start vork

1 files changed, 40 insertions, 13 deletions

diff --git a/src/a3.c b/src/a3.c
index 0c36fdd..265b0de 100644
--- a/src/a3.c
+++ b/src/a3.c
@@ -17,24 +17,33 @@ algo_t a3 = {
 #define M24_SIZE              3
 #define M24_RANGE             (1<<(M24_SIZE*8))
 
+#define M32_SIZE              4
+
 #define CROSSOVER             (1<<15)
 #define CROSSOVER_MASK        (~(CROSSOVER-1))
 
-static uint8_t     *m24;
-static uint32_t    *m32;
+//static uint8_t     *m24;
+//static uint8_t     *m32;
+static uint8_t     *m;
+static uint8_t     *m_32ik; //TMP
 static size_t       m24_n;
 static size_t       m32_n;
+static size_t       m32_off;
 
 static void         inner     (uint32_t x);
 static void         inner_bias(uint32_t x);
 
 void a3_init(uint64_t n){
-  t     = 0;
-  m24_n = n > CROSSOVER ? CROSSOVER : n;
-  m32_n = n-m24_n;
-  m24   = calloc(sizeof(uint8_t)*M24_SIZE*m24_n,1);
-  m32   = calloc(sizeof(uint32_t)*m32_n,1);
-  if(!m24 || !m32) fuk("memory\n");
+  size_t size;
+  t       = 0;
+  m24_n   = n > CROSSOVER ? CROSSOVER : n;
+  m32_n   = n-m24_n;
+  size    = sizeof(uint8_t)*M24_SIZE*m24_n;
+  m32_off = size;
+  size   += sizeof(uint8_t)*M32_SIZE*m32_n;
+  m       = calloc(size,1);
+  m_32ik  = m+m32_off;
+  if(!m) fuk("memory\n");
 }
 
 void a3_loop(uint64_t x0,uint64_t x1){
@@ -59,14 +68,23 @@ static void inner(uint32_t x){
 
   //if(t >= CROSSOVER){
   if(t & CROSSOVER_MASK){
+    /*
     uint64_t i = t-m24_n;
     r       = m32[i];
     m32[i]  = x;
+    */
+
+    uint8_t *p;
+
+    r = 0;
+    p = m_32ik+(t*M32_SIZE);
+    memcpy(&r,p,M32_SIZE);
+    memcpy(p,&x,M32_SIZE);
   }else{
     uint8_t *p;
 
     r = 0;
-    p = m24+(t*M24_SIZE);
+    p = m+(t*M24_SIZE);
     memcpy(&r,p,M24_SIZE);
     memcpy(p,&x,M24_SIZE);
   }
@@ -78,18 +96,27 @@ static void inner_bias(uint32_t x){
 
   //if(t >= CROSSOVER){
   if(t & CROSSOVER_MASK){
-    uint64_t i = t-m24_n;
-    r       = m32[i];
-    m32[i]  = x;
+    uint8_t *p;
+
+    r = 0;
+    p = m_32ik+(t*M32_SIZE);
+    memcpy(&r,p,M32_SIZE);
+    memcpy(p,&x,M32_SIZE);
     t = r == 0 ? 0 : x-r;
   }else{
     uint8_t *p;
 
     r = 0;
-    p = m24+(t*M24_SIZE);
+    p = m+(t*M24_SIZE);
     memcpy(&r,p,M24_SIZE);
     memcpy(p,&x,M24_SIZE);
     t = (x-r) & 0x00FFFFFF;
   }
+
+  return;
+  uint32_t c24;
+
+  c24 = ((int32_t)t-CROSSOVER)>>31; //FFFFetc.. if t < CROSSOVER
+  
 }