Changeset 3441

Timestamp:

Sep 7, 2013, 3:05:51 PM (6 years ago)

Author:

linmengl

Message:

make avx2 can run now

Location:

trunk/lib

Files:

• bitblock256.hpp (modified) (6 diffs)
• idisa256.hpp (modified) (1 diff)
• idisa_cpp/idisa_avx.cpp (modified) (106 diffs)
• idisa_cpp/idisa_avx2.cpp (added)
• idisa_cpp/idisa_sse2.cpp (modified) (2 diffs)

trunk/lib/bitblock256.hpp

@@ -18 +18 @@
 union ubitblock {
         bitblock256_t _256;
-        bitblock256_t _128[sizeof(bitblock256_t)/sizeof(bitblock256_t)];
+        bitblock128_t _128[sizeof(bitblock256_t)/sizeof(bitblock256_t)];
         uint64_t _64[sizeof(bitblock256_t)/sizeof(uint64_t)];
         uint32_t _32[sizeof(bitblock256_t)/sizeof(uint32_t)];
@@ -28 +28 @@
 typedef bitblock256_t carry_t;
 
+static IDISA_ALWAYS_INLINE void add_ci_co(bitblock256_t x, bitblock256_t y, carry_t carry_in, carry_t & carry_out, bitblock256_t & sum);
+static IDISA_ALWAYS_INLINE void sub_bi_bo(bitblock256_t x, bitblock256_t y, carry_t borrow_in, carry_t & borrow_out, bitblock256_t & difference);
+static IDISA_ALWAYS_INLINE void adv_ci_co(bitblock256_t cursor, bitblock256_t carry_in, bitblock256_t & carry_out, bitblock256_t & rslt);
+
+
+
+
 static IDISA_ALWAYS_INLINE bitblock256_t carry2bitblock(carry_t carry);
 static IDISA_ALWAYS_INLINE carry_t bitblock2carry(bitblock256_t carry);
@@ -46 +53 @@
 static IDISA_ALWAYS_INLINE carry_t bitblock2carry(bitblock256_t carry) {  return carry;}
 
-static inline void add_ci_co(bitblock256_t x, bitblock256_t y, carry_t carry_in, carry_t & carry_out, bitblock256_t & sum);
-static inline void sub_bi_bo(bitblock256_t x, bitblock256_t y, carry_t borrow_in, carry_t & borrow_out, bitblock256_t & difference);
-static IDISA_ALWAYS_INLINE void adv_ci_co(bitblock256_t cursor, bitblock256_t carry_in, bitblock256_t & carry_out, bitblock256_t & rslt);
-
-
 static inline void add_ci_co(bitblock256_t x, bitblock256_t y, carry_t carry_in, carry_t & carry_out, bitblock256_t & sum) {
-bitblock256_t all_ones = simd256<1>::constant<1>();
-bitblock256_t gen = simd_and(x, y);
-bitblock256_t prop = simd_xor(x, y);
-bitblock256_t partial_sum = simd256<64>::add(x, y);
-bitblock256_t carry = simd_or(gen, simd_andc(prop, partial_sum));
-bitblock256_t bubble = simd256<64>::eq(partial_sum, all_ones);
-uint64_t carry_mask = hsimd256<64>::signmask(carry) * 2 + convert(carry_in);
-uint64_t bubble_mask = hsimd256<64>::signmask(bubble);
-uint64_t carry_scan_thru_bubbles = (carry_mask + bubble_mask) &~ bubble_mask;
-uint64_t increments = carry_scan_thru_bubbles | (carry_scan_thru_bubbles - carry_mask);
-carry_out = convert(increments >> 4);
-uint64_t spread = 0x0000200040008001 * increments & 0x0001000100010001;
-sum = simd256<64>::add(partial_sum, _mm256_cvtepu8_epi64(avx_select_lo128(convert(spread))));
+  bitblock256_t all_ones = simd256<1>::constant<1>();
+  bitblock256_t gen = simd_and(x, y);
+  bitblock256_t prop = simd_xor(x, y);
+  bitblock256_t partial_sum = simd256<64>::add(x, y);
+  bitblock256_t carry = simd_or(gen, simd_andc(prop, partial_sum));
+  bitblock256_t bubble = simd256<64>::eq(partial_sum, all_ones);
+  uint64_t carry_mask = hsimd256<64>::signmask(carry) * 2 + convert(carry_in);
+  uint64_t bubble_mask = hsimd256<64>::signmask(bubble);
+  uint64_t carry_scan_thru_bubbles = (carry_mask + bubble_mask) &~ bubble_mask;
+  uint64_t increments = carry_scan_thru_bubbles | (carry_scan_thru_bubbles - carry_mask);
+  carry_out = convert(increments >> 4);
+  uint64_t spread = 0x0000200040008001 * increments & 0x0001000100010001;
+  sum = simd256<64>::add(partial_sum, _mm256_cvtepu8_epi64(avx_select_lo128(convert(spread))));
 }
 
@@ -74 +76 @@
         difference = simd256<128>::sub(partial, b1);
         borrow_out = simd_or(gen, simd_and(prop, difference));
-
 }
 
@@ -83 +84 @@
         rslt = simd_or(simd256<64>::add(cursor, cursor), low_bits);
 }
-
 
 
@@ -251 +251 @@
 }
 
-IDISA_ALWAYS_INLINE uint64_t convert (bitblock256_t v)
+IDISA_ALWAYS_INLINE uint64_t convert(bitblock256_t v)
 {
   return (uint64_t) mvmd256<64>::extract<0>(v);

trunk/lib/idisa256.hpp

@@ -9 +9 @@
 #define IDISA256_HPP
 
+#if defined USE_AVX
 #include "idisa_cpp/idisa_avx.cpp"
+#else
+#include "idisa_cpp/idisa_avx2.cpp"
 #endif
+                
+#endif

