oke

4 files changed, 96 insertions, 23 deletions

diff --git a/makefile b/makefile
index b8fa6a6..fff6a5a 100644
--- a/makefile
+++ b/makefile
@@ -55,7 +55,7 @@ ifneq ($(run),)
   run_cmd  := @echo "run     $(bin)"
   run_cmd  += && cd bin
   run_cmd  += && echo "----------------------------------------------------------------" 
-  run_cmd  += && time ./$(out_file) -a5 100000000
+  run_cmd  += && time ./$(out_file) -a6 100000000
   run_cmd  += ; echo "----------------------------------------------------------------"
   run_cmd  += && cd ..
 endif
diff --git a/src/a5.c b/src/a5.c
index f36bfac..46ddcfd 100644
--- a/src/a5.c
+++ b/src/a5.c
@@ -16,7 +16,7 @@ algo_t a5 = {
 
 static uint32_t    *lastp;
 static uint64_t    *index;
-static uint32_t     filled;
+static uint64_t     filled;
 
 void a5_init(uint64_t n){
   size_t  lastp_size;
@@ -28,28 +28,17 @@ void a5_init(uint64_t n){
   lastp_size  = n*sizeof(uint32_t);
   index       = calloc(index_size,1);
   lastp       = calloc(lastp_size,1);
+  /*
+  index       = AALLOC(64,index_size);
+  lastp       = AALLOC(64,lastp_size);
+  memset(index,0x00,index_size);
+  memset(lastp,0x00,lastp_size);
+  */
 }
 
 void a5_loop(uint64_t x0,uint64_t x1){
-  uint32_t x1_32 = x1;
-
-  for(uint32_t x = x0;x < x1_32;x++){
-    uint32_t  next;
-
-    /*
-    uint64_t *index_ptr;
-    int       index_bit;
-
-    index_ptr = index+(t>>6);
-    index_bit = t & 63;
-
-    if(t < filled || ((*index_ptr>>index_bit) & 1)){
-      next        = x-lastp[t];
-    }else{
-      next        = 0;
-      *index_ptr |= 1ULL<<index_bit;
-    }
-    */
+  for(uint64_t x = x0;x < x1;x++){
+    uint64_t  next;
 
     if(t < filled){
       next = x-lastp[t]; 
diff --git a/src/a6.c b/src/a6.c
new file mode 100644
index 0000000..c865d20
--- /dev/null
+++ b/src/a6.c
@@ -0,0 +1,84 @@
+
+#include "def.h"
+#include <stdint.h>
+#include <stdlib.h>
+#include <string.h>
+#include <stdio.h>
+
+void                a6_init(uint64_t n);
+void                a6_loop(uint64_t x0,uint64_t x1);
+
+algo_t a6 = {
+  .name = "a6 (NT5 tek)",
+  .nmax = 1ULL<<32,
+  .init = a6_init,
+  .loop = a6_loop,
+};
+
+static uint32_t    *lastp;
+static uint64_t    *index;
+static uint64_t    *index2;
+static uint64_t     filled;
+
+void a6_init(uint64_t n){
+  size_t  index_n;
+  size_t  index2_n;
+  size_t  index_size;
+  size_t  index2_size;
+  size_t  lastp_size;
+
+  t           = 0;
+  filled      = 0;
+  index_n     = (n-1)/64+1;
+  index2_n    = (index_n-1)/64+1;
+  index_size  = index_n*sizeof(uint64_t);
+  index2_size = index2_n*sizeof(uint64_t);
+  lastp_size  = n*sizeof(uint32_t);
+  index       = calloc(index_size,1);
+  index2      = calloc(index2_size,1);
+  lastp       = calloc(lastp_size,1);
+}
+
+void a6_loop(uint64_t x0,uint64_t x1){
+  for(uint64_t x = x0;x < x1;x++){
+    uint64_t  next;
+
+    if(t < filled){
+      next = x-lastp[t]; 
+    }else{
+      uint64_t *index2_ptr;
+      uint64_t  index2_bit;
+      uint64_t *index_ptr;
+      uint64_t  index_bit;
+
+      index2_ptr = index2+(t/64/64);
+      index2_bit = 1ULL<<(t/64%64);
+
+      if(*index2_ptr & index2_bit){
+        next = x-lastp[t]; 
+      }else{
+        index_ptr = index+(t/64);
+        index_bit = 1ULL<<(t%64);
+
+        if(*index_ptr & index_bit){
+          next = x-lastp[t]; 
+        }else{
+          next        = 0;
+          *index_ptr |= index_bit;
+          if(*index_ptr == 0xFFFFFFFFFFFFFFFF){
+            *index2_ptr |= index2_bit;
+          }
+        }
+      }
+    }
+
+    lastp[t] = x;
+
+    if(t == filled){
+      while(lastp[++filled]);
+    }
+
+    t = next;
+  }
+}
+
diff --git a/src/def.h b/src/def.h
index 4053dcf..0a7acab 100644
--- a/src/def.h
+++ b/src/def.h
@@ -3,14 +3,12 @@
 
 #include <stdint.h>
 
-/*
 #ifdef _WIN32
   #include <windows.h>
   #define AALLOC(_a,_s) _aligned_malloc(_s,_a)
 #else
   #define AALLOC(_a,_s) aligned_alloc(_a,_s)
 #endif
-*/
 
 typedef struct algo_t{
   char const       *name;
@@ -28,6 +26,7 @@ extern algo_t       a2;
 extern algo_t       a3;
 extern algo_t       a4;
 extern algo_t       a5;
+extern algo_t       a6;
 
 extern algo_t       nt5;
 
@@ -38,6 +37,7 @@ static algo_t *algos[] = {
   [3] = &a3,
   [4] = &a4,
   [5] = &a5,
+  [6] = &a6,
 };
 
 #define ALGO_CUNT   (sizeof(algos)/sizeof(algo_t *))