diff --git a/benchmarks/CMakeLists.txt b/benchmarks/CMakeLists.txt
index debfc0ca..f56a5f5b 100644
--- a/benchmarks/CMakeLists.txt
+++ b/benchmarks/CMakeLists.txt
@@ -1,4 +1,4 @@
-add_executable(benchmarks benchmarks.cpp shufti.cpp truffle.cpp noodle.cpp)
-set_source_files_properties(shufti.cpp PROPERTIES COMPILE_FLAGS
+add_executable(benchmarks benchmarks.cpp noodle.cpp)
+set_source_files_properties(benchmarks.cpp PROPERTIES COMPILE_FLAGS
     "-Wall -Wno-unused-variable")
 target_link_libraries(benchmarks hs)
\ No newline at end of file
diff --git a/benchmarks/benchmarks.cpp b/benchmarks/benchmarks.cpp
index be224ba9..ce680334 100644
--- a/benchmarks/benchmarks.cpp
+++ b/benchmarks/benchmarks.cpp
@@ -1,25 +1,161 @@
-#include "benchmarks.hpp"
 #include <iostream>
-#include <string>
-#include <string.h>
-#include <time.h>
-#include <functional>
-#include <vector>
+#include <chrono>
+#include <cstring>
+#include <ctime>
+#include <cstdlib>
 #include <memory>
 
+#include "nfa/shufti.h"
+#include "nfa/shufticompile.h"
+#include "nfa/truffle.h"
+#include "nfa/trufflecompile.h"
+#include "benchmarks.hpp"
+
 #define MAX_LOOPS    500000000
 #define MAX_MATCHES  10
 
+/*
+void shuffle_init(){
+    m128 lo, hi;
+    ue2::CharReach chars;
+    chars.set('a');
+    shuftiBuildMasks(chars, (u8 *)&lo, (u8 *)&hi);
+    std::unique_ptr<u8 []> kt1 ( new u8[size] );
+    memset(kt1.get(),'b',size);  
+}
+*/
+
+/*
+void truffle_init(){
+     m128 lo, hi;
+    ue2::CharReach chars;
+    chars.set('a');
+    truffleBuildMasks(chars, (u8 *)&lo, (u8 *)&hi);
+    std::unique_ptr<u8 []> kt1 ( new u8[size] );
+    memset(kt1.get(),'b',size); 
+}
+*/
+
+/*
+struct hlmMatchEntry {
+    size_t to;
+    u32 id;
+    hlmMatchEntry(size_t end, u32 identifier) :
+            to(end), id(identifier) {}
+};
+
+std::vector<hlmMatchEntry> ctxt;
+
+static
+hwlmcb_rv_t hlmSimpleCallback(size_t to, u32 id,
+                              UNUSED struct hs_scratch *scratch) {
+    DEBUG_PRINTF("match @%zu = %u\n", to, id);
+
+    ctxt.push_back(hlmMatchEntry(to, id));
+
+    return HWLM_CONTINUE_MATCHING;
+}
+
+void noodle_init(){
+    ctxt.clear();
+    std::unique_ptr<u8 []> data ( new u8[size] );
+    memset(data.get(), 'a', size);
+    double total_sec = 0.0;
+    u64a transferred_size = 0;
+    double avg_time = 0.0;
+    double max_bw = 0.0;
+    double bandwitdh = 0.0;
+    u32 id = 1000;
+    ue2::hwlmLiteral lit(std::string(lit_str, lit_len), nocase, id);
+    auto n = ue2::noodBuildTable(lit);
+    assert(n != nullptr);
+    struct hs_scratch scratch;
+}
+*/
+
+void run_benchmarks(int size, int loops, int M, bool has_match, std::function <const u8 *(m128, m128, const u8 *, const u8 *)> function) {
+    m128 lo, hi;
+    ue2::CharReach chars;
+    chars.set('a');
+    shuftiBuildMasks(chars, (u8 *)&lo, (u8 *)&hi);
+    std::unique_ptr<u8 []> kt1 ( new u8[size] );
+    memset(kt1.get(),'b',size);
+    double total_sec = 0.0;            
+    u64a transferred_size = 0;
+    double bandwidth = 0.0;
+    double max_bw = 0.0;
+    double avg_time = 0.0;
+    if (has_match) {
+        int pos = 0;
+        for(int j = 0; j < M; j++) {
+            kt1[pos] = 'b';
+            pos = (j*size) / M ;
+            kt1[pos] = 'a';
+            unsigned long act_size = 0;
+            auto start = std::chrono::steady_clock::now();
+            for(int i = 0; i < loops; i++) { 
+                const u8 *res = function(lo, hi, kt1.get(), kt1.get() + size);
+                act_size += res - kt1.get();
+            }