trunk/lib/idisa_cpp/idisa_avx.cpp

@@ -16 +16 @@
 
 typedef __m256 bitblock256_t;
+                        
 template <uint32_t fw>
 class simd256
@@ -25 +26 @@
         static IDISA_ALWAYS_INLINE bitblock256_t umult(bitblock256_t arg1, bitblock256_t arg2);
         static IDISA_ALWAYS_INLINE bitblock256_t ult(bitblock256_t arg1, bitblock256_t arg2);
-        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock256_t srli(bitblock256_t arg1);
+        static IDISA_ALWAYS_INLINE bitblock256_t all(bitblock256_t arg1);
+        template <uint16_t sh> static IDISA_ALWAYS_INLINE bitblock256_t srli(bitblock256_t arg1);
         static IDISA_ALWAYS_INLINE bitblock256_t ctz(bitblock256_t arg1);
         static IDISA_ALWAYS_INLINE bitblock256_t eq(bitblock256_t arg1, bitblock256_t arg2);
@@ -31 +33 @@
         static IDISA_ALWAYS_INLINE bitblock256_t neg(bitblock256_t arg1);
         static IDISA_ALWAYS_INLINE bitblock256_t himask();
-        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock256_t slli(bitblock256_t arg1);
+        template <uint16_t sh> static IDISA_ALWAYS_INLINE bitblock256_t slli(bitblock256_t arg1);
         static IDISA_ALWAYS_INLINE bitblock256_t ifh(bitblock256_t arg1, bitblock256_t arg2, bitblock256_t arg3);
         static IDISA_ALWAYS_INLINE bitblock256_t sub(bitblock256_t arg1, bitblock256_t arg2);
@@ -37 +39 @@
         static IDISA_ALWAYS_INLINE bitblock256_t lomask();
         static IDISA_ALWAYS_INLINE bitblock256_t umin(bitblock256_t arg1, bitblock256_t arg2);
-        template <uint64_t val> static IDISA_ALWAYS_INLINE bitblock256_t constant();
+        template <typename FieldType<fw>::T val> static IDISA_ALWAYS_INLINE bitblock256_t constant();
         static IDISA_ALWAYS_INLINE bitblock256_t min(bitblock256_t arg1, bitblock256_t arg2);
         static IDISA_ALWAYS_INLINE bitblock256_t umax(bitblock256_t arg1, bitblock256_t arg2);
         static IDISA_ALWAYS_INLINE bitblock256_t abs(bitblock256_t arg1);
         static IDISA_ALWAYS_INLINE bitblock256_t xor_hl(bitblock256_t arg1);
-        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock256_t srai(bitblock256_t arg1);
+        static IDISA_ALWAYS_INLINE bitblock256_t any(bitblock256_t arg1);
+        template <uint16_t sh> static IDISA_ALWAYS_INLINE bitblock256_t srai(bitblock256_t arg1);
         static IDISA_ALWAYS_INLINE bitblock256_t lt(bitblock256_t arg1, bitblock256_t arg2);
         static IDISA_ALWAYS_INLINE bitblock256_t add(bitblock256_t arg1, bitblock256_t arg2);
@@ -56 +59 @@
         static IDISA_ALWAYS_INLINE bitblock256_t packss(bitblock256_t arg1, bitblock256_t arg2);
         static IDISA_ALWAYS_INLINE bitblock256_t packh(bitblock256_t arg1, bitblock256_t arg2);
-        static IDISA_ALWAYS_INLINE uint64_t signmask(bitblock256_t arg1);
+        static IDISA_ALWAYS_INLINE typename FieldType<256/fw>::T signmask(bitblock256_t arg1);
         static IDISA_ALWAYS_INLINE bitblock256_t packl(bitblock256_t arg1, bitblock256_t arg2);
         static IDISA_ALWAYS_INLINE bitblock256_t min_hl(bitblock256_t arg1, bitblock256_t arg2);
@@ -78 +81 @@
 {
 public:
-        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock256_t dsrli(bitblock256_t arg1, bitblock256_t arg2);
-        static IDISA_ALWAYS_INLINE bitblock256_t fill(uint64_t val1);
-        template <uint64_t pos> static IDISA_ALWAYS_INLINE uint64_t extract(bitblock256_t arg1);
-        template <uint64_t pos> static IDISA_ALWAYS_INLINE bitblock256_t splat(bitblock256_t arg1);
-        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock256_t slli(bitblock256_t arg1);
-        static IDISA_ALWAYS_INLINE bitblock256_t fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4);
-        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock256_t srli(bitblock256_t arg1);
-        static IDISA_ALWAYS_INLINE bitblock256_t fill2(uint64_t val1, uint64_t val2);
-        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock256_t dslli(bitblock256_t arg1, bitblock256_t arg2);
-        static IDISA_ALWAYS_INLINE bitblock256_t fill8(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8);
-        static IDISA_ALWAYS_INLINE bitblock256_t fill16(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8, uint64_t val9, uint64_t val10, uint64_t val11, uint64_t val12, uint64_t val13, uint64_t val14, uint64_t val15, uint64_t val16);
+        template <uint16_t sh> static IDISA_ALWAYS_INLINE bitblock256_t dsrli(bitblock256_t arg1, bitblock256_t arg2);
+        static IDISA_ALWAYS_INLINE bitblock256_t fill(typename FieldType<fw>::T val1);
+        template <uint8_t pos> static IDISA_ALWAYS_INLINE typename FieldType<fw>::T extract(bitblock256_t arg1);
+        template <uint16_t pos> static IDISA_ALWAYS_INLINE bitblock256_t splat(bitblock256_t arg1);
+        template <uint16_t sh> static IDISA_ALWAYS_INLINE bitblock256_t slli(bitblock256_t arg1);
+        static IDISA_ALWAYS_INLINE bitblock256_t fill4(typename FieldType<fw>::T val1, typename FieldType<fw>::T val2, typename FieldType<fw>::T val3, typename FieldType<fw>::T val4);
+        template <uint16_t sh> static IDISA_ALWAYS_INLINE bitblock256_t srli(bitblock256_t arg1);
+        static IDISA_ALWAYS_INLINE bitblock256_t fill2(typename FieldType<fw>::T val1, typename FieldType<fw>::T val2);
+        template <uint16_t sh> static IDISA_ALWAYS_INLINE bitblock256_t dslli(bitblock256_t arg1, bitblock256_t arg2);
+        static IDISA_ALWAYS_INLINE bitblock256_t fill8(typename FieldType<fw>::T val1, typename FieldType<fw>::T val2, typename FieldType<fw>::T val3, typename FieldType<fw>::T val4, typename FieldType<fw>::T val5, typename FieldType<fw>::T val6, typename FieldType<fw>::T val7, typename FieldType<fw>::T val8);
+        static IDISA_ALWAYS_INLINE bitblock256_t fill16(typename FieldType<fw>::T val1, typename FieldType<fw>::T val2, typename FieldType<fw>::T val3, typename FieldType<fw>::T val4, typename FieldType<fw>::T val5, typename FieldType<fw>::T val6, typename FieldType<fw>::T val7, typename FieldType<fw>::T val8, typename FieldType<fw>::T val9, typename FieldType<fw>::T val10, typename FieldType<fw>::T val11, typename FieldType<fw>::T val12, typename FieldType<fw>::T val13, typename FieldType<fw>::T val14, typename FieldType<fw>::T val15, typename FieldType<fw>::T val16);
 };
 
