create a "bundle" from all the C files (better inlining opportunities for the C compiler); also add a simple benchmark for the C FFT implementation

reference/src/FRI/Types.hs

@@ -268,6 +268,7 @@ instance Binary FriProof where
 
 --------------------------------------------------------------------------------
 
+{-
 estimateFriProofSize :: FriConfig -> Int
 estimateFriProofSize friConfig@(MkFriConfig{..}) = total where
 
@@ -308,3 +309,5 @@ data FriProof = MkFriProof
   , proofQueryRounds     :: [FriQueryRound]    -- ^ query rounds
   , proofPowWitness      :: F                  -- ^ witness showing that the prover did PoW
   }
+
+-}

reference/src/NTT/FFT.hs

@@ -1,14 +1,14 @@
 
 {-# LANGUAGE CPP #-}
 
-#ifdef USE_NAIVE_HASKELL
+-- #ifdef USE_NAIVE_HASKELL
 
 module NTT.FFT ( module NTT.FFT.Slow ) where
 import NTT.FFT.Slow
 
-#else
-
-module NTT.FFT ( module NTT.FFT.Fast ) where
-import NTT.FFT.Fast
-
-#endif
+-- #else
+-- 
+-- module NTT.FFT ( module NTT.FFT.Fast ) where
+-- import NTT.FFT.Fast
+-- 
+-- #endif

reference/src/NTT/FFT/Fast.hs

@@ -16,7 +16,7 @@ import System.IO.Unsafe
 
 import Data.Flat
 
-import NTT.Poly
+import NTT.Poly.Flat
 import NTT.Subgroup
 import Field.Goldilocks
 import Misc

reference/src/NTT/Poly.hs

@@ -1,13 +1,13 @@
 {-# LANGUAGE CPP #-}
 
-#ifdef USE_NAIVE_HASKELL
+-- #ifdef USE_NAIVE_HASKELL
 
 module NTT.Poly ( module NTT.Poly.Naive ) where
 import NTT.Poly.Naive
 
-#else
-
-module NTT.Poly ( module NTT.Poly.Flat ) where
-import NTT.Poly.Flat
-
-#endif
+-- #else
+-- 
+-- module NTT.Poly ( module NTT.Poly.Flat ) where
+-- import NTT.Poly.Flat
+-- 
+-- #endif

reference/src/NTT/Poly/Flat.hs

@@ -26,7 +26,13 @@ import Data.Flat as L
 
 --------------------------------------------------------------------------------
 
-newtype Poly a = MkPoly (L.FlatArray a)
+newtype Poly a 
+  = MkPoly (L.FlatArray a) 
+  deriving Show
+
+-- TEMPORARY HACK
+instance (Eq a, Flat a) => Eq (Poly a) where
+  p == q = (coeffs p == coeffs q)
 
 pattern XPoly n arr = MkPoly (L.MkFlatArray n arr)
 

reference/src/cbits/.gitignore

@@ -0,0 +1 @@
+a.out

reference/src/cbits/bench_fft.c

@@ -0,0 +1,59 @@
+
+#include <stdint.h>
+#include <stdlib.h>
+#include <stdio.h>
+
+#include <time.h>
+
+//#include "goldilocks.h"
+//#include "ntt.h"
+#include "bundle.h"
+
+//------------------------------------------------------------------------------
+
+const int LOGN    = 20;
+const int N       = (1<<LOGN);
+const int NREPEAT = 20;
+
+const uint64_t generator_size_32 = 0x185629dcda58878cULL;
+
+//------------------------------------------------------------------------------
+
+void print_time(const char *prefix, clock_t start, clock_t end) {
+  double cpu_time_used = ((double) (end - start)) / CLOCKS_PER_SEC;
+  double msec_per_megabyte = 1000 * cpu_time_used * 1024 * 1024 / (8 * N * NREPEAT );
+  double megabyte_per_sec  = 1000 / msec_per_megabyte;
+  printf("repeating %d %s of size 2^%d took %0.4f seconds (%0.0f MB/sec)\n", NREPEAT, prefix, LOGN, cpu_time_used, megabyte_per_sec);
+}
+
+int main() {
+
+  clock_t start, end;
+ 
+  uint64_t *src = (uint64_t*) malloc( 8 * N );     
+  uint64_t *tgt = (uint64_t*) malloc( 8 * N );  
+
+  for(int i=0; i<N; i++) { 
+    src[i] = 101 + 7*i;
+  }   
+
+  uint64_t gen = goldilocks_pow( generator_size_32 , 1<<(32-LOGN) );
+
+  start = clock();
+  for(int k=0; k<(NREPEAT/2); k++) { 
+    goldilocks_ntt_forward( LOGN, gen, src, tgt );
+    goldilocks_ntt_forward( LOGN, gen, tgt, src );
+  }
+  end = clock();
+  print_time("forward NTT", start, end);
+
+  start = clock();
+  for(int k=0; k<(NREPEAT/2); k++) { 
+    goldilocks_ntt_inverse( LOGN, gen, src, tgt );
+    goldilocks_ntt_inverse( LOGN, gen, tgt, src );
+  }
+  end = clock();
+  print_time("inverse NTT", start, end);
+
+  return 0;
+}