+            auto end = std::chrono::steady_clock::now();
+            double dt = std::chrono::duration_cast<std::chrono::microseconds>(end - start).count();
+            total_sec += dt;
+            /*convert microseconds to seconds*/
+            total_sec /= 1000000.0;
+            /*calculate bandwidth*/
+            bandwidth  += (act_size / dt) * 1000000.0;
+            /*convert to MB/s*/
+            bandwidth  = bandwidth  / 1048576.0;
+            max_bw = std::max(bandwidth ,max_bw);
+            /*calculate average time*/
+            avg_time += total_sec / loops;
+        }
+        avg_time /= M;
+        bandwidth /= M;
+        /*convert average time to us*/
+        avg_time *= 1000000.0;
+        printf(KMAG "case with %u matches, %u * %u iterations," KBLU " total elapsed time =" RST " %.3f s, " 
+               KBLU "average time per call =" RST " %.3f μs," KBLU " max bandwidth = " RST " %.3f MB/s," KBLU " average bandwidth =" RST " %.3f MB/s \n",
+               M, size ,loops, total_sec, avg_time, max_bw, bandwidth);
+    } else {
+        auto start = std::chrono::steady_clock::now();
+        for (int i = 0; i < loops; i++) {
+            function(lo, hi, kt1.get(), kt1.get() + size);
+        }
+        auto end = std::chrono::steady_clock::now();
+        total_sec += std::chrono::duration_cast<std::chrono::microseconds>(end - start).count();
+        /*calculate transferred size*/
+        transferred_size = size * loops;
+        /*calculate average time*/
+        avg_time = total_sec / loops;
+        /*convert microseconds to seconds*/
+        total_sec /= 1000000.0;
+        /*calculate maximum bandwidth*/
+        max_bw = transferred_size / total_sec;
+        /*convert to MB/s*/
+        max_bw /= 1048576.0;
+        printf(KMAG "case without matches, %u * %u iterations," KBLU " total elapsed time =" RST " %.3f s, " 
+               KBLU "average time per call =" RST " %.3f μs ," KBLU " bandwidth = " RST " %.3f MB/s \n",
+               size ,loops, total_sec, avg_time, max_bw);
+    }
+}
+
+
 int main(){
-    std::function<void(int,int,int,bool)>  functions[] = { shufti_benchmarks, rshufti_benchmarks, truffle_benchmarks, rtruffle_benchmarks };
+    std::function <const u8 *(m128, m128, const u8 *, const u8 *)> functions[] = {shuftiExec, rshuftiExec, truffleExec, rtruffleExec};
     int sizes[] =  {  16000,  32000,  64000, 120000, 1600000, 2000000, 2500000, 3500000, 150000000, 250000000, 350000000, 500000000 };
+    std::string labels[] = {"\x1B[33m shuftiExec Benchmarks \x1B[0m\n", "\x1B[33m rshuftiExec Benchmarks \x1B[0m\n",
+                            "\x1B[33m triffleExec Benchmarks \x1B[0m\n", "\x1B[33m triffleExec Benchmarks \x1B[0m\n"};
     const char charset[] = "aAaAaAaAAAaaaaAAAAaaaaAAAAAAaaaAAaaa"; 
+    
     for (size_t i = 0; i < std::size(sizes); i++) {
         for(size_t j = 0; j < std::size(functions); j++) {
-            functions[j](sizes[i], MAX_LOOPS / sizes[i], MAX_MATCHES, false);
-            functions[j](sizes[i], MAX_LOOPS / sizes[i], MAX_MATCHES, true);  
+            std::cout << labels[j];
+            run_benchmarks(sizes[i], MAX_LOOPS / sizes[i], MAX_MATCHES, false, functions[j]);
+            run_benchmarks(sizes[i], MAX_LOOPS / sizes[i], MAX_MATCHES, true, functions[j]);
         } 
     }
+    
     for(size_t i=0; i < std::size(sizes); i++){
         //we imitate the noodle unit tests
         for (int char_len = 1; char_len < 9; char_len++) {
diff --git a/benchmarks/benchmarks.hpp b/benchmarks/benchmarks.hpp
index 894ed9ce..621dfb25 100644
