Pig breeding companies keep relatively small populations of pure sire and dam lines that are selected to improve the performance of crossbred animals. This design of the pig breeding industry presents challenges to the implementation of genomic selection, which requires large data sets to obtain highly accurate genomic breeding values. The objective of this study was to evaluate the impact of different reference sets (across population and multipopulation) on the accuracy of genomic breeding values in 3 purebred pig populations and to assess the potential of using crossbreed performance in genomic prediction. Data consisted of phenotypes and genotypes on animals from 3 purebred populations (sire line [SL] 1, = 1,146; SL2, = 682; and SL3, = 1,264) and 3 crossbred pig populations (Terminal cross [TER] 1, = 183; TER2, = 106; and TER3, = 177). Animals were genotyped using the Illumina Porcine SNP60 Beadchip. For each purebred population, within-, across-, and multipopulation predictions were considered. In addition, data from the paternal purebred populations were used as a reference set to predict the performance of crossbred animals. Backfat thickness phenotypes were precorrected for fixed effects and subsequently included in the genomic BLUP model. A genomic relationship matrix that accounted for the differences in allele frequencies between lines was implemented. Accuracies of genomic EBV obtained within the 3 different sire lines varied considerably. For within-population prediction, SL1 showed higher values (0.80) than SL2 (0.61) and SL3 (0.67). Multipopulation predictions had accuracies similar to within-population accuracies for the validation in SL1. For SL2 and SL3, the accuracies of multipopulation prediction were similar to the within-population prediction when the reference set was composed by 900 animals (600 of the target line plus 300 of another line). For across-population predictions, the accuracy was mostly close to zero. The accuracies of predicting crossbreed performance were similar for the 3 different crossbred populations (ranging from 0.25 to 0.29). In summary, the differences in accuracy of the within-population scenarios may be due to line divergences in heritability and genetic architecture of the trait. Within- and multipopulation predictions yield similar accuracies. Across-population prediction accuracy was negligible. The moderate accuracy of prediction of crossbreed performance appears to be a result of the relationship between the crossbreed and its parental lines.