vork

1 files changed, 31 insertions, 44 deletions

diff --git a/src/a3.c b/src/a3.c
index 61c1b71..8645e97 100644
--- a/src/a3.c
+++ b/src/a3.c
@@ -14,22 +14,21 @@ algo_t a3 = {
   .loop = a3_loop,
 };
 
-#define CROSSOVER   24576
-#define M24_SIZE    3
-#define M24_RANGE   (1<<24)
+#define CROSSOVER             32768
+#define CROSSOVER_MASK        0xFFFFFFFFFFFF8000
+#define M24_SIZE              3
+#define M24_RANGE             (1<<24)
 
 static uint8_t     *m24;
 static uint32_t    *m32;
 static size_t       m24_n;
 static size_t       m32_n;
-static uint32_t     bias;
 
 static void         inner     (uint32_t x);
 static void         inner_bias(uint32_t x);
 
 void a3_init(uint64_t n){
   t     = 0;
-  bias  = 0;
   m24_n = n > CROSSOVER ? CROSSOVER : n;
   m32_n = n-m24_n;
   m24   = calloc(sizeof(uint8_t)*M24_SIZE*m24_n,1);
@@ -38,68 +37,56 @@ void a3_init(uint64_t n){
 }
 
 void a3_loop(uint64_t x0,uint64_t x1){
-  uint32_t x1_32 = x1;
-  printf("MAIN %10u -> %10u\n\n",(uint32_t)x0,x1_32);
-  for(uint32_t x = x0;x < x1_32;){
-    if(x+M24_RANGE >= x1_32){
-      printf("LAST %10u -> %10u\n",x,x1_32);
-      for(;x < x1_32;x++){
-        inner_bias(x); //TODO blabla
-      }
-      break;
-    }else{
-      printf("     %10u -> %10u\n",x,x+M24_RANGE);
-      uint32_t max = x+M24_RANGE;
-      if(bias == 0){
-        for(;x < max;x++){
-          inner(x);
-        }
-      }else{
-        for(;x < max;x++){
-          inner_bias(x);
-        }
-      }
-      bias += M24_RANGE;
-    }
+  uint32_t x;
+  uint32_t x1_32;
+  uint32_t x1_first;
+
+  x1_32     = x1;
+  x1_first  = x0+M24_RANGE;
+  if(x1_first > x1_32) x1_first = x1_32;
+
+  for(x = x0;x < x1_first;x++){
+    inner(x);
+  }
+  for(;x < x1_32;x++){
+    inner_bias(x);
   }
 }
 
 static void inner(uint32_t x){
   uint32_t r;
 
-  if(t < CROSSOVER){
+  if(t & CROSSOVER_MASK){
+    uint64_t i = t-m24_n;
+    r       = m32[i];
+    m32[i]  = x;
+  }else{
     uint8_t *p;
 
     r = 0;
     p = m24+(t*M24_SIZE);
     memcpy(&r,p,M24_SIZE);
     memcpy(p,&x,M24_SIZE);
-    t = r == 0 ? 0 : x-r;
-  }else{
-    uint64_t i = t-m24_n;
-    r       = m32[i];
-    m32[i]  = x;
-    t = r == 0 ? 0 : x-r;
   }
+  t = r == 0 ? 0 : x-r;
 }
 
 static void inner_bias(uint32_t x){
   uint32_t r;
 
-  if(t < CROSSOVER){
+  if(t & CROSSOVER_MASK){
+    uint64_t i = t-m24_n;
+    r       = m32[i];
+    m32[i]  = x;
+    t = r == 0 ? 0 : x-r;
+  }else{
     uint8_t *p;
 
     r = 0;
     p = m24+(t*M24_SIZE);
-    uint32_t i = x & 0x00FFFFFF;
     memcpy(&r,p,M24_SIZE);
-    memcpy(p,&i,M24_SIZE);
-    t = (i-r) & 0x00FFFFFF;
-  }else{
-    uint64_t i = t-m24_n;
-    r       = m32[i];
-    m32[i]  = x;
-    t = r == 0 ? 0 : x-r;
+    memcpy(p,&x,M24_SIZE);
+    t = (x-r) & 0x00FFFFFF;
   }
 }