--- a/benchmarks/benchmarks.hpp
+++ b/benchmarks/benchmarks.hpp
@@ -1,3 +1,5 @@
+#include <functional>
+
 /*define colour control characters*/
 #define RST  "\x1B[0m"
 #define KRED  "\x1B[31m"
@@ -8,8 +10,6 @@
 #define KCYN  "\x1B[36m"
 #define KWHT  "\x1B[37m"
 
-void shufti_benchmarks(int size, int loops, int M, bool has_match);
-void rshufti_benchmarks(int size, int loops, int M, bool has_match);
-void truffle_benchmarks(int size, int loops, int M, bool has_match);
-void rtruffle_benchmarks(int size, int loops, int M, bool has_match);
-void noodle_benchmarks(int size, int M, const char *lit_str, int lit_len, char nocase);
\ No newline at end of file
+
+void noodle_benchmarks(int size, int M, const char *lit_str, int lit_len, char nocase);
+void run_benchmarks(int size, int loops, int M, bool has_match, std::function <const u8 *(m128, m128, const u8 *, const u8 *)> function);
\ No newline at end of file
diff --git a/benchmarks/shufti.cpp b/benchmarks/shufti.cpp
deleted file mode 100644
index f6c2be5c..00000000
--- a/benchmarks/shufti.cpp
+++ /dev/null
@@ -1,149 +0,0 @@
-#include "nfa/shufti.h"
-#include "nfa/shufticompile.h"
-#include "benchmarks.hpp"
-#include <iostream>
-#include <chrono>
-#include <cstring>
-#include <ctime>
-#include <cstdlib>
-#include <memory>
-
-void shufti_benchmarks(int size, int loops, int M, bool has_match) { 
-    m128 lo, hi;
-    ue2::CharReach chars;
-    chars.set('a');
-    int ret = shuftiBuildMasks(chars, (u8 *)&lo, (u8 *)&hi);
-    std::unique_ptr<u8 []> kt1 ( new u8[size] );
-    memset(kt1.get(),'b',size);
-    double total_sec = 0.0;            
-    u64a transferred_size = 0;
-    double bandwitdh = 0.0;
-    double max_bw = 0.0;
-    double avg_time = 0.0;
-    if (has_match) {
-        int pos = 0;
-        for(int j = 0; j < M; j++) {
-            kt1[pos] = 'b';
-            srand (time(NULL));
-            pos = rand() % size + 0;
-            kt1[pos] = 'a';
-            unsigned long act_size = 0;
-            auto start = std::chrono::steady_clock::now();
-            for(int i = 0; i < loops; i++) { 
-                const u8 *res = shuftiExec(lo, hi, kt1.get(), kt1.get() + size);
-                act_size += res - kt1.get();
-            }
-            auto end = std::chrono::steady_clock::now();
-            double dt = std::chrono::duration_cast<std::chrono::microseconds>(end - start).count();
-            total_sec += dt;
-            /*convert microseconds to seconds*/
-            total_sec /= 1000000.0;
-            /*calculate bandwidth*/
-            bandwitdh  += act_size / total_sec;
-            /*convert to MB/s*/
-            bandwitdh  = bandwitdh  / 1048576.0;
-            max_bw = std::max(bandwitdh ,max_bw);
-            /*calculate average time*/
-            avg_time += total_sec / loops;
-        }
-        avg_time /= M;
-        bandwitdh /= M;
-        /*convert average time to us*/
-        avg_time *= 1000000.0;
-        printf(KMAG "ShuftiExec: case with %u matches, %u * %u iterations," KBLU " total elapsed time =" RST " %.3f s, " 
-               KBLU "average time per call =" RST " %.3f μs," KBLU " bandwidth = " RST " %.3f MB/s," KBLU " average bandwidth =" RST " %.3f MB/s \n",
-               M, size ,loops, total_sec, avg_time, max_bw, bandwitdh);
-    } else {
-        auto start = std::chrono::steady_clock::now();
-        for (int i = 0; i < loops; i++) {
-            shuftiExec(lo, hi, kt1.get(), kt1.get() + size);
-        }
-        auto end = std::chrono::steady_clock::now();