@@ -95 +98 @@
 public:
         static IDISA_ALWAYS_INLINE bitblock256_t load_unaligned(const bitblock256_t* arg1);
-        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock256_t srli(bitblock256_t arg1);
+        template <uint16_t sh> static IDISA_ALWAYS_INLINE bitblock256_t srli(bitblock256_t arg1);
         static IDISA_ALWAYS_INLINE void store_aligned(bitblock256_t arg1, bitblock256_t* arg2);
         static IDISA_ALWAYS_INLINE bool all(bitblock256_t arg1);
         static IDISA_ALWAYS_INLINE bool any(bitblock256_t arg1);
-        static IDISA_ALWAYS_INLINE uint64_t popcount(bitblock256_t arg1);
-        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock256_t slli(bitblock256_t arg1);
+        static IDISA_ALWAYS_INLINE uint16_t popcount(bitblock256_t arg1);
+        template <uint16_t sh> static IDISA_ALWAYS_INLINE bitblock256_t slli(bitblock256_t arg1);
         static IDISA_ALWAYS_INLINE bitblock256_t load_aligned(const bitblock256_t* arg1);
         static IDISA_ALWAYS_INLINE void store_unaligned(bitblock256_t arg1, bitblock256_t* arg2);
@@ -110 +113 @@
 IDISA_ALWAYS_INLINE bitblock256_t simd_andc(bitblock256_t arg1, bitblock256_t arg2);
 IDISA_ALWAYS_INLINE bitblock256_t simd_or(bitblock256_t arg1, bitblock256_t arg2);
+IDISA_ALWAYS_INLINE bitblock256_t simd_and(bitblock256_t arg1, bitblock256_t arg2);
 IDISA_ALWAYS_INLINE bitblock256_t simd_xor(bitblock256_t arg1, bitblock256_t arg2);
-IDISA_ALWAYS_INLINE bitblock256_t simd_and(bitblock256_t arg1, bitblock256_t arg2);
 template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<1>::max(bitblock256_t arg1, bitblock256_t arg2);
 template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::max(bitblock256_t arg1, bitblock256_t arg2);
@@ -156 +159 @@
 template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::ult(bitblock256_t arg1, bitblock256_t arg2);
 template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::ult(bitblock256_t arg1, bitblock256_t arg2);
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<1>::lt(bitblock256_t arg1, bitblock256_t arg2);
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::lt(bitblock256_t arg1, bitblock256_t arg2);
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::lt(bitblock256_t arg1, bitblock256_t arg2);
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::lt(bitblock256_t arg1, bitblock256_t arg2);
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::lt(bitblock256_t arg1, bitblock256_t arg2);
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::lt(bitblock256_t arg1, bitblock256_t arg2);
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::lt(bitblock256_t arg1, bitblock256_t arg2);
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::lt(bitblock256_t arg1, bitblock256_t arg2);
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::lt(bitblock256_t arg1, bitblock256_t arg2);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::srli(bitblock256_t arg1);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::srli(bitblock256_t arg1);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::srli(bitblock256_t arg1);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::srli(bitblock256_t arg1);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::srli(bitblock256_t arg1);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::srli(bitblock256_t arg1);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::srli(bitblock256_t arg1);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::srli(bitblock256_t arg1);
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::all(bitblock256_t arg1);
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::all(bitblock256_t arg1);
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::all(bitblock256_t arg1);
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::all(bitblock256_t arg1);
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::all(bitblock256_t arg1);
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::all(bitblock256_t arg1);
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::all(bitblock256_t arg1);
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::all(bitblock256_t arg1);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::srli(bitblock256_t arg1);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::srli(bitblock256_t arg1);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::srli(bitblock256_t arg1);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::srli(bitblock256_t arg1);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::srli(bitblock256_t arg1);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::srli(bitblock256_t arg1);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::srli(bitblock256_t arg1);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::srli(bitblock256_t arg1);
 template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<1>::ctz(bitblock256_t arg1);
 template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::ctz(bitblock256_t arg1);
@@ -208 +210 @@
 template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::popcount(bitblock256_t arg1);
 template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::popcount(bitblock256_t arg1);
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::any(bitblock256_t arg1);
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::any(bitblock256_t arg1);
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::any(bitblock256_t arg1);
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::any(bitblock256_t arg1);
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::any(bitblock256_t arg1);
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::any(bitblock256_t arg1);
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::any(bitblock256_t arg1);
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::any(bitblock256_t arg1);
 template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::neg(bitblock256_t arg1);
 template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::neg(bitblock256_t arg1);
@@ -216 +226 @@
 template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::neg(bitblock256_t arg1);
 template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::neg(bitblock256_t arg1);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::slli(bitblock256_t arg1);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::slli(bitblock256_t arg1);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::slli(bitblock256_t arg1);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::slli(bitblock256_t arg1);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::slli(bitblock256_t arg1);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::slli(bitblock256_t arg1);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::slli(bitblock256_t arg1);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::slli(bitblock256_t arg1);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::slli(bitblock256_t arg1);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::slli(bitblock256_t arg1);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::slli(bitblock256_t arg1);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::slli(bitblock256_t arg1);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::slli(bitblock256_t arg1);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::slli(bitblock256_t arg1);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::slli(bitblock256_t arg1);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::slli(bitblock256_t arg1);
 template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<1>::ifh(bitblock256_t arg1, bitblock256_t arg2, bitblock256_t arg3);
 template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::ifh(bitblock256_t arg1, bitblock256_t arg2, bitblock256_t arg3);
@@ -250 +260 @@
 template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::add_hl(bitblock256_t arg1);
 template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::add_hl(bitblock256_t arg1);
-template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock256_t simd256<1>::constant();
-template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::constant();
-template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::constant();
-template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::constant();
-template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::constant();
-template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::constant();
-template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::constant();
-template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::constant();
-template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::constant();
+template <> template <FieldType<1>::T val> IDISA_ALWAYS_INLINE bitblock256_t simd256<1>::constant();
+template <> template <FieldType<2>::T val> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::constant();
+template <> template <FieldType<4>::T val> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::constant();
+template <> template <FieldType<8>::T val> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::constant();
+template <> template <FieldType<16>::T val> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::constant();
