BenchTasty.hs

{-# LANGUAGE OverloadedStrings #-}
{-# OPTIONS_GHC -fno-full-laziness -fno-cse #-}
-- | tasty-bench based microbenchmarks for hot paths affected by
-- Phase 1-7 perf optimizations (-O2, StrictData, INLINE).
--
-- Run with:
--
-- > OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 \
-- >   cabal run bench-tasty -- --csv bench/results/tasty.csv
--
-- The CSV output is comparable across builds (same data fixtures and
-- single-thread BLAS). Use @--baseline=path.csv --fail-if-slower=N@ to
-- guard against regressions.
module Main where

import qualified Numeric.LinearAlgebra   as LA
import           Test.Tasty.Bench

import           Hanalyze.Model.Core              (coefficients)
import           Hanalyze.Model.GLM               (Family (..), LinkFn (..), fitGLMFull)
import qualified Hanalyze.Model.Regularized       as Reg
import           Hanalyze.Model.Regularized       (Penalty (..), rfBeta)
import qualified Hanalyze.Model.KernelRegression            as Kn
import qualified Hanalyze.Stat.Cholesky           as Chol
import qualified Hanalyze.Stat.KernelDist         as KD
import           Data.Maybe              (fromMaybe)

import           BenchUtil               (readCsvXY)

makeSpd :: Int -> LA.Matrix Double
makeSpd n =
  let g = LA.build (n, n)
            (\i j -> exp (-(((i - j) * (i - j)) / fromIntegral n)))
  in g + LA.scale 1e-3 (LA.ident n)

main :: IO ()
main = do
  (xLogi, yLogi) <- readCsvXY "bench/data/logistic_n10000_p20.csv"
  (xKR,   yKR)   <- readCsvXY "bench/data/kernel_n1000_p5.csv"
  (xLas,  yLas)  <- readCsvXY "bench/data/lm_n10000_p50.csv"
  let yKRMat   = LA.asColumn yKR
      spd500   = makeSpd 500
      spdRhs   = LA.asColumn (LA.fromList (replicate 500 1.0))
      kdInput2k = LA.fromLists
        [[fromIntegral i + 0.1 * fromIntegral j | j <- [0 .. 4]] | i <- [0 .. 1999]]

  defaultMain
    [ bgroup "regression"
        [ bench "GLM_logit_n10000_p20" $
            nf (\() -> LA.sumElements
                         (coefficients (fst (fitGLMFull Binomial Logit
                                              xLogi yLogi)))) ()
        , bench "Lasso_n10000_p50_lam0.1" $
            nf (\() -> LA.sumElements
                         (rfBeta (Reg.fitRegularized (L1 0.1) xLas yLas))) ()
        ]
    , bgroup "kernel"
        [ bench "KernelRidgeMV_n1000_p5_RBF" $
            nf (\() -> LA.sumElements
                         (Kn.krmvAlpha (Kn.kernelRidgeMV Kn.Gaussian 1.0 1e-3
                                          xKR yKRMat))) ()
        , bench "pairwiseSqDist_n2000_p5" $
            nf (LA.sumElements . KD.pairwiseSqDist) kdInput2k
        , bench "gramMatrixMV_n1000_p5_RBF" $
            nf (\() -> LA.sumElements (Kn.gramMatrixMV Kn.Gaussian 1.0 xKR)) ()
        ]
    , bgroup "cholesky"
        [ bench "cholSolve_n500" $
            nf (\() -> LA.sumElements
                         (fromMaybe (LA.scalar 0)
                                    (Chol.cholSolve spd500 spdRhs))) ()
        ]
    ]