vok

1 files changed, 52 insertions, 14 deletions

diff --git a/src/a3.c b/src/a3.c
index dfe9b34..16fcd9f 100644
--- a/src/a3.c
+++ b/src/a3.c
@@ -2,37 +2,75 @@
 #include <stdint.h>
 #include <stdlib.h>
 #include <string.h>
+#include <stdio.h>
 
 void                a3_init(uint64_t n);
 void                a3_loop(uint64_t x0,uint64_t x1);
 
 algo_t a3 = {
-  .name = "a3",
+  .name = "a3 aka SDI0",
   .nmax = 1ULL<<32,
   .init = a3_init,
   .loop = a3_loop,
 };
 
-static uint16_t    *m_fast; //mammory
-static uint32_t    *m_slow; //mammory
+#define CROSSOVER   24576
+#define M24_SIZE    3
+#define M24_RANGE   (1<<24)
 
-void a3_init(uint64_t n){
-  size_t size; //size
+static uint8_t     *m24;
+static uint32_t    *m32;
+static size_t       m24_n;
+static size_t       m32_n;
+
+static void         inner(uint32_t x);
 
-  t       = 0;
-  size    = sizeof(uint32_t)*n;
-  m_slow  = calloc(size,1);
-  if(!m_slow) fuk("memory\n");
+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");
 }
 
 void a3_loop(uint64_t x0,uint64_t x1){
   uint32_t x1_32 = x1;
-  for(uint32_t x = x0;x < x1_32;x++){
-    uint32_t r; //index readed
+  printf("MAIN %10u -> %10u\n",(uint32_t)x0,x1_32);
+  for(uint32_t x = x0;x < x1_32;){
+    if(x+M24_RANGE >= x1_32){
+      printf("     %10u -> %10u\n",x,x+x1_32);
+      for(;x < x1_32;x++){
+        inner(x);
+      }
+      break;
+    }else{
+      printf("     %10u -> %10u\n",x,x+M24_RANGE);
+      for(;x < x+M24_RANGE;x++){
+        inner(x);
+      }
+      //scale back
+    }
+  }
+}
+
+static void inner(uint32_t x){
+  return;
+  uint32_t r;
 
-    r         = m_slow[t];
-    m_slow[t] = x;
-    t         = r == 0 ? 0 : x-r;
+  if(t < CROSSOVER){
+    uint8_t *p;
+
+    r = 0;
+    p = m24+(t*M24_SIZE);
+    memcpy(&r,p,M24_SIZE);
+    memcpy(p,&x,M24_SIZE);
+  }else{
+    uint64_t i = t-m24_n;
+    r       = m32[i];
+    m32[i]  = x;
   }
+
+  t         = r == 0 ? 0 : x-r;
 }