+template <> template <FieldType<32>::T val> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::constant();
+template <> template <FieldType<64>::T val> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::constant();
+template <> template <FieldType<128>::T val> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::constant();
+template <> template <FieldType<256>::T val> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::constant();
 template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<1>::min(bitblock256_t arg1, bitblock256_t arg2);
 template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::min(bitblock256_t arg1, bitblock256_t arg2);
@@ -276 +286 @@
 template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::lomask();
 template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::lomask();
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<1>::add(bitblock256_t arg1, bitblock256_t arg2);
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::add(bitblock256_t arg1, bitblock256_t arg2);
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::add(bitblock256_t arg1, bitblock256_t arg2);
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::add(bitblock256_t arg1, bitblock256_t arg2);
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::add(bitblock256_t arg1, bitblock256_t arg2);
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::add(bitblock256_t arg1, bitblock256_t arg2);
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::add(bitblock256_t arg1, bitblock256_t arg2);
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::add(bitblock256_t arg1, bitblock256_t arg2);
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::add(bitblock256_t arg1, bitblock256_t arg2);
 template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<1>::umin(bitblock256_t arg1, bitblock256_t arg2);
 template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::umin(bitblock256_t arg1, bitblock256_t arg2);
@@ -302 +321 @@
 template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::eq(bitblock256_t arg1, bitblock256_t arg2);
 template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::eq(bitblock256_t arg1, bitblock256_t arg2);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::srai(bitblock256_t arg1);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::srai(bitblock256_t arg1);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::srai(bitblock256_t arg1);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::srai(bitblock256_t arg1);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::srai(bitblock256_t arg1);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::srai(bitblock256_t arg1);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::srai(bitblock256_t arg1);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::srai(bitblock256_t arg1);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::srai(bitblock256_t arg1);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::srai(bitblock256_t arg1);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::srai(bitblock256_t arg1);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::srai(bitblock256_t arg1);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::srai(bitblock256_t arg1);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::srai(bitblock256_t arg1);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::srai(bitblock256_t arg1);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::srai(bitblock256_t arg1);
 template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::himask();
 template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::himask();
@@ -318 +337 @@
 template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::himask();
 template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::himask();
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<1>::add(bitblock256_t arg1, bitblock256_t arg2);
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::add(bitblock256_t arg1, bitblock256_t arg2);
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::add(bitblock256_t arg1, bitblock256_t arg2);
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::add(bitblock256_t arg1, bitblock256_t arg2);
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::add(bitblock256_t arg1, bitblock256_t arg2);
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::add(bitblock256_t arg1, bitblock256_t arg2);
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::add(bitblock256_t arg1, bitblock256_t arg2);
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::add(bitblock256_t arg1, bitblock256_t arg2);
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::add(bitblock256_t arg1, bitblock256_t arg2);
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<1>::lt(bitblock256_t arg1, bitblock256_t arg2);
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::lt(bitblock256_t arg1, bitblock256_t arg2);
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::lt(bitblock256_t arg1, bitblock256_t arg2);
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::lt(bitblock256_t arg1, bitblock256_t arg2);
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::lt(bitblock256_t arg1, bitblock256_t arg2);
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::lt(bitblock256_t arg1, bitblock256_t arg2);
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::lt(bitblock256_t arg1, bitblock256_t arg2);
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::lt(bitblock256_t arg1, bitblock256_t arg2);
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::lt(bitblock256_t arg1, bitblock256_t arg2);
 template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<1>::umax(bitblock256_t arg1, bitblock256_t arg2);
 template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::umax(bitblock256_t arg1, bitblock256_t arg2);
@@ -360 +379 @@
 template <> IDISA_ALWAYS_INLINE bitblock256_t hsimd256<128>::packss(bitblock256_t arg1, bitblock256_t arg2);
 template <> IDISA_ALWAYS_INLINE bitblock256_t hsimd256<256>::packss(bitblock256_t arg1, bitblock256_t arg2);
-template <> IDISA_ALWAYS_INLINE uint64_t hsimd256<8>::signmask(bitblock256_t arg1);
-template <> IDISA_ALWAYS_INLINE uint64_t hsimd256<16>::signmask(bitblock256_t arg1);
-template <> IDISA_ALWAYS_INLINE uint64_t hsimd256<32>::signmask(bitblock256_t arg1);
-template <> IDISA_ALWAYS_INLINE uint64_t hsimd256<64>::signmask(bitblock256_t arg1);
-template <> IDISA_ALWAYS_INLINE uint64_t hsimd256<128>::signmask(bitblock256_t arg1);
-template <> IDISA_ALWAYS_INLINE uint64_t hsimd256<256>::signmask(bitblock256_t arg1);
+template <> IDISA_ALWAYS_INLINE FieldType<256/8>::T hsimd256<8>::signmask(bitblock256_t arg1);
+template <> IDISA_ALWAYS_INLINE FieldType<256/16>::T hsimd256<16>::signmask(bitblock256_t arg1);
+template <> IDISA_ALWAYS_INLINE FieldType<256/32>::T hsimd256<32>::signmask(bitblock256_t arg1);
+template <> IDISA_ALWAYS_INLINE FieldType<256/64>::T hsimd256<64>::signmask(bitblock256_t arg1);
+template <> IDISA_ALWAYS_INLINE FieldType<256/128>::T hsimd256<128>::signmask(bitblock256_t arg1);
+template <> IDISA_ALWAYS_INLINE FieldType<256/256>::T hsimd256<256>::signmask(bitblock256_t arg1);
 template <> IDISA_ALWAYS_INLINE bitblock256_t hsimd256<2>::packl(bitblock256_t arg1, bitblock256_t arg2);
 template <> IDISA_ALWAYS_INLINE bitblock256_t hsimd256<4>::packl(bitblock256_t arg1, bitblock256_t arg2);
@@ -446 +465 @@
 template <> IDISA_ALWAYS_INLINE bitblock256_t esimd256<64>::signextendl(bitblock256_t arg1);
 template <> IDISA_ALWAYS_INLINE bitblock256_t esimd256<128>::signextendl(bitblock256_t arg1);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<2>::dsrli(bitblock256_t arg1, bitblock256_t arg2);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<4>::dsrli(bitblock256_t arg1, bitblock256_t arg2);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<8>::dsrli(bitblock256_t arg1, bitblock256_t arg2);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<16>::dsrli(bitblock256_t arg1, bitblock256_t arg2);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<32>::dsrli(bitblock256_t arg1, bitblock256_t arg2);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<64>::dsrli(bitblock256_t arg1, bitblock256_t arg2);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<128>::dsrli(bitblock256_t arg1, bitblock256_t arg2);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<256>::dsrli(bitblock256_t arg1, bitblock256_t arg2);
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<1>::fill(uint64_t val1);
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<2>::fill(uint64_t val1);
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<4>::fill(uint64_t val1);
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<8>::fill(uint64_t val1);
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<16>::fill(uint64_t val1);
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<32>::fill(uint64_t val1);
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<64>::fill(uint64_t val1);
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<128>::fill(uint64_t val1);
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<256>::fill(uint64_t val1);
-template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd256<1>::extract(bitblock256_t arg1);
