Evolving GPU Machine Code
Authors: Cleomar Pereira da Silva, Douglas Mota Dias, Cristiana Bentes, Marco Aurélio Cavalcanti Pacheco, Leandro Fontoura Cupertino
JMLR 2015 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We obtained up to 2.74 trillion GP operations per second for the 20-bit Boolean Multiplexer benchmark. We also compared our approach with the other three GPU-based acceleration methodologies implemented for quantum-inspired linear GP. Significant gains in performance were obtained. (...) 7. Experiments and Results |
| Researcher Affiliation | Academia | Cleomar Pereira da Silva EMAIL Department of Electrical Engineering Pontifical Catholic University of Rio de Janeiro (PUC-Rio) Rio de Janeiro, RJ 22451-900, Brazil Department of Education Development Federal Institute of Education, Science and Technology Catarinense (IFC) Videira, SC 89560-000, Brazil (...) Leandro Fontoura Cupertino EMAIL Toulouse Institute of Computer Science Research (IRIT) University of Toulouse 118 Route de Narbonne F-31062 Toulouse Cedex 9, France |
| Pseudocode | Yes | Algorithm 1: Pseudo-code for the GP interpreter for a GPU based on quantum-inspired LGP. |
| Open Source Code | No | The paper describes the methodology and implementation details (e.g., "Our procedure creates a PTX program...") but does not provide any explicit statement about releasing the source code or a link to a code repository for their GMGP implementation. |
| Open Datasets | Yes | We used five widely used GP benchmarks: two symbolic regression problems, Mexican Hat and Salutowicz; one time-series forecasting problem, Mackey-Glass; one image processing problem, Sobel filter; and one Boolean regression problem, 20-bit Multiplexer. (...) The Mackey-Glass benchmark (Jang and Sun, 1993) is a chaotic time-series prediction benchmark (...) We used six 512 × 512 images taken from the USC-SIPI image repository (Weber, 1997). (...) The 20-bit Boolean Multiplexer benchmark (Langdon, 2010b, 2011). |
| Dataset Splits | Yes | This sampling generates the training, validation, and testing data sets. (...) Figure 10 presents the three gray-scale images used for training. Figure 11 shows the two images used for validation. Figure 12 shows, for the same image, the original image in gray scale, the resulting image after the Sobel filter is applied by the GIMP tool, and the output image produced by the GMGP evolved filter. |
| Hardware Specification | Yes | The GPU used in our experiments was the Ge Force GTX TITAN. This processor has 2,688 CUDA cores (at 837 MHz) and 6 GB of RAM (no ECC) with a memory bandwidth of 288.4 GB/s through a 384-bit data bus. (...) GMGP creates the individuals on CPU using a single-threaded code running on a single core of an Intel Xeon CPU X5690 processor, with 32 KB of L1 data cache, 1.5 M of L2 cache, 12 MB of L3 cache, and 24 GB of RAM, running at 3.46 GHz. |
| Software Dependencies | Yes | The GP methodologies were implemented in C, CUDA 5.5, and PTX 3.2. The compilers used were gcc 4.4.7, nvcc release 5.5, V5.5.0, and ptxas release 5.5, V5.5.0. |
| Experiment Setup | Yes | Table 5 presents the parameters used when executing the Mexican Hat and Salutowicz benchmarks. (...) Table 9: Parameter settings for the Mackey-Glass benchmark. (...) Table 11: Parameter settings for the Sobel filter. (...) Table 14: Parameter settings for the 20-bit Boolean Multiplexer benchmark. |