vork

6 files changed, 44 insertions, 47 deletions

diff --git a/makefile b/makefile
index bc7ebc9..c7670c9 100644
--- a/makefile
+++ b/makefile
@@ -10,7 +10,7 @@ src_dir    := src
 CFLAGS     := -std=c11 -fdata-sections -ffunction-sections -D_POSIX_C_SOURCE=200809L
 CFLAGS     += -pedantic -Wall -Wextra -Wcast-qual -Wstrict-aliasing -Wpointer-arith -Wstrict-prototypes -Wmissing-prototypes -Wvla -Wno-parentheses -Wno-unused-variable -Wno-unused-function -Wno-unused-parameter -Wno-unused-label
 LDFLAGS    := -Wl,--gc-sections
-LDLIBS     := -lokelib2
+LDLIBS     := 
 ARFLAGS    := -rcs
 
 compile     = $(CC) $(CFLAGS) -MMD -c $< -o $@
@@ -27,7 +27,7 @@ bin        := $(bin_dir)/$(out_file)
 ifneq ($(MSYSTEM),)
   bin      := $(bin).exe
   LDFLAGS  += -static
-  RELFLAGS := -s -mwindows
+  RELFLAGS := -s
 else
   RELFLAGS := 
 endif
@@ -37,7 +37,7 @@ out        := $(bin)
 ifeq ($(build),debug)
   CFLAGS   += -g
 else ifeq ($(build),release)
-  CFLAGS   += -Ofast -march=native -mtune=native -fomit-frame-pointer
+  CFLAGS   += -Ofast -march=native -mtune=native
   LDFLAGS  += $(RELFLAGS) 
 else ifeq ($(build),analyze)
   CC       := clang
@@ -55,7 +55,7 @@ ifneq ($(run),)
   run_cmd  := @echo "run     $(bin)"
   run_cmd  += && cd bin
   run_cmd  += && echo "----------------------------------------------------------------" 
-  run_cmd  += && time ./$(out_file) 10000000
+  run_cmd  += && time ./$(out_file) 100000000
   run_cmd  += ; echo "----------------------------------------------------------------"
   run_cmd  += && cd ..
 endif
diff --git a/src/a0.c b/src/a0.c
index bee7ab9..3631ebc 100644
--- a/src/a0.c
+++ b/src/a0.c
@@ -3,14 +3,14 @@
 #include <stdlib.h>
 #include <string.h>
 
-void                a0_init       (uint64_t n);
-void                a0_loop       (uint64_t x0,uint64_t x1);
+void                a0_init(uint64_t n);
+void                a0_loop(uint64_t x0,uint64_t x1);
 
 algo_t a0 = {
-  .name     = "a0",
-  .nmax     = UINT32_MAX,
-  .init     = a0_init,
-  .loop     = a0_loop,
+  .name = "a0",
+  .nmax = 1ULL<<32,
+  .init = a0_init,
+  .loop = a0_loop,
 };
 
 static uint32_t    *m; //mammory
@@ -28,6 +28,7 @@ void a0_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
+
     r     = m[t];
     m[t]  = x;
     t     = r == 0 ? 0 : x-r;
diff --git a/src/a1.c b/src/a1.c
index e47b0b0..f2b0539 100644
--- a/src/a1.c
+++ b/src/a1.c
@@ -5,14 +5,14 @@
 
 #define BITS        36
 
-void                a1_init       (uint64_t n);
-void                a1_loop       (uint64_t x0,uint64_t x1);
+void                a1_init(uint64_t n);
+void                a1_loop(uint64_t x0,uint64_t x1);
 
 algo_t a1 = {
-  .name     = "a1",
-  .nmax     = 1ULL<<BITS,
-  .init     = a1_init,
-  .loop     = a1_loop,
+  .name = "a1",
+  .nmax = 1ULL<<BITS,
+  .init = a1_init,
+  .loop = a1_loop,
 };
 
 static uint8_t     *m; //mammory
@@ -21,7 +21,7 @@ void a1_init(uint64_t n){
   size_t size; //size
 
   t     = 0;
-  size  = sizeof(uint8_t)*n*BITS>>3;
+  size  = (sizeof(uint8_t)*n*BITS)>>3;
   m     = calloc(size,1);
   if(!m) fuk("memory\n");
 }
@@ -37,12 +37,12 @@ void a1_loop(uint64_t x0,uint64_t x1){
     w = x;
     memcpy(&r,p,5);
     if(t & 1){
-      r >>= 4;
       w <<= 4;
-      w  |= p[0] & 0x0F;
+      w  |= r & 0x000000000000000F;
+      r >>= 4;
     }else{
+      w  |= r & 0x000000F000000000;
       r  &= 0x0000000FFFFFFFFF; 
-      w  |= (uint64_t)(p[4] & 0xF0)<<32;
     }
     memcpy(p,&w,5);
     t = r == 0 ? 0 : x-r;
diff --git a/src/a2.c b/src/a2.c
index 13dcc87..47524e9 100644
--- a/src/a2.c
+++ b/src/a2.c
@@ -5,14 +5,14 @@
 