-        total_sec += std::chrono::duration_cast<std::chrono::microseconds>(end - start).count();
-        /*calculate transferred size*/
-        transferred_size = size * loops;
-        /*calculate average time*/
-        avg_time = total_sec / loops;
-        /*convert microseconds to seconds*/
-        total_sec /= 1000000.0;
-        /*calculate maximum bandwidth*/
-        max_bw = transferred_size / total_sec;
-        /*convert to MB/s*/
-        max_bw /= 1048576.0;
-        /*calculate average bandwidth*/
-        bandwitdh = max_bw / loops;
-        printf(KMAG "ShuftiExec: case without matches, %u * %u iterations," KBLU " total elapsed time =" RST " %.3f s, " 
-               KBLU "average time per call =" RST " %.3f μs ," KBLU " bandwidth = " RST " %.3f MB/s," KBLU " average bandwidth =" RST " %.3f MB/s \n",
-               size ,loops, total_sec, avg_time, max_bw, bandwitdh);
-    }
-}
-
-void rshufti_benchmarks(int size, int loops, int M, bool has_match) {
-    m128 lo, hi;
-    ue2::CharReach chars;
-    chars.set('a');
-    int ret = shuftiBuildMasks(chars, (u8 *)&lo, (u8 *)&hi);
-    std::unique_ptr<u8 []> kt1 ( new u8[size] );
-    memset(kt1.get(),'b',size);
-    double total_sec = 0.0;            
-    u64a transferred_size = 0;
-    double bandwitdh = 0.0;
-    double max_bw = 0.0;
-    double avg_time = 0.0;
-    if (has_match) {
-        int pos = 0;
-        for(int j = 0; j < M; j++) {
-            kt1[pos] = 'b';
-            srand (time(NULL));
-            pos = rand() % size + 0;
-            kt1[pos] = 'a';
-            unsigned long act_size = 0;
-            auto start = std::chrono::steady_clock::now();
-            for(int i = 0; i < loops; i++) {
-                const u8 *res = rshuftiExec(lo, hi, kt1.get(), kt1.get() + size);
-                act_size += res - kt1.get();
-            }
-            auto end = std::chrono::steady_clock::now();
-            double dt = std::chrono::duration_cast<std::chrono::microseconds>(end - start).count();
-            total_sec += dt;
-            /*convert microseconds to seconds*/
-            total_sec /= 1000000.0;
-            /*calculate bandwidth*/
-            bandwitdh  += act_size / total_sec;
-            /*convert to MB/s*/
-            bandwitdh  = bandwitdh  / 1048576.0;
-            max_bw = std::max(bandwitdh ,max_bw);
-            /*calculate average time*/
-            avg_time += total_sec / loops;
-        }
-        avg_time /= M;
-        bandwitdh /= M;
-        /*convert average time to μs*/
-        avg_time *= 1000000.0;
-        printf(KMAG "rShuftiExec: case with %u matches, %u * %u iterations," KBLU " total elapsed time =" RST " %.3f s, " 
-               KBLU "average time per call =" RST " %.3f μs," KBLU " bandwidth = " RST " %.3f MB/s," KBLU " average bandwidth =" RST " %.3f MB/s \n",
-               M, size ,loops, total_sec, avg_time, max_bw, bandwitdh);
-    } else {
-        auto start = std::chrono::steady_clock::now();
-        for (int i = 0; i < loops; i++) {
-            rshuftiExec(lo, hi, kt1.get(), kt1.get() + size);
-        }
-        auto end = std::chrono::steady_clock::now();
-        total_sec += std::chrono::duration_cast<std::chrono::microseconds>(end - start).count();
-        /*calculate transferred size*/
-        transferred_size = size * loops;
-        /*calculate average time*/
-        avg_time = total_sec / loops;
-        /*convert microseconds to seconds*/
-        total_sec /= 1000000.0;
-        /*calculate maximum bandwidth*/
-        max_bw = transferred_size / total_sec;
-        /*convert to MB/s*/