-template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd256<2>::extract(bitblock256_t arg1);
-template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd256<4>::extract(bitblock256_t arg1);
-template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd256<8>::extract(bitblock256_t arg1);
-template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd256<16>::extract(bitblock256_t arg1);
-template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd256<32>::extract(bitblock256_t arg1);
-template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd256<64>::extract(bitblock256_t arg1);
-template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<1>::splat(bitblock256_t arg1);
-template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<2>::splat(bitblock256_t arg1);
-template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<4>::splat(bitblock256_t arg1);
-template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<8>::splat(bitblock256_t arg1);
-template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<16>::splat(bitblock256_t arg1);
-template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<32>::splat(bitblock256_t arg1);
-template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<64>::splat(bitblock256_t arg1);
-template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<128>::splat(bitblock256_t arg1);
-template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<256>::splat(bitblock256_t arg1);
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<1>::fill16(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8, uint64_t val9, uint64_t val10, uint64_t val11, uint64_t val12, uint64_t val13, uint64_t val14, uint64_t val15, uint64_t val16);
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<2>::fill16(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8, uint64_t val9, uint64_t val10, uint64_t val11, uint64_t val12, uint64_t val13, uint64_t val14, uint64_t val15, uint64_t val16);
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<4>::fill16(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8, uint64_t val9, uint64_t val10, uint64_t val11, uint64_t val12, uint64_t val13, uint64_t val14, uint64_t val15, uint64_t val16);
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<8>::fill16(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8, uint64_t val9, uint64_t val10, uint64_t val11, uint64_t val12, uint64_t val13, uint64_t val14, uint64_t val15, uint64_t val16);
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<16>::fill16(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8, uint64_t val9, uint64_t val10, uint64_t val11, uint64_t val12, uint64_t val13, uint64_t val14, uint64_t val15, uint64_t val16);
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<1>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4);
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<2>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4);
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<4>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4);
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<8>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4);
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<16>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4);
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<32>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4);
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<64>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<2>::srli(bitblock256_t arg1);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<4>::srli(bitblock256_t arg1);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<8>::srli(bitblock256_t arg1);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<16>::srli(bitblock256_t arg1);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<32>::srli(bitblock256_t arg1);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<64>::srli(bitblock256_t arg1);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<128>::srli(bitblock256_t arg1);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<256>::srli(bitblock256_t arg1);
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<1>::fill2(uint64_t val1, uint64_t val2);
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<2>::fill2(uint64_t val1, uint64_t val2);
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<4>::fill2(uint64_t val1, uint64_t val2);
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<8>::fill2(uint64_t val1, uint64_t val2);
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<16>::fill2(uint64_t val1, uint64_t val2);
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<32>::fill2(uint64_t val1, uint64_t val2);
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<64>::fill2(uint64_t val1, uint64_t val2);
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<128>::fill2(uint64_t val1, uint64_t val2);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<2>::dslli(bitblock256_t arg1, bitblock256_t arg2);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<4>::dslli(bitblock256_t arg1, bitblock256_t arg2);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<8>::dslli(bitblock256_t arg1, bitblock256_t arg2);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<16>::dslli(bitblock256_t arg1, bitblock256_t arg2);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<32>::dslli(bitblock256_t arg1, bitblock256_t arg2);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<64>::dslli(bitblock256_t arg1, bitblock256_t arg2);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<128>::dslli(bitblock256_t arg1, bitblock256_t arg2);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<256>::dslli(bitblock256_t arg1, bitblock256_t arg2);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<2>::slli(bitblock256_t arg1);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<4>::slli(bitblock256_t arg1);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<8>::slli(bitblock256_t arg1);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<16>::slli(bitblock256_t arg1);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<32>::slli(bitblock256_t arg1);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<64>::slli(bitblock256_t arg1);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<128>::slli(bitblock256_t arg1);
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<256>::slli(bitblock256_t arg1);
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<1>::fill8(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8);
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<2>::fill8(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8);
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<4>::fill8(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8);
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<8>::fill8(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8);
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<16>::fill8(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8);
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<32>::fill8(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<2>::dsrli(bitblock256_t arg1, bitblock256_t arg2);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<4>::dsrli(bitblock256_t arg1, bitblock256_t arg2);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<8>::dsrli(bitblock256_t arg1, bitblock256_t arg2);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<16>::dsrli(bitblock256_t arg1, bitblock256_t arg2);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<32>::dsrli(bitblock256_t arg1, bitblock256_t arg2);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<64>::dsrli(bitblock256_t arg1, bitblock256_t arg2);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<128>::dsrli(bitblock256_t arg1, bitblock256_t arg2);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<256>::dsrli(bitblock256_t arg1, bitblock256_t arg2);
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<1>::fill(FieldType<1>::T val1);
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<2>::fill(FieldType<2>::T val1);
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<4>::fill(FieldType<4>::T val1);
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<8>::fill(FieldType<8>::T val1);
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<16>::fill(FieldType<16>::T val1);
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<32>::fill(FieldType<32>::T val1);
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<64>::fill(FieldType<64>::T val1);
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<128>::fill(FieldType<128>::T val1);
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<256>::fill(FieldType<256>::T val1);
+template <> template <uint8_t pos> IDISA_ALWAYS_INLINE FieldType<1>::T mvmd256<1>::extract(bitblock256_t arg1);
+template <> template <uint8_t pos> IDISA_ALWAYS_INLINE FieldType<2>::T mvmd256<2>::extract(bitblock256_t arg1);
+template <> template <uint8_t pos> IDISA_ALWAYS_INLINE FieldType<4>::T mvmd256<4>::extract(bitblock256_t arg1);
+template <> template <uint8_t pos> IDISA_ALWAYS_INLINE FieldType<8>::T mvmd256<8>::extract(bitblock256_t arg1);
