Pasta / Halo2 MSM bench

.github/workflows/ci.yml

@@ -160,6 +160,10 @@ jobs:
         run: |
           sudo dpkg --add-architecture i386
           sudo apt-fast update -qq
+
+          # Try to fix "E: Unable to correct problems, you have held broken packages."
+          sudo apt-fast clean
+
           sudo DEBIAN_FRONTEND='noninteractive' apt-fast install \
             --no-install-recommends -yq \
             gcc-multilib g++-multilib \
@@ -216,11 +220,12 @@ jobs:
           nimble refresh --verbose -y
           nimble install --verbose -y gmp jsony asynctools
 
-      - name: Print Nim version
-        if: runner.os != 'Windows'
+      - name: Print Nim & compiler versions
         shell: bash
+        # gcc is an alias to Clang on MacOS
         run: |
           nim -v
+          gcc -v 
 
       - name: Run Constantine tests (UNIX with Assembly)
         if: runner.os != 'Windows' && matrix.target.BACKEND == 'ASM'
@@ -235,9 +240,9 @@ jobs:
         shell: bash
         run: |
           cd constantine
-          nimble bindings_no_asm      --verbose
-          nimble test_bindings        --verbose
-          nimble test_parallel_no_asm --verbose
+          CTT_ASM=0 nimble bindings      --verbose
+          nimble test_bindings           --verbose
+          CTT_ASM=0 nimble test_parallel --verbose
       - name: Run Constantine tests (Windows with Assembly)
         # So "test_bindings" uses C and can find GMP
         # but nim-gmp cannot find GMP on Windows CI
@@ -255,6 +260,6 @@ jobs:
         shell: msys2 {0}
         run: |
           cd constantine
-          nimble bindings_no_asm             --verbose
-          nimble test_bindings               --verbose
-          nimble test_parallel_no_gmp_no_asm --verbose
+          CTT_ASM=0 nimble bindings             --verbose
+          nimble test_bindings                  --verbose
+          CTT_ASM=0 nimble test_parallel_no_gmp --verbose

benchmarks/bench_blueprint.nim

@@ -60,7 +60,7 @@ echo "  release: ", defined(release)
 echo "  danger: ", defined(danger)
 echo "  inline assembly: ", UseASM_X86_64
 
-when (sizeof(int) == 4) or defined(Ctt32):
+when (sizeof(int) == 4) or defined(CTT_32):
   echo "⚠️ Warning: using Constantine with 32-bit limbs"
 else:
   echo "Using Constantine with 64-bit limbs"

benchmarks/bench_ec_g1_msm_pasta.nim

@@ -0,0 +1,50 @@
+# Constantine
+# Copyright (c) 2018-2019    Status Research & Development GmbH
+# Copyright (c) 2020-Present Mamy André-Ratsimbazafy
+# Licensed and distributed under either of
+#   * MIT license (license terms in the root directory or at http://opensource.org/licenses/MIT).
+#   * Apache v2 license (license terms in the root directory or at http://www.apache.org/licenses/LICENSE-2.0).
+# at your option. This file may not be copied, modified, or distributed except according to those terms.
+
+import
+  # Internals
+  ../constantine/math/config/curves,
+  ../constantine/math/arithmetic,
+  ../constantine/math/elliptic/[
+    ec_shortweierstrass_projective,
+    ec_shortweierstrass_jacobian],
+  # Helpers
+  ../helpers/prng_unsafe,
+  ./bench_elliptic_parallel_template
+
+# ############################################################
+#
+#               Benchmark of the G1 group of
+#            Short Weierstrass elliptic curves
+#          in (homogeneous) projective coordinates
+#
+# ############################################################
+
+
+const Iters = 10_000
+const AvailableCurves = [
+  Pallas, Vesta
+]
+
+# const testNumPoints = [10, 100, 1000, 10000, 100000]
+# const testNumPoints = [64, 128, 256, 512, 1024, 2048, 4096, 8192, 16384, 32768, 65536, 131072]
+const testNumPoints = [1 shl 8, 1 shl 9, 1 shl 10, 1 shl 11, 1 shl 12, 1 shl 13, 1 shl 14, 1 shl 15, 1 shl 16, 1 shl 17, 1 shl 22]
+
+proc main() =
+  separator()
+  staticFor i, 0, AvailableCurves.len:
+    const curve = AvailableCurves[i]
+    separator()
+    for numPoints in testNumPoints:
+      let batchIters = max(1, Iters div numPoints)
+      msmParallelBench(ECP_ShortW_Jac[Fp[curve], G1], numPoints, batchIters)
+      separator()
+    separator()
+
+main()
+notes()