-        max_bw /= 1048576.0;
-        /*calculate average bandwidth*/
-        bandwitdh = max_bw / loops;
-        printf(KMAG "rShuftiExec: case without matches, %u * %u iterations," KBLU " total elapsed time =" RST " %.3f s, " 
-               KBLU "average time per call =" RST " %.3f μs ," KBLU " bandwidth = " RST " %.3f MB/s," KBLU " average bandwidth =" RST " %.3f MB/s \n",
-               size ,loops, total_sec, avg_time, max_bw, bandwitdh);
-    }
-}
diff --git a/benchmarks/truffle.cpp b/benchmarks/truffle.cpp
deleted file mode 100644
index d521c2b1..00000000
--- a/benchmarks/truffle.cpp
+++ /dev/null
@@ -1,149 +0,0 @@
-#include "nfa/truffle.h"
-#include "nfa/trufflecompile.h"
-#include "benchmarks.hpp"
-#include <iostream>
-#include <chrono>
-#include <cstring>
-#include <ctime>
-#include <memory>
-
-void truffle_benchmarks(int size, int loops, int M, bool has_match) {
-    m128 lo, hi;
-    ue2::CharReach chars;
-    chars.set('a');
-    truffleBuildMasks(chars, (u8 *)&lo, (u8 *)&hi);
-    std::unique_ptr<u8 []> kt1 ( new u8[size] );
-    memset(kt1.get(),'b',size);
-    double total_sec = 0.0;            
-    u64a transferred_size = 0;
-    double bandwitdh = 0.0;
-    double max_bw = 0.0;
-    double avg_time = 0.0;
-    if (has_match) {
-        int pos = 0;
-        for(int j = 0; j < M; j++) {
-            kt1[pos] = 'b';
-            srand (time(NULL));
-            pos = rand() % size + 0;
-            kt1[pos] = 'a';
-            unsigned long act_size = 0;
-            auto start = std::chrono::steady_clock::now();
-            for(int i = 0; i < loops; i++) {
-                const u8 *res = truffleExec(lo, hi, kt1.get(), kt1.get() + size);
-                act_size += res - kt1.get();
-            }
-            auto end = std::chrono::steady_clock::now();
-            double dt = std::chrono::duration_cast<std::chrono::microseconds>(end - start).count();
-            total_sec += dt;
-            /*convert microseconds to seconds*/
-            total_sec /= 1000000.0;
-            /*calculate bandwidth*/
-            bandwitdh  += act_size / total_sec;
-            /*convert to MB/s*/
-            bandwitdh  = bandwitdh  / 1048576.0;
-            max_bw = std::max(bandwitdh ,max_bw);
-            /*calculate average time*/
-            avg_time += total_sec / loops;
-        }
-        avg_time /= M;
-        bandwitdh /= M;
-        /*convert average time to us*/
-        avg_time *= 1000000.0;
-        printf(KMAG "TruffleExec: case with %u matches, %u * %u iterations," KBLU " total elapsed time =" RST " %.3f s, " 
-               KBLU "average time per call =" RST " %.3f μs," KBLU " bandwidth = " RST " %.3f MB/s," KBLU " average bandwidth =" RST " %.3f MB/s \n",
-               M, size ,loops, total_sec, avg_time, max_bw, bandwitdh);
-    } else {
-        auto start = std::chrono::steady_clock::now();
-        for (int i = 0; i < loops; i++) {
-            truffleExec(lo, hi, kt1.get(), kt1.get() + size);
-        }
-        auto end = std::chrono::steady_clock::now();
-        total_sec += std::chrono::duration_cast<std::chrono::microseconds>(end - start).count();
-        /*calculate transferred size*/
-        transferred_size = size * loops;
-        /*calculate average time*/
-        avg_time = total_sec / loops;
-        /*convert microseconds to seconds*/
-        total_sec /= 1000000.0;
-        /*calculate maximum bandwidth*/
-        max_bw = transferred_size / total_sec;
-        /*convert to MB/s*/
-        max_bw /= 1048576.0;
-        /*calculate average bandwidth*/