+template <> template <uint8_t pos> IDISA_ALWAYS_INLINE FieldType<16>::T mvmd256<16>::extract(bitblock256_t arg1);
+template <> template <uint8_t pos> IDISA_ALWAYS_INLINE FieldType<32>::T mvmd256<32>::extract(bitblock256_t arg1);
+template <> template <uint8_t pos> IDISA_ALWAYS_INLINE FieldType<64>::T mvmd256<64>::extract(bitblock256_t arg1);
+template <> template <uint16_t pos> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<1>::splat(bitblock256_t arg1);
+template <> template <uint16_t pos> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<2>::splat(bitblock256_t arg1);
+template <> template <uint16_t pos> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<4>::splat(bitblock256_t arg1);
+template <> template <uint16_t pos> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<8>::splat(bitblock256_t arg1);
+template <> template <uint16_t pos> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<16>::splat(bitblock256_t arg1);
+template <> template <uint16_t pos> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<32>::splat(bitblock256_t arg1);
+template <> template <uint16_t pos> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<64>::splat(bitblock256_t arg1);
+template <> template <uint16_t pos> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<128>::splat(bitblock256_t arg1);
+template <> template <uint16_t pos> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<256>::splat(bitblock256_t arg1);
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<1>::fill16(FieldType<1>::T val1, FieldType<1>::T val2, FieldType<1>::T val3, FieldType<1>::T val4, FieldType<1>::T val5, FieldType<1>::T val6, FieldType<1>::T val7, FieldType<1>::T val8, FieldType<1>::T val9, FieldType<1>::T val10, FieldType<1>::T val11, FieldType<1>::T val12, FieldType<1>::T val13, FieldType<1>::T val14, FieldType<1>::T val15, FieldType<1>::T val16);
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<2>::fill16(FieldType<2>::T val1, FieldType<2>::T val2, FieldType<2>::T val3, FieldType<2>::T val4, FieldType<2>::T val5, FieldType<2>::T val6, FieldType<2>::T val7, FieldType<2>::T val8, FieldType<2>::T val9, FieldType<2>::T val10, FieldType<2>::T val11, FieldType<2>::T val12, FieldType<2>::T val13, FieldType<2>::T val14, FieldType<2>::T val15, FieldType<2>::T val16);
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<4>::fill16(FieldType<4>::T val1, FieldType<4>::T val2, FieldType<4>::T val3, FieldType<4>::T val4, FieldType<4>::T val5, FieldType<4>::T val6, FieldType<4>::T val7, FieldType<4>::T val8, FieldType<4>::T val9, FieldType<4>::T val10, FieldType<4>::T val11, FieldType<4>::T val12, FieldType<4>::T val13, FieldType<4>::T val14, FieldType<4>::T val15, FieldType<4>::T val16);
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<8>::fill16(FieldType<8>::T val1, FieldType<8>::T val2, FieldType<8>::T val3, FieldType<8>::T val4, FieldType<8>::T val5, FieldType<8>::T val6, FieldType<8>::T val7, FieldType<8>::T val8, FieldType<8>::T val9, FieldType<8>::T val10, FieldType<8>::T val11, FieldType<8>::T val12, FieldType<8>::T val13, FieldType<8>::T val14, FieldType<8>::T val15, FieldType<8>::T val16);
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<16>::fill16(FieldType<16>::T val1, FieldType<16>::T val2, FieldType<16>::T val3, FieldType<16>::T val4, FieldType<16>::T val5, FieldType<16>::T val6, FieldType<16>::T val7, FieldType<16>::T val8, FieldType<16>::T val9, FieldType<16>::T val10, FieldType<16>::T val11, FieldType<16>::T val12, FieldType<16>::T val13, FieldType<16>::T val14, FieldType<16>::T val15, FieldType<16>::T val16);
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<1>::fill4(FieldType<1>::T val1, FieldType<1>::T val2, FieldType<1>::T val3, FieldType<1>::T val4);
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<2>::fill4(FieldType<2>::T val1, FieldType<2>::T val2, FieldType<2>::T val3, FieldType<2>::T val4);
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<4>::fill4(FieldType<4>::T val1, FieldType<4>::T val2, FieldType<4>::T val3, FieldType<4>::T val4);
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<8>::fill4(FieldType<8>::T val1, FieldType<8>::T val2, FieldType<8>::T val3, FieldType<8>::T val4);
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<16>::fill4(FieldType<16>::T val1, FieldType<16>::T val2, FieldType<16>::T val3, FieldType<16>::T val4);
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<32>::fill4(FieldType<32>::T val1, FieldType<32>::T val2, FieldType<32>::T val3, FieldType<32>::T val4);
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<64>::fill4(FieldType<64>::T val1, FieldType<64>::T val2, FieldType<64>::T val3, FieldType<64>::T val4);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<2>::srli(bitblock256_t arg1);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<4>::srli(bitblock256_t arg1);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<8>::srli(bitblock256_t arg1);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<16>::srli(bitblock256_t arg1);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<32>::srli(bitblock256_t arg1);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<64>::srli(bitblock256_t arg1);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<128>::srli(bitblock256_t arg1);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<256>::srli(bitblock256_t arg1);
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<1>::fill2(FieldType<1>::T val1, FieldType<1>::T val2);
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<2>::fill2(FieldType<2>::T val1, FieldType<2>::T val2);
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<4>::fill2(FieldType<4>::T val1, FieldType<4>::T val2);
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<8>::fill2(FieldType<8>::T val1, FieldType<8>::T val2);
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<16>::fill2(FieldType<16>::T val1, FieldType<16>::T val2);
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<32>::fill2(FieldType<32>::T val1, FieldType<32>::T val2);
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<64>::fill2(FieldType<64>::T val1, FieldType<64>::T val2);
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<128>::fill2(FieldType<128>::T val1, FieldType<128>::T val2);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<2>::dslli(bitblock256_t arg1, bitblock256_t arg2);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<4>::dslli(bitblock256_t arg1, bitblock256_t arg2);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<8>::dslli(bitblock256_t arg1, bitblock256_t arg2);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<16>::dslli(bitblock256_t arg1, bitblock256_t arg2);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<32>::dslli(bitblock256_t arg1, bitblock256_t arg2);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<64>::dslli(bitblock256_t arg1, bitblock256_t arg2);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<128>::dslli(bitblock256_t arg1, bitblock256_t arg2);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<256>::dslli(bitblock256_t arg1, bitblock256_t arg2);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<2>::slli(bitblock256_t arg1);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<4>::slli(bitblock256_t arg1);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<8>::slli(bitblock256_t arg1);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<16>::slli(bitblock256_t arg1);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<32>::slli(bitblock256_t arg1);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<64>::slli(bitblock256_t arg1);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<128>::slli(bitblock256_t arg1);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<256>::slli(bitblock256_t arg1);
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<1>::fill8(FieldType<1>::T val1, FieldType<1>::T val2, FieldType<1>::T val3, FieldType<1>::T val4, FieldType<1>::T val5, FieldType<1>::T val6, FieldType<1>::T val7, FieldType<1>::T val8);
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<2>::fill8(FieldType<2>::T val1, FieldType<2>::T val2, FieldType<2>::T val3, FieldType<2>::T val4, FieldType<2>::T val5, FieldType<2>::T val6, FieldType<2>::T val7, FieldType<2>::T val8);
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<4>::fill8(FieldType<4>::T val1, FieldType<4>::T val2, FieldType<4>::T val3, FieldType<4>::T val4, FieldType<4>::T val5, FieldType<4>::T val6, FieldType<4>::T val7, FieldType<4>::T val8);
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<8>::fill8(FieldType<8>::T val1, FieldType<8>::T val2, FieldType<8>::T val3, FieldType<8>::T val4, FieldType<8>::T val5, FieldType<8>::T val6, FieldType<8>::T val7, FieldType<8>::T val8);
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<16>::fill8(FieldType<16>::T val1, FieldType<16>::T val2, FieldType<16>::T val3, FieldType<16>::T val4, FieldType<16>::T val5, FieldType<16>::T val6, FieldType<16>::T val7, FieldType<16>::T val8);
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<32>::fill8(FieldType<32>::T val1, FieldType<32>::T val2, FieldType<32>::T val3, FieldType<32>::T val4, FieldType<32>::T val5, FieldType<32>::T val6, FieldType<32>::T val7, FieldType<32>::T val8);
 