benchmarks/bench_ec_g1_msm_pasta.nim.cfg

@@ -0,0 +1 @@
+--threads:on

benchmarks/bench_elliptic_parallel_template.nim

@@ -21,7 +21,7 @@ import
     ec_multi_scalar_mul_parallel],
   ../constantine/math/constants/zoo_subgroups,
   # Threadpool
-  ../constantine/threadpool/threadpool,
+  ../constantine/threadpool/[threadpool, partitioners],
   # Helpers
   ../helpers/prng_unsafe,
   ./bench_elliptic_template,
@@ -55,11 +55,32 @@ proc msmParallelBench*(EC: typedesc, numPoints: int, iters: int) =
   var points = newSeq[ECP_ShortW_Aff[EC.F, EC.G]](numPoints)
   var scalars = newSeq[BigInt[bits]](numPoints)
 
-  for i in 0 ..< numPoints:
-    var tmp = rng.random_unsafe(EC)
-    tmp.clearCofactor()
-    points[i].affine(tmp)
-    scalars[i] = rng.random_unsafe(BigInt[bits])
+  # Creating millions of points and clearing their cofactor takes a long long time
+  var tp = Threadpool.new()
+
+  proc genCoefPointPairs(rngSeed: uint64, start, len: int, points: ptr ECP_ShortW_Aff[EC.F, EC.G], scalars: ptr BigInt[bits]) {.nimcall.} =
+    let points = cast[ptr UncheckedArray[ECP_ShortW_Aff[EC.F, EC.G]]](points) # TODO use views to reduce verbosity
+    let scalars = cast[ptr UncheckedArray[BigInt[bits]]](scalars)
+
+    # RNGs are not threadsafe, create a threadlocal one seeded from the global RNG
+    var threadRng: RngState
+    threadRng.seed(rngSeed)
+
+    for i in start ..< start + len:
+      var tmp = threadRng.random_unsafe(EC)
+      tmp.clearCofactor()
+      points[i].affine(tmp)
+      scalars[i] = rng.random_unsafe(BigInt[bits])
+
+  let chunks = balancedChunksPrioNumber(0, numPoints, tp.numThreads)
+
+  syncScope:
+    for (id, start, size) in items(chunks):
+      tp.spawn genCoefPointPairs(rng.next(), start, size, points[0].addr, scalars[0].addr)
+
+  # Even if child threads are sleeping, it seems like perf is lower when there are threads around
+  # maybe because the kernel has more overhead or time quantum to keep track off so shut them down.
+  tp.shutdown()
 
   var r{.noInit.}: EC
   var startNaive, stopNaive, startMSMbaseline, stopMSMbaseline, startMSMopt, stopMSMopt, startMSMpara, stopMSMpara: MonoTime
@@ -88,7 +109,7 @@ proc msmParallelBench*(EC: typedesc, numPoints: int, iters: int) =
     stopMSMopt = getMonotime()
 
   block:
-    var tp = Threadpool.new()
+    tp = Threadpool.new()
 
     startMSMpara = getMonotime()
     bench("EC multi-scalar-mul" & align($tp.numThreads & " threads", 11) & align($numPoints, 10) & " (" & $bits & "-bit coefs, points)", EC, iters):

benchmarks/bench_elliptic_template.nim

@@ -45,7 +45,7 @@ macro fixEllipticDisplay(EC: typedesc): untyped =
   var name = $instantiated[1][0] # EllipticEquationFormCoordinates
   let fieldName = $instantiated[1][1][0]
   let curveName = $Curve(instantiated[1][1][1].intVal)
-  name.add "[" & fieldName & "[" & curveName & ']'
+  name.add "[" & fieldName & "[" & curveName & "]]"
   result = newLit name
 
 proc report(op, elliptic: string, start, stop: MonoTime, startClk, stopClk: int64, iters: int) =

benchmarks/bench_fp_double_precision.nim

@@ -61,7 +61,7 @@ echo "  release: ", defined(release)
 echo "  danger: ", defined(danger)
 echo "  inline assembly: ", UseASM_X86_64
 