reference/src/cbits/bundle.c

@@ -0,0 +1,9 @@
+
+// gcc can do better inlineing if everything is in the same file
+// this can win like 33% speedup for the naive FFT algorithm...
+
+#include "goldilocks.c"
+#include "goldilocks_ext.c"
+#include "monolith.c"
+#include "ntt.c"
+#include "short_dft.c"

reference/src/cbits/bundle.h

@@ -0,0 +1,6 @@
+
+#include "goldilocks.h"
+#include "goldilocks_ext.h"
+#include "monolith.h"
+#include "ntt.h"
+#include "short_dft.h"

reference/src/cbits/compile.sh

@@ -5,4 +5,6 @@ gcc -c -O2 goldilocks_ext.c
 gcc -c -O2 monolith.c
 gcc -c -O2 ntt.c
 gcc -c -O2 short_dft.c
+gcc -c -O2 bundle.c
 
+gcc -O2 bench_fft.c bundle.c

reference/src/cbits/goldilocks.h

@@ -1,6 +1,8 @@
 
 // the "Goldilocks" prime field of size `p = 2^64 - 2^32 + 1`
 
+#ifndef _GOLDILOCKS_H_INCLUDED_
+
 #include <stdint.h>
 
 //------------------------------------------------------------------------------
@@ -44,3 +46,5 @@ void goldilocks_convert_31_bytes_to_4_field_elements (      const uint8_t *ptr,
 void goldilocks_convert_bytes_to_field_elements ( int rate, const uint8_t *ptr, uint64_t *felts );
 
 //------------------------------------------------------------------------------
+
+#endif // _GOLDILOCKS_H_INCLUDED_

reference/src/cbits/goldilocks_ext.h

@@ -1,6 +1,8 @@
 
 // quadratic field extension F[x] = F(x) / (x^2 - 7) over the Goldilocks field
 
+#ifndef _GOLDILOCKS_EXT_H_INCLUDED_
+
 #include <stdint.h>
 #include "goldilocks.h"
 
@@ -21,3 +23,4 @@ void goldilocks_ext_pow(const uint64_t *b , int e             , uint64_t *out);
 
 //------------------------------------------------------------------------------
 
+#endif // _GOLDILOCKS_EXT_H_INCLUDED_

reference/src/cbits/monolith.c

@@ -1,4 +1,6 @@
 
+// Monolith permutation and hash function
+
 #include <assert.h>
 #include <string.h>
 
@@ -242,3 +244,4 @@ void goldilocks_monolith_bytes_digest(int rate, int N, const uint8_t *input, uin
 }
 
 //------------------------------------------------------------------------------
+

reference/src/cbits/monolith.h

@@ -1,4 +1,7 @@
 
+
+#ifndef _MONOLITH_H_INCLUDED_
+
 #include <stdint.h>
 
 //------------------------------------------------------------------------------
@@ -11,3 +14,5 @@ void goldilocks_monolith_bytes_digest    (int rate, int N, const uint8_t  *input
 void goldilocks_monolith_felts_digest    (int rate, int N, const uint64_t *input, uint64_t *hash);
 
 //------------------------------------------------------------------------------
+
+#endif // _MONOLITH_H_INCLUDED_

reference/src/cbits/ntt.c

@@ -1,4 +1,5 @@
 
+
 #include <stdint.h>
 #include <stdlib.h>
 #include <string.h>

reference/src/cbits/ntt.h

@@ -1,4 +1,6 @@
 
+#ifndef _NTT_H_INCLUDED_
+
 #include <stdint.h>
 
 //------------------------------------------------------------------------------
@@ -12,3 +14,5 @@ void goldilocks_ntt_inverse           (              int m, uint64_t gen, const
 void goldilocks_ntt_inverse_shifted   (uint64_t eta, int m, uint64_t gen, const uint64_t *src, uint64_t *tgt);
 
 //------------------------------------------------------------------------------
+
+#endif // _NTT_H_INCLUDED_

reference/src/cbits/short_dft.h

@@ -1,4 +1,6 @@
 
+#ifndef _SHORT_DFT_H_INCLUDED_
+
 #include <stdint.h>
 
 //------------------------------------------------------------------------------
@@ -22,3 +24,5 @@ void short_inv_DFT_size_16_rescaled( int src_stride, int tgt_stride, uint64_t *s
 // void short_inv_DFT_size_4_ext_rescaled( int src_stride, int tgt_stride, uint64_t *src, uint64_t *tgt );
 
 //------------------------------------------------------------------------------
+
+#endif // _SHORT_DFT_H_INCLUDED_

reference/src/runi.sh

@@ -1,3 +1,3 @@
 #!/bin/bash
 
-ghci testMain.hs cbits/goldilocks.o cbits/goldilocks_ext.o cbits/monolith.o cbits/ntt.o cbits/short_dft.o
+ghci testMain.hs cbits/bundle.o