 //Implementation Part
@@ -577 +596 @@
 
 //The total number of operations is 1.0
+IDISA_ALWAYS_INLINE bitblock256_t simd_and(bitblock256_t arg1, bitblock256_t arg2)
+{
+        return _mm256_and_ps(arg1, arg2);
+}
+
+//The total number of operations is 1.0
 IDISA_ALWAYS_INLINE bitblock256_t simd_xor(bitblock256_t arg1, bitblock256_t arg2)
 {
         return _mm256_xor_ps(arg1, arg2);
-}
-
-//The total number of operations is 1.0
-IDISA_ALWAYS_INLINE bitblock256_t simd_and(bitblock256_t arg1, bitblock256_t arg2)
-{
-        return _mm256_and_ps(arg1, arg2);
 }
 
@@ -962 +981 @@
 }
 
+//The total number of operations is 12.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::all(bitblock256_t arg1)
+{
+        bitblock256_t f0 = simd_and(arg1, simd256<2>::srli<1>(arg1));
+        return simd_or(f0, simd256<2>::slli<1>(f0));
+}
+
+//The total number of operations is 17.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::all(bitblock256_t arg1)
+{
+        return simd256<4>::eq(arg1, simd256<8>::constant<255>());
+}
+
+//The total number of operations is 5.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::all(bitblock256_t arg1)
+{
+        return simd256<8>::eq(arg1, simd256<8>::constant<255>());
+}
+
+//The total number of operations is 5.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::all(bitblock256_t arg1)
+{
+        return simd256<16>::eq(arg1, simd256<8>::constant<255>());
+}
+
+//The total number of operations is 5.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::all(bitblock256_t arg1)
+{
+        return simd256<32>::eq(arg1, simd256<8>::constant<255>());
+}
+
+//The total number of operations is 5.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::all(bitblock256_t arg1)
+{
+        return simd256<64>::eq(arg1, simd256<8>::constant<255>());
+}
+
+//The total number of operations is 23.6666666667
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::all(bitblock256_t arg1)
+{
+        return simd256<128>::eq(arg1, simd256<8>::constant<255>());
+}
+
+//The total number of operations is 2.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::all(bitblock256_t arg1)
+{
+        return ((bitblock256::all(arg1)) ? simd256<8>::constant<255>() : simd256<8>::constant<0>());
+}
+
+//The total number of operations is 5.0
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::srli(bitblock256_t arg1)
+{
+        return simd_and(simd256<32>::srli<sh>(arg1), simd256<2>::constant<((3)>>sh)>());
+}
+
+//The total number of operations is 5.0
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::srli(bitblock256_t arg1)
+{
+        return simd_and(simd256<32>::srli<sh>(arg1), simd256<4>::constant<((15)>>sh)>());
+}
+
+//The total number of operations is 5.0
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::srli(bitblock256_t arg1)
+{
+        return simd_and(simd256<32>::srli<sh>(arg1), simd256<8>::constant<((255)>>sh)>());
+}
+
+//The total number of operations is 4.0
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::srli(bitblock256_t arg1)
+{
+        return avx_general_combine256(_mm_srli_epi16(avx_select_hi128(arg1), (int32_t)(sh)), _mm_srli_epi16(avx_select_lo128(arg1), (int32_t)(sh)));
+}
+
+//The total number of operations is 4.0
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::srli(bitblock256_t arg1)
+{
+        return avx_general_combine256(_mm_srli_epi32(avx_select_hi128(arg1), (int32_t)(sh)), _mm_srli_epi32(avx_select_lo128(arg1), (int32_t)(sh)));
+}
+
+//The total number of operations is 4.0
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::srli(bitblock256_t arg1)
+{
+        return avx_general_combine256(_mm_srli_epi64(avx_select_hi128(arg1), (int32_t)(sh)), _mm_srli_epi64(avx_select_lo128(arg1), (int32_t)(sh)));
+}
+
+//The total number of operations is 8.33333333333
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::srli(bitblock256_t arg1)
+{
+        return (((sh%8) == 0) ? avx_byte_shift_right(arg1, (sh/8)) : ((sh >= 64) ? simd256<64>::srli<(sh&63)>(avx_byte_shift_right(arg1, 8)) : simd_or(simd256<64>::srli<sh>(arg1), avx_byte_shift_right(simd256<64>::slli<((128-sh)&63)>(arg1), 8))));
+}
+
+//The total number of operations is 14.5
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::srli(bitblock256_t arg1)
+{
+        return ((sh < 128) ? simd_or(simd256<128>::srli<sh>(arg1), simd256<128>::slli<(128-sh)>(((bitblock256_t)(_mm256_castsi128_si256(avx_select_hi128(arg1)))))) : simd256<128>::srli<(sh-128)>(avx_move_hi128_to_lo128(arg1)));
+}
+
+//The total number of operations is 1.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<1>::ctz(bitblock256_t arg1)
+{
+        return simd_not(arg1);
+}
+
+//The total number of operations is 27.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::ctz(bitblock256_t arg1)
+{
+        return simd256<2>::popcount(simd_andc(simd256<2>::sub(arg1, simd256<2>::constant<1>()), arg1));
+}
+
+//The total number of operations is 36.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::ctz(bitblock256_t arg1)
+{
+        return simd256<4>::popcount(simd_andc(simd256<4>::sub(arg1, simd256<4>::constant<1>()), arg1));
+}
+
+//The total number of operations is 38.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::ctz(bitblock256_t arg1)
+{
+        return simd256<8>::popcount(simd_andc(simd256<8>::sub(arg1, simd256<8>::constant<1>()), arg1));
+}
+
+//The total number of operations is 48.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::ctz(bitblock256_t arg1)
+{
+        return simd256<16>::popcount(simd_andc(simd256<16>::sub(arg1, simd256<16>::constant<1>()), arg1));
+}
+
+//The total number of operations is 58.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::ctz(bitblock256_t arg1)
+{
+        return simd256<32>::popcount(simd_andc(simd256<32>::sub(arg1, simd256<32>::constant<1>()), arg1));
+}
+
+//The total number of operations is 44.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::ctz(bitblock256_t arg1)
+{
+        return simd256<64>::popcount(simd_andc(simd256<64>::sub(arg1, simd256<64>::constant<1>()), arg1));
+}
+
+//The total number of operations is 101.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::ctz(bitblock256_t arg1)
+{
+        return simd256<128>::popcount(simd_andc(simd256<128>::sub(arg1, simd256<128>::constant<1>()), arg1));
+}
+
+//The total number of operations is 192.166666667
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::ctz(bitblock256_t arg1)
+{
+        return simd256<256>::popcount(simd_andc(simd256<256>::sub(arg1, simd256<256>::constant<1>()), arg1));
+}
+
+//The total number of operations is 1.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<1>::ugt(bitblock256_t arg1, bitblock256_t arg2)
+{
+        return simd_andc(arg1, arg2);
+}
+
+//The total number of operations is 23.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::ugt(bitblock256_t arg1, bitblock256_t arg2)
+{
+        bitblock256_t tmpAns = simd256<(1)>::ugt(arg1, arg2);
+        bitblock256_t mask = simd_and(tmpAns, simd256<2>::srli<(1)>(simd256<(1)>::eq(arg1, arg2)));
+        mask = simd_or(mask, simd256<2>::slli<(1)>(mask));
+        return simd_or(simd256<2>::srai<(1)>(tmpAns), mask);
+}
+
+//The total number of operations is 20.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::ugt(bitblock256_t arg1, bitblock256_t arg2)
+{
+        return simd256<1>::ifh(simd256<(8)>::himask(), simd256<(8)>::ugt(simd_and(simd256<(8)>::himask(), arg1), arg2), simd256<(8)>::ugt(simd_andc(arg1, simd256<(8)>::himask()), simd_andc(arg2, simd256<(8)>::himask())));
+}
+
+//The total number of operations is 7.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::ugt(bitblock256_t arg1, bitblock256_t arg2)
+{
+        bitblock256_t high_bit = simd256<8>::constant<(128)>();
+        return simd256<8>::gt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
+}
+
+//The total number of operations is 7.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::ugt(bitblock256_t arg1, bitblock256_t arg2)
+{
+        bitblock256_t high_bit = simd256<16>::constant<(32768)>();
+        return simd256<16>::gt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
+}
+
+//The total number of operations is 7.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::ugt(bitblock256_t arg1, bitblock256_t arg2)
+{
+        bitblock256_t high_bit = simd256<32>::constant<(2147483648ULL)>();
+        return simd256<32>::gt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
+}
+
+//The total number of operations is 7.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::ugt(bitblock256_t arg1, bitblock256_t arg2)
+{
+        bitblock256_t high_bit = simd256<64>::constant<(9223372036854775808ULL)>();
+        return simd256<64>::gt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
+}
+
+//The total number of operations is 60.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::ugt(bitblock256_t arg1, bitblock256_t arg2)
+{
+        bitblock256_t tmpAns = simd256<(64)>::ugt(arg1, arg2);
+        bitblock256_t mask = simd_and(tmpAns, simd256<128>::srli<(64)>(simd256<(64)>::eq(arg1, arg2)));
+        mask = simd_or(mask, simd256<128>::slli<(64)>(mask));
+        return simd_or(simd256<128>::srai<(64)>(tmpAns), mask);
+}
+
+//The total number of operations is 174.166666667
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::ugt(bitblock256_t arg1, bitblock256_t arg2)
+{
+        bitblock256_t tmpAns = simd256<(128)>::ugt(arg1, arg2);
+        bitblock256_t mask = simd_and(tmpAns, simd256<256>::srli<(128)>(simd256<(128)>::eq(arg1, arg2)));