 #define WIDTH       3
 
-void                a2_init       (uint64_t n);
-void                a2_loop       (uint64_t x0,uint64_t x1);
+void                a2_init(uint64_t n);
+void                a2_loop(uint64_t x0,uint64_t x1);
 
 algo_t a2 = {
-  .name     = "a2",
-  .nmax     = 1<<24,
-  .init     = a2_init,
-  .loop     = a2_loop,
+  .name = "a2",
+  .nmax = 1<<(WIDTH*8),
+  .init = a2_init,
+  .loop = a2_loop,
 };
 
 static uint8_t     *m; //mammory
@@ -33,7 +33,7 @@ void a2_loop(uint64_t x0,uint64_t x1){
     uint8_t  *p; //addr of elemten
 
     r = 0;
-    p = &m[t*WIDTH];
+    p = m+(t*WIDTH);
     memcpy(&r,p,WIDTH);
     memcpy(p,&x,WIDTH);
     t = r == 0 ? 0 : x-r;
diff --git a/src/def.h b/src/def.h
index 68f0c41..2ce6a94 100644
--- a/src/def.h
+++ b/src/def.h
@@ -3,12 +3,14 @@
 
 #include <stdint.h>
 
+/*
 #ifdef _WIN32
   #include <windows.h>
-  #define aalloc(_a,_s) _aligned_malloc(_s,_a)
+  #define AALLOC(_a,_s) _aligned_malloc(_s,_a)
 #else
-  #define aalloc(_a,_s) aligned_alloc(_a,_s)
+  #define AALLOC(_a,_s) aligned_alloc(_a,_s)
 #endif
+*/
 
 typedef struct algo_t{
   char const       *name;
diff --git a/src/main.c b/src/main.c
index 69a8007..9beefbd 100644
--- a/src/main.c
+++ b/src/main.c
@@ -25,7 +25,7 @@ void fuk(char const *fmt,...){
 
 static void perse(int p){
   ftime_t t = ftimer_time(&ft);
-  float   f = (float)t/FTIME_SEC;
+  float   f = ftime_float(t);
   printf("%9.3f\t%3u%%\n",f,p);
   fflush(stdout);
 }
@@ -35,13 +35,12 @@ static void calc_fast(uint64_t n){
 }
 
 static void calc_perse(uint64_t n){
-  uint64_t    agurk;
-  uint64_t    outer;
-  uint64_t    last;
-  uint64_t    x0;
-  uint64_t    x1;
+  uint64_t agurk;
+  uint64_t x0;
+  uint64_t x1;
 
   agurk = n/100;
+  perse(0);
   for(uint64_t y = 0;y < 100;y++){
     x0 = y*agurk;
     x1 = x0+agurk;
@@ -55,7 +54,7 @@ static void calc_perse(uint64_t n){
 static void usake(void){
   printf("usake: numper [options] <n>\n");
   printf("  -p      display progress percentage\n");
-  printf("  -a<n>   use algorithm n\n");
+  printf("  -a<n>   use algorithm n (0,1,2)\n");
   fflush(stdout);
   exit(0);
 }
@@ -72,11 +71,11 @@ int main(int argc,char **argv){
   calc    = calc_fast;
   algo_i  = -1;
   
-  if(argc < 2){
-    usake();
-  }
-
   #define OPT(O) (strcmp(argv[x],O) == 0)
+  if(argc < 2) usake();
+  n = strtoull(argv[argc-1],0,10);
+  if((int64_t)n <= 0) usake();
+
   for(int x = 1;x < argc-1;x++){
     if(OPT("-p")){
       calc = calc_perse;
@@ -84,16 +83,13 @@ int main(int argc,char **argv){
              argv[x][1] == 'a'){
       algo_i = (size_t)argv[x][2]-(size_t)'0';
       if(algo_i >= ALGO_CUNT){
-        fuk("unrecognized algorithm: a%zu\n",algo_i);
+        fuk("unrecognized algorithm\n");
       }
     }else{
       fuk("unrecognized option: %s\n",argv[x]);
     }
   }
 
-  n = strtoull(argv[argc-1],0,10);
-  if((int64_t)n <= 0) usake();
-
   if(algo_i == (size_t)-1){
     if     (n <= a2.nmax) algo_i = 2;
     else if(n <= a0.nmax) algo_i = 0;
@@ -103,21 +99,19 @@ int main(int argc,char **argv){
   algo = algos[algo_i];
 
   if(n > algo->nmax){
-    fuk("n exceeds algorithm nmax\n");
+    fuk("n exceeds algorithm a%zu nmax (%llu > %llu)\n",algo_i,n,algo->nmax);
   }
 
   printf("algo  = %s\n",  algo->name);
   printf("n     = %llu\n",n);
   fflush(stdout);
 
-  //for(size_t x = 0;x < 10;x++){
   algo->init(n);
   ftimer_init(&ft);
   calc(n);
 
   printf("t     = %llu\n",t);
   fflush(stdout);
-  //}
 
   return 0;
 }