oneMKL.MatrixCal Performance Test

library(oneMKL.MatrixCal)
library(data.table)
library(nanotime)

threadList <- 2^(0:5)
dimList <- 2^(8:13)
resDT_vanilla <- data.table(
  method = "Vanilla",
  dimension = dimList,
  thread = 1,
  time = rep(0, length(dimList)),
  replications = rep(0, length(dimList))
)
resDT_mkl <- data.table(
  method = "oneMKL.MatrixCal",
  dimension = rep(dimList, length(threadList)),
  thread = rep(threadList, each=length(dimList)),
  time = rep(0, length(threadList) * length(dimList)),
  replications = rep(0, length(threadList) * length(dimList))
)

for (i in seq_along(dimList)) {
  set.seed(100)
  
  d <- dimList[i]
  replication <- ifelse(d >= 2048, 10, 50)
  
  A <- matrix(rnorm(d*d), d, d)
  B <- matrix(rnorm(d), d, d/16)
  
  st <- nanotime(Sys.time())
  for (k in 1:replication) {
    A%*%B
    # solve(A, B)
  }
  et <- nanotime(Sys.time())
  resDT_vanilla[dimension == d, `:=`(time = as.integer(et - st) / 1000000 /replication, replications = replication)]
  
  for (t in threadList) {
    setMKLThreads(t)
    st <- nanotime(Sys.time())
    for (k in 1:replication) {
      fMatProd(A, B)
      # fMatSolve(A, B)
    }
    et <- nanotime(Sys.time())
    resDT_mkl[(dimension == d) & (thread == t), `:=`(time = as.integer(et - st) / 1000000 /replication, replications = replication)]
  }
}

fwrite(rbind(resDT_vanilla, resDT_mkl), "matrix_multiply_results.csv")
# fwrite(rbind(resDT_vanilla, resDT_mkl), "solve_results.csv")