evaluation. For time measurement, use gpuTKTime start and gpuTKTime stop functions (You can find details in libgputk/README.md).
