|
| 1 | +import PartIiProject.Bench.Common |
| 2 | +import PartIiProject.Bench.Table |
| 3 | +import PartIiProject.Optimisations |
| 4 | +import PartIiProject.SyntaxSDQLProg |
| 5 | +import Lean |
| 6 | +import Std |
| 7 | + |
| 8 | +open PartIiProject |
| 9 | +open PartIiProject.Optimisations |
| 10 | +open System |
| 11 | + |
| 12 | +namespace OptimisationPerformanceComparison |
| 13 | + |
| 14 | +def outDir : FilePath := FilePath.mk ".sdql-opt-perf-out" |
| 15 | +def runtimeSrc : FilePath := FilePath.mk "sdql_runtime.rs" |
| 16 | + |
| 17 | +def iters : Nat := 3 |
| 18 | +def dictN : Int := 200000 |
| 19 | +def memoN : Int := 100000 |
| 20 | +def memoM : Int := 1000 |
| 21 | + |
| 22 | +unsafe structure BenchCase where |
| 23 | + name : String |
| 24 | + prog : SProg2 |
| 25 | + opts : List Optimisation |
| 26 | + env : List (String × String) := [] |
| 27 | + |
| 28 | +structure Reading where |
| 29 | + name : String |
| 30 | + unoptMs : Nat |
| 31 | + optMs : Nat |
| 32 | + |
| 33 | +def meanNat (xs : List Nat) : Nat := |
| 34 | + match xs with |
| 35 | + | [] => 0 |
| 36 | + | _ => |
| 37 | + let total := xs.foldl (fun s x => s + x) 0 |
| 38 | + total / xs.length |
| 39 | + |
| 40 | +def timeBinaryAvgMs (binPath : FilePath) (iters : Nat) (env : List (String × String)) : |
| 41 | + IO (Except String Nat) := do |
| 42 | + if iters == 0 then |
| 43 | + return .error "iters must be > 0" |
| 44 | + let mut acc : List Nat := [] |
| 45 | + for _ in [0:iters] do |
| 46 | + match ← PartIiProject.Bench.timeBinaryMs binPath env with |
| 47 | + | .ok ms => acc := ms :: acc |
| 48 | + | .error e => return .error e |
| 49 | + return .ok (meanNat acc) |
| 50 | + |
| 51 | +unsafe def runCase (b : BenchCase) : IO (Except String Reading) := do |
| 52 | + let baseCp := ToCore2.trProg2 b.prog |
| 53 | + let optTerm := applyOptimisationsLoc b.opts baseCp.term |
| 54 | + let optCp : Prog2 := { baseCp with term := optTerm } |
| 55 | + |
| 56 | + let unoptBin ← |
| 57 | + match ← PartIiProject.Bench.compileProg2ToBin outDir runtimeSrc s!"{b.name}_unopt" baseCp with |
| 58 | + | .ok p => pure p |
| 59 | + | .error e => return .error s!"Unoptimised compile failed:\n{e}" |
| 60 | + |
| 61 | + let optBin ← |
| 62 | + match ← PartIiProject.Bench.compileProg2ToBin outDir runtimeSrc s!"{b.name}_opt" optCp with |
| 63 | + | .ok p => pure p |
| 64 | + | .error e => return .error s!"Optimised compile failed:\n{e}" |
| 65 | + |
| 66 | + let unoptMs ← |
| 67 | + match ← timeBinaryAvgMs unoptBin iters b.env with |
| 68 | + | .ok ms => pure ms |
| 69 | + | .error e => return .error s!"Unoptimised run failed:\n{e}" |
| 70 | + |
| 71 | + let optMs ← |
| 72 | + match ← timeBinaryAvgMs optBin iters b.env with |
| 73 | + | .ok ms => pure ms |
| 74 | + | .error e => return .error s!"Optimised run failed:\n{e}" |
| 75 | + |
| 76 | + return .ok { name := b.name, unoptMs := unoptMs, optMs := optMs } |
| 77 | + |
| 78 | +unsafe def p_vertical_loop_fusion_key_map : SProg2 := |
| 79 | + [SDQLProg2 { { int -> int } }| |
| 80 | + let dictN = 200000 in |
| 81 | + let d = sum( <i, _> <- range(dictN) ) { i -> i } in |
| 82 | + let y = sum( <x, x_v> <- d ) { x + 1 -> x_v } in |
| 83 | + sum( <x, x_v> <- y ) { x + 2 -> x_v } |
| 84 | + ] |
| 85 | + |
| 86 | +unsafe def p_vertical_loop_fusion_value_map : SProg2 := |
| 87 | + [SDQLProg2 { { int -> int } }| |
| 88 | + let dictN = 200000 in |
| 89 | + let d = sum( <i, _> <- range(dictN) ) { i -> i } in |
| 90 | + let y = sum( <x, x_v> <- d ) { x -> x_v + 1 } in |
| 91 | + sum( <x, x_v> <- y ) { x -> x_v + 2 } |
| 92 | + ] |
| 93 | + |
| 94 | +unsafe def p_horizontal_loop_fusion : SProg2 := |
| 95 | + [SDQLProg2 { int }| |
| 96 | + let dictN = 200000 in |
| 97 | + let d = sum( <i, _> <- range(dictN) ) { i -> i } in |
| 98 | + let y1 = sum( <_, v> <- d ) v in |
| 99 | + let y2 = sum( <_, v> <- d ) (v + 1) in |
| 100 | + y1 + y2 |
| 101 | + ] |
| 102 | + |
| 103 | +unsafe def p_loop_factorization_left : SProg2 := |
| 104 | + [SDQLProg2 { int }| |
| 105 | + let dictN = 200000 in |
| 106 | + let d = sum( <i, _> <- range(dictN) ) { i -> i } in |
| 107 | + sum( <_, v> <- d ) (2 * v) |
| 108 | + ] |
| 109 | + |
