Have you considered Salmon? It works on a similar methodology, and can output gene, and transcript level counts. Specific to your question, your methodology seems right, if you sum the TPMs of ensembl transcripts a, b, and c, for associated ensembl gene x, then that will work.

Edit: R example.

foo <- data.frame(gene=c(rep("A",3),
                         rep("B",2),
                         rep("C",1),
                         rep("D",4)),
                  transcript=c(paste0("A", 1:3),
                               paste0("B", 1:2),
                               paste0("C", 1),
                               paste0("D", 1:4)))
doo <- data.frame(SampleA = sample(1:100, 10),
                  SampleB = sample(1:100, 10),
                  SampleC = sample(1:100, 10))
rownames(doo) <- foo$transcript

out <- lapply(unique(foo$gene),
              function(x) {
                tmp       <- foo[foo$gene == x,]
                tmp_count <- doo[match(tmp$transcript,
                                       rownames(doo)),]
                tmp_out   <- colSums(tmp_count)
                return(tmp_out)
              })

gene_counts <- matrix(unlist(out), 
                      ncol  = ncol(doo), 
                      byrow = T)
rownames(gene_counts) <- unique(foo$gene)
colnames(gene_counts) <- colnames(doo)