-when (sizeof(int) == 4) or defined(Ctt32):
+when (sizeof(int) == 4) or defined(CTT_32):
   echo "⚠️ Warning: using Constantine with 32-bit limbs"
 else:
   echo "Using Constantine with 64-bit limbs"

bindings_generators/constantine_bls12_381.nim

@@ -12,10 +12,10 @@ type
   bls12381_fr = Fr[BLS12_381]
   bls12381_fp = Fp[BLS12_381]
   bls12381_fp2 = Fp2[BLS12_381]
-  bls12381_ec_g1_aff = ECP_ShortW_Aff[Fp[BLS12_381], G1] 
+  bls12381_ec_g1_aff = ECP_ShortW_Aff[Fp[BLS12_381], G1]
   bls12381_ec_g1_jac = ECP_ShortW_Jac[Fp[BLS12_381], G1]
   bls12381_ec_g1_prj = ECP_ShortW_Prj[Fp[BLS12_381], G1]
-  bls12381_ec_g2_aff = ECP_ShortW_Aff[Fp2[BLS12_381], G2] 
+  bls12381_ec_g2_aff = ECP_ShortW_Aff[Fp2[BLS12_381], G2]
   bls12381_ec_g2_jac = ECP_ShortW_Jac[Fp2[BLS12_381], G2]
   bls12381_ec_g2_prj = ECP_ShortW_Prj[Fp2[BLS12_381], G2]
 
@@ -33,9 +33,9 @@ collectBindings(cBindings):
   genBindings_EC_ShortW_NonAffine(bls12381_ec_g2_prj, bls12381_ec_g2_aff, bls12381_fp2)
 
 # Write header
-when isMainModule and defined(CttGenerateHeaders):
+when isMainModule and defined(CTT_GENERATE_HEADERS):
   import std/[os, strformat]
-  
+
   proc main() =
     # echo "Running bindings generation for " & getAppFilename().extractFilename()
 

bindings_generators/constantine_pasta.nim

@@ -13,10 +13,10 @@ type
   pallas_fp = Fp[Pallas]
   vesta_fr = Fr[Vesta]
   vesta_fp = Fp[Vesta]
-  pallas_ec_aff = ECP_ShortW_Aff[Fp[Pallas], G1] 
+  pallas_ec_aff = ECP_ShortW_Aff[Fp[Pallas], G1]
   pallas_ec_jac = ECP_ShortW_Jac[Fp[Pallas], G1]
   pallas_ec_prj = ECP_ShortW_Prj[Fp[Pallas], G1]
-  vesta_ec_aff = ECP_ShortW_Aff[Fp[Vesta], G1] 
+  vesta_ec_aff = ECP_ShortW_Aff[Fp[Vesta], G1]
   vesta_ec_jac = ECP_ShortW_Jac[Fp[Vesta], G1]
   vesta_ec_prj = ECP_ShortW_Prj[Fp[Vesta], G1]
 
@@ -35,9 +35,9 @@ collectBindings(cBindings):
   genBindings_EC_ShortW_NonAffine(vesta_ec_prj, vesta_ec_aff, vesta_fp)
 
 # Write header
-when isMainModule and defined(CttGenerateHeaders):
+when isMainModule and defined(CTT_GENERATE_HEADERS):
   import std/[os, strformat]
-  
+
   proc main() =
     # echo "Running bindings generation for " & getAppFilename().extractFilename()
 

bindings_generators/gen_header.nim

@@ -203,7 +203,7 @@ macro collectBindings*(cBindingsStr: untyped, body: typed): untyped =
 
       cBindings &= ");"
 
-  if defined(CttGenerateHeaders):
+  if defined(CTT_GENERATE_HEADERS):
     result = newConstStmt(cBindingsStr, newLit cBindings)
   else:
     result = body