| 110 | +unsafe def p_loop_factorization_right : SProg2 := |
| 111 | + [SDQLProg2 { int }| |
| 112 | + let dictN = 200000 in |
| 113 | + let d = sum( <i, _> <- range(dictN) ) { i -> i } in |
| 114 | + sum( <_, v> <- d ) (v * 2) |
| 115 | + ] |
| 116 | + |
| 117 | +unsafe def p_loop_invariant_code_motion : SProg2 := |
| 118 | + [SDQLProg2 { int }| |
| 119 | + let dictN = 200000 in |
| 120 | + let d = sum( <i, _> <- range(dictN) ) { i -> i } in |
| 121 | + sum( <_, v> <- d ) (let y = 5 in v + y) |
| 122 | + ] |
| 123 | + |
| 124 | +-- Memoization lookup + hoisting: without hoisting this is not a win; include code motion. |
| 125 | +unsafe def p_loop_memoization_lookup : SProg2 := |
| 126 | + [SDQLProg2 { int }| |
| 127 | + let memoN = 100000 in |
| 128 | + let memoM = 1000 in |
| 129 | + let d = sum( <i, _> <- range(memoN) ) { i -> i } in |
| 130 | + sum( <i, _> <- range(memoM) ) (sum( <k, v> <- d ) (if k == i then v)) |
| 131 | + ] |
| 132 | + |
| 133 | +-- Partition + hoisting: then-branch depends on `i`, so lookup-memoization cannot hoist `f`. |
| 134 | +unsafe def p_loop_memoization_partition : SProg2 := |
| 135 | + [SDQLProg2 { int }| |
| 136 | + let memoN = 100000 in |
| 137 | + let memoM = 1000 in |
| 138 | + let d = sum( <i, _> <- range(memoN) ) { i -> i } in |
| 139 | + sum( <i, _> <- range(memoM) ) (sum( <k, v> <- d ) (if k == i then v + i)) |
| 140 | + ] |
| 141 | + |
| 142 | +unsafe def mkCases : List BenchCase := |
| 143 | + [ { name := "vertical_loop_fusion_key_map" |
| 144 | + prog := p_vertical_loop_fusion_key_map |
| 145 | + opts := [verticalLoopFusionKeyMap2, verticalLoopFusionValueMap2] |
| 146 | + } |
| 147 | + , { name := "vertical_loop_fusion_value_map" |
| 148 | + prog := p_vertical_loop_fusion_value_map |
| 149 | + opts := [verticalLoopFusionKeyMap2, verticalLoopFusionValueMap2] |
| 150 | + } |
| 151 | + , { name := "horizontal_loop_fusion" |
| 152 | + prog := p_horizontal_loop_fusion |
| 153 | + opts := [horizontalLoopFusion2] |
| 154 | + } |
| 155 | + , { name := "loop_factorization_left" |
| 156 | + prog := p_loop_factorization_left |
| 157 | + opts := [loopFactorizationLeft2] |
| 158 | + } |
| 159 | + , { name := "loop_factorization_right" |
| 160 | + prog := p_loop_factorization_right |
| 161 | + opts := [loopFactorizationRight2] |
| 162 | + } |
| 163 | + , { name := "loop_invariant_code_motion" |
| 164 | + prog := p_loop_invariant_code_motion |
| 165 | + opts := [loopInvariantCodeMotion2] |
| 166 | + } |
| 167 | + , { name := "loop_memoization_lookup" |
| 168 | + prog := p_loop_memoization_lookup |
| 169 | + opts := [loopMemoizationLookup2, loopInvariantCodeMotion2] |
| 170 | + } |
| 171 | + , { name := "loop_memoization_partition" |
| 172 | + prog := p_loop_memoization_partition |
| 173 | + opts := [loopMemoizationPartition2, loopInvariantCodeMotion2] |
| 174 | + } |
| 175 | + ] |
| 176 | + |
| 177 | +unsafe def main (_args : List String) : IO UInt32 := do |
| 178 | + if (← outDir.pathExists) then |
| 179 | + IO.FS.removeDirAll outDir |
| 180 | + IO.FS.createDirAll outDir |
| 181 | + |
| 182 | + let mut readings : List Reading := [] |
| 183 | + let mut failures : List (String × String) := [] |
| 184 | + |
| 185 | + for b in mkCases do |
| 186 | + match ← runCase b with |
| 187 | + | .ok r => readings := readings.concat r |
| 188 | + | .error e => failures := failures.concat (b.name, e) |
| 189 | + |
| 190 | + PartIiProject.Bench.printMsComparisonTable |
| 191 | + s!"SDQL optimisation performance comparison (mean of {iters} run(s); wall-clock ms)" |
| 192 | + "unopt" "opt" "opt/unopt" |
| 193 | + (readings.map fun r => |
| 194 | + { name := r.name, leftMs := r.unoptMs, rightMs := r.optMs }) |
| 195 | + (preamble := [s!"Params: dictN={dictN}, memoN={memoN}, memoM={memoM}"]) |
| 196 | + |
| 197 | + if !failures.isEmpty then |
| 198 | + IO.eprintln "" |
| 199 | + IO.eprintln "Failures:" |
| 200 | + for (nm, err) in failures do |
| 201 | + IO.eprintln s!"- {nm}: {err}" |
| 202 | + return 1 |
| 203 | + return 0 |
| 204 | + |
| 205 | +end OptimisationPerformanceComparison |
| 206 | + |
| 207 | +unsafe def main (args : List String) : IO UInt32 := |
| 208 | + OptimisationPerformanceComparison.main args |
0 commit comments