-        bandwitdh = max_bw / loops;
-        printf(KMAG "TruffleExec case without matches, %u * %u iterations," KBLU " total elapsed time =" RST " %.3f s, " 
-               KBLU "average time per call =" RST " %.3f μs ," KBLU " bandwidth = " RST " %.3f MB/s," KBLU " average bandwidth =" RST " %.3f MB/s \n",
-               size ,loops, total_sec, avg_time, max_bw, bandwitdh);
-    }
-}
-
-
-void rtruffle_benchmarks(int size, int loops, int M, bool has_match) {  
-     m128 lo, hi;
-    ue2::CharReach chars;
-    chars.set('a');
-    truffleBuildMasks(chars, (u8 *)&lo, (u8 *)&hi);
-    std::unique_ptr<u8 []> kt1 ( new u8[size] );
-    memset(kt1.get(),'b',size);
-    double total_sec = 0.0;            
-    u64a transferred_size = 0;
-    double bandwitdh = 0.0;
-    double max_bw = 0.0;
-    double avg_time = 0.0;
-    if (has_match) {
-        int pos = 0;
-        for(int j = 0; j < M; j++) {
-            kt1[pos] = 'b';
-            srand (time(NULL));
-            pos = rand() % size + 0;
-            kt1[pos] = 'a';
-            unsigned long act_size = 0;
-            auto start = std::chrono::steady_clock::now();
-            for(int i = 0; i < loops; i++) {
-                const u8 *res = rtruffleExec(lo, hi, kt1.get(), kt1.get() + size);
-                act_size += res - kt1.get();
-            }
-            auto end = std::chrono::steady_clock::now();
-            double dt = std::chrono::duration_cast<std::chrono::microseconds>(end - start).count();
-            total_sec += dt;
-            /*convert microseconds to seconds*/
-            total_sec /= 1000000.0;
-            /*calculate bandwidth*/
-            bandwitdh  += act_size / total_sec;
-            /*convert to MB/s*/
-            bandwitdh  = bandwitdh  / 1048576.0;
-            max_bw = std::max(bandwitdh ,max_bw);
-            /*calculate average time*/
-            avg_time += total_sec / loops;
-        }
-        avg_time /= M;
-        bandwitdh /= M;
-        /*convert average time to us*/
-        avg_time *= 1000000.0;
-        printf(KMAG "rTruffleExec: case with %u matches, %u * %u iterations," KBLU " total elapsed time =" RST " %.3f s, " 
-               KBLU "average time per call =" RST " %.3f μs," KBLU " bandwidth = " RST " %.3f MB/s," KBLU " average bandwidth =" RST " %.3f MB/s \n",
-               M, size ,loops, total_sec, avg_time, max_bw, bandwitdh);
-    } else {
-        auto start = std::chrono::steady_clock::now();
-        for (int i = 0; i < loops; i++) {
-            rtruffleExec(lo, hi, kt1.get(), kt1.get() + size);
-        }
-        auto end = std::chrono::steady_clock::now();
-        total_sec += std::chrono::duration_cast<std::chrono::microseconds>(end - start).count();
-        /*calculate transferred size*/
-        transferred_size = size * loops;
-        /*calculate average time*/
-        avg_time = total_sec / loops;
-        /*convert microseconds to seconds*/
-        total_sec /= 1000000.0;
-        /*calculate maximum bandwidth*/
-        max_bw = transferred_size / total_sec;
-        /*convert to MB/s*/
-        max_bw /= 1048576.0;
-        /*calculate average bandwidth*/
-        bandwitdh = max_bw / loops;
-        printf(KMAG "rTruffleExec case without matches, %u * %u iterations," KBLU " total elapsed time =" RST " %.3f s, " 
-               KBLU "average time per call =" RST " %.3f μs ," KBLU " bandwidth = " RST " %.3f MB/s," KBLU " average bandwidth =" RST " %.3f MB/s \n",
-               size ,loops, total_sec, avg_time, max_bw, bandwitdh);
-    }
-}