v_mul

5 files changed, 94 insertions, 14 deletions

diff --git a/inc/okeffa/iron.h b/inc/okeffa/iron.h
index 71b3aa5..e31b680 100644
--- a/inc/okeffa/iron.h
+++ b/inc/okeffa/iron.h
@@ -1,11 +1,9 @@
 #ifndef OKEFFA_IRON_H
 #define OKEFFA_IRON_H
 
-//32bit system
-//#define IRON_MACHINEBITNESSLOG2         5
-
-//64bit system
-#define IRON_MACHINEBITNESSLOG2         6
+//#define IRON_MACHINEBITNESSLOG2         4 //16bit system
+//#define IRON_MACHINEBITNESSLOG2         5 //32bit system
+#define IRON_MACHINEBITNESSLOG2         6 //64bit system
 
 #define IRON_MACHINEBITNESS             (1<<IRON_MACHINEBITNESSLOG2)
 #define IRON_BYTEBITS                   8
diff --git a/inc/okeffa/w_mul.h b/inc/okeffa/w_mul.h
new file mode 100644
index 0000000..b5dfe4f
--- /dev/null
+++ b/inc/okeffa/w_mul.h
@@ -0,0 +1,69 @@
+#ifndef OKEFFA_W_MUL_H
+#define OKEFFA_W_MUL_H
+
+#include <okeffa/word.h>
+
+inline word_t       hw_mul  (hword_t a,hword_t b);
+
+inline hword_t      w_lo    (word_t a);
+inline hword_t      w_hi    (word_t a);
+inline void         w_mul   (word_t a,word_t b,word_t *o_lo,word_t *o_hi);
+
+////////////////////////////////////////////////////////////////
+
+#ifdef HW_MUL_IRON
+inline word_t hw_mul(hword_t a,hword_t b){
+  return (word_t)a*(word_t)b;
+}
+#else
+
+#define _HW_MUL_BIT   \
+  gm    = -(bs & 1);  \
+  ab   += as & gm;    \
+  bs  >>= 1;          \
+  as  <<= 1;
+
+inline word_t hw_mul(hword_t a,hword_t b){
+  word_t as = a;
+  word_t bs = b;
+  word_t gm;
+  word_t ab = 0;
+  for(size_t x = 0;x < HWORD_BYTENESS;x++){
+    _HW_MUL_BIT;
+    _HW_MUL_BIT;
+    _HW_MUL_BIT;
+    _HW_MUL_BIT;
+    _HW_MUL_BIT;
+    _HW_MUL_BIT;
+    _HW_MUL_BIT;
+    _HW_MUL_BIT;
+  }
+  return ab;
+}
+#undef _HW_MUL_BIT
+#endif
+
+inline hword_t w_lo(word_t a){
+  return (hword_t)a;
+}
+
+inline hword_t w_hi(word_t a){
+  return (hword_t)(a>>HWORD_BITNESS);
+}
+
+inline void w_mul(word_t a,word_t b,word_t *o_lo,word_t *o_hi){
+  hword_t al  = w_lo(a);
+  hword_t ah  = w_hi(a);
+  hword_t bl  = w_lo(b);
+  hword_t bh  = w_hi(b);
+  word_t  ll  = hw_mul(al,bl);
+  word_t  lh  = hw_mul(al,bh);
+  word_t  hl  = hw_mul(ah,bl);
+  word_t  hh  = hw_mul(ah,bh);
+  word_t  cl  = w_hi(w_hi(ll)+w_lo(lh)+w_lo(hl));
+  *o_lo       = ll+((lh+hl)<<HWORD_BITNESS);
+  *o_hi       = hh+w_hi(hl)+w_hi(lh)+cl;
+}
+
+#endif
+
diff --git a/inc/okeffa/word.h b/inc/okeffa/word.h
index dc8bdbf..1083061 100644
--- a/inc/okeffa/word.h
+++ b/inc/okeffa/word.h
@@ -9,18 +9,23 @@
 #define WORD_BITNESSLOG2      IRON_MACHINEBITNESSLOG2
 #define WORD_BYTENESS         (WORD_BITNESS/IRON_BYTEBITS)
 
-#if   WORD_BITNESS ==  8
-typedef uint8_t  word_t;
-#elif WORD_BITNESS == 16
-typedef uint16_t word_t;
+#define HWORD_BITNESS         (WORD_BITNESS/2)
+#define HWORD_BITNESSLOG2     (WORD_BITNESSLOG2/2)
+#define HWORD_BYTENESS        (WORD_BYTENESS/2)
+
+#if   WORD_BITNESS == 16
+typedef uint16_t    word_t;
+typedef uint8_t     hword_t;
 #elif WORD_BITNESS == 32
-typedef uint32_t word_t;
+typedef uint32_t    word_t;
+typedef uint16_t    hword_t;
 #elif WORD_BITNESS == 64
-typedef uint64_t word_t;
+typedef uint64_t    word_t;
+typedef uint32_t    hword_t;
 #else
 #error cant into this bitness
 #endif
 
-typedef word_t wbool_t;
+typedef word_t      wbool_t;
 
 #endif
diff --git a/makefile b/makefile
index d38dd30..0a5fbbd 100644
--- a/makefile
+++ b/makefile
@@ -65,9 +65,9 @@ out        := $(bin_dir)/$(out)
 
 ifneq ($(run),)
   run_cmd  := @echo "run     $(out)" 1>&2
-  #run_cmd  += && ./$(out) 256 16
+  run_cmd  += && ./$(out) 256 16
   #run_cmd  += && time $(out) 2048 16 < prog/ch7
-  run_cmd  += && prog/ch8.sh $(out)
+  #run_cmd  += && prog/ch8.sh $(out)
 endif
 
 ifneq ($(debug),)
diff --git a/src/w_mul.c b/src/w_mul.c
new file mode 100644
index 0000000..fb2715e
--- /dev/null
+++ b/src/w_mul.c
@@ -0,0 +1,8 @@
+#include <okeffa/w_mul.h>
+
+extern word_t       hw_mul  (hword_t a,hword_t b);
+
+extern hword_t      w_lo    (word_t a);
+extern hword_t      w_hi    (word_t a);
+extern void         w_mul   (word_t a,word_t b,word_t *o_lo,word_t *o_hi);
+