+        mask = simd_or(mask, simd256<256>::slli<(128)>(mask));
+        return simd_or(simd256<256>::srai<(128)>(tmpAns), mask);
+}
+
+//The total number of operations is 7.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::xor_hl(bitblock256_t arg1)
+{
+        return simd_xor(simd256<2>::srli<(1)>(arg1), simd_and(arg1, simd256<2>::lomask()));
+}
+
+//The total number of operations is 7.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::xor_hl(bitblock256_t arg1)
+{
+        return simd_xor(simd256<4>::srli<(2)>(arg1), simd_and(arg1, simd256<4>::lomask()));
+}
+
+//The total number of operations is 7.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::xor_hl(bitblock256_t arg1)
+{
+        return simd_xor(simd256<8>::srli<(4)>(arg1), simd_and(arg1, simd256<8>::lomask()));
+}
+
+//The total number of operations is 6.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::xor_hl(bitblock256_t arg1)
+{
+        return simd_xor(simd256<16>::srli<(8)>(arg1), simd_and(arg1, simd256<16>::lomask()));
+}
+
+//The total number of operations is 6.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::xor_hl(bitblock256_t arg1)
+{
+        return simd_xor(simd256<32>::srli<(16)>(arg1), simd_and(arg1, simd256<32>::lomask()));
+}
+
+//The total number of operations is 6.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::xor_hl(bitblock256_t arg1)
+{
+        return simd_xor(simd256<64>::srli<(32)>(arg1), simd_and(arg1, simd256<64>::lomask()));
+}
+
+//The total number of operations is 10.3333333333
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::xor_hl(bitblock256_t arg1)
+{
+        return simd_xor(simd256<128>::srli<(64)>(arg1), simd_and(arg1, simd256<128>::lomask()));
+}
+
+//The total number of operations is 16.5
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::xor_hl(bitblock256_t arg1)
+{
+        return simd_xor(simd256<256>::srli<(128)>(arg1), simd_and(arg1, simd256<256>::lomask()));
+}
+
+//The total number of operations is 0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<1>::popcount(bitblock256_t arg1)
+{
+        return arg1;
+}
+
+//The total number of operations is 10.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::popcount(bitblock256_t arg1)
+{
+        return simd256<2>::add_hl(simd256<(1)>::popcount(arg1));
+}
+
+//The total number of operations is 21.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::popcount(bitblock256_t arg1)
+{
+        return simd256<4>::add_hl(simd256<(2)>::popcount(arg1));
+}
+
+//The total number of operations is 32.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::popcount(bitblock256_t arg1)
+{
+        return simd256<8>::add_hl(simd256<(4)>::popcount(arg1));
+}
+
+//The total number of operations is 42.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::popcount(bitblock256_t arg1)
+{
+        return simd256<16>::add_hl(simd256<(8)>::popcount(arg1));
+}
+
+//The total number of operations is 52.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::popcount(bitblock256_t arg1)
+{
+        return simd256<32>::add_hl(simd256<(16)>::popcount(arg1));
+}
+
+//The total number of operations is 38.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::popcount(bitblock256_t arg1)
+{
+        bitblock256_t tmpAns = simd256<8>::popcount(arg1);
+        return avx_general_combine256(_mm_sad_epu8(avx_select_hi128(tmpAns), _mm_set1_epi32((int32_t)(0))), _mm_sad_epu8(avx_select_lo128(tmpAns), _mm_set1_epi32((int32_t)(0))));
+}
+
+//The total number of operations is 73.6666666667
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::popcount(bitblock256_t arg1)
+{
+        return simd256<128>::add_hl(simd256<(64)>::popcount(arg1));
+}
+
+//The total number of operations is 115.5
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::popcount(bitblock256_t arg1)
+{
+        bitblock256_t tmpAns = simd256<(128)>::popcount(arg1);
+        return simd256<(128)>::add(simd_and(tmpAns, simd256<256>::lomask()), simd256<256>::srli<(128)>(tmpAns));
+}
+
+//The total number of operations is 14.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::any(bitblock256_t arg1)
+{
+        bitblock256_t t0 = simd256<2>::srli<1>(arg1);
+        bitblock256_t f0 = simd_or(t0, simd_and(arg1, simd_xor(t0, simd256<8>::constant<255>())));
+        return simd_or(f0, simd256<2>::slli<1>(f0));
+}
+
+//The total number of operations is 20.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::any(bitblock256_t arg1)
+{
+        return simd256<4>::ugt(arg1, simd256<8>::constant<0>());
+}
+
+//The total number of operations is 7.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::any(bitblock256_t arg1)
+{
+        return simd256<8>::ugt(arg1, simd256<8>::constant<0>());
+}
+
+//The total number of operations is 7.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::any(bitblock256_t arg1)
+{
+        return simd256<16>::ugt(arg1, simd256<8>::constant<0>());
+}
+
+//The total number of operations is 7.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::any(bitblock256_t arg1)
+{
+        return simd256<32>::ugt(arg1, simd256<8>::constant<0>());
+}
+
+//The total number of operations is 7.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::any(bitblock256_t arg1)
+{
+        return simd256<64>::ugt(arg1, simd256<8>::constant<0>());
+}
+
+//The total number of operations is 60.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::any(bitblock256_t arg1)
+{
+        return simd256<128>::ugt(arg1, simd256<8>::constant<0>());
+}
+
+//The total number of operations is 1.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::any(bitblock256_t arg1)
+{
+        return ((bitblock256::any(arg1)) ? simd256<8>::constant<255>() : simd256<8>::constant<0>());
+}
+
+//The total number of operations is 16.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::neg(bitblock256_t arg1)
+{
+        return simd256<2>::sub(simd256<2>::constant<0>(), arg1);
+}
+
+//The total number of operations is 14.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::neg(bitblock256_t arg1)
+{
+        return simd256<4>::sub(simd256<4>::constant<0>(), arg1);
+}
+
+//The total number of operations is 5.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::neg(bitblock256_t arg1)
+{
+        return simd256<8>::sub(simd256<8>::constant<0>(), arg1);
+}
+
+//The total number of operations is 5.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::neg(bitblock256_t arg1)
+{
+        return simd256<16>::sub(simd256<16>::constant<0>(), arg1);
+}
+
+//The total number of operations is 5.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::neg(bitblock256_t arg1)
+{
+        return simd256<32>::sub(simd256<32>::constant<0>(), arg1);
+}
+
+//The total number of operations is 5.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::neg(bitblock256_t arg1)
+{
+        return simd256<64>::sub(simd256<64>::constant<0>(), arg1);
+}
+
+//The total number of operations is 26.3333333333
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::neg(bitblock256_t arg1)
+{
+        return simd256<128>::sub(simd256<128>::constant<0>(), arg1);
+}
+
+//The total number of operations is 75.6666666667
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::neg(bitblock256_t arg1)
+{
+        return simd256<256>::sub(simd256<256>::constant<0>(), arg1);
+}
+
+//The total number of operations is 5.0
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::slli(bitblock256_t arg1)
+{
+        return simd_and(simd256<32>::slli<sh>(arg1), simd256<2>::constant<(((3)<<sh)&(3))>());
+}
+
+//The total number of operations is 5.0
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::slli(bitblock256_t arg1)
+{
+        return simd_and(simd256<32>::slli<sh>(arg1), simd256<4>::constant<(((15)<<sh)&(15))>());
+}
+
+//The total number of operations is 5.0
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::slli(bitblock256_t arg1)
+{
+        return simd_and(simd256<32>::slli<sh>(arg1), simd256<8>::constant<(((255)<<sh)&(255))>());
+}
+
+//The total number of operations is 4.0
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::slli(bitblock256_t arg1)
+{
+        return avx_general_combine256(_mm_slli_epi16(avx_select_hi128(arg1), (int32_t)(sh)), _mm_slli_epi16(avx_select_lo128(arg1), (int32_t)(sh)));
+}
+
+//The total number of operations is 4.0
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::slli(bitblock256_t arg1)
+{
+        return avx_general_combine256(_mm_slli_epi32(avx_select_hi128(arg1), (int32_t)(sh)), _mm_slli_epi32(avx_select_lo128(arg1), (int32_t)(sh)));
+}
+
+//The total number of operations is 4.0
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::slli(bitblock256_t arg1)
+{
+        return avx_general_combine256(_mm_slli_epi64(avx_select_hi128(arg1), (int32_t)(sh)), _mm_slli_epi64(avx_select_lo128(arg1), (int32_t)(sh)));
+}
+
+//The total number of operations is 8.33333333333
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::slli(bitblock256_t arg1)
+{
+        return (((sh%8) == 0) ? avx_byte_shift_left(arg1, (sh/8)) : ((sh >= 64) ? simd256<64>::slli<(sh&63)>(avx_byte_shift_left(arg1, 8)) : simd_or(simd256<64>::slli<sh>(arg1), avx_byte_shift_left(simd256<64>::srli<((128-sh)&63)>(arg1), 8))));
+}
+
+//The total number of operations is 14.0
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::slli(bitblock256_t arg1)
+{
+        return ((sh < 128) ? simd_or(simd256<128>::slli<sh>(arg1), avx_move_lo128_to_hi128(simd256<128>::srli<(128-sh)>(arg1))) : simd256<128>::slli<(sh-128)>(avx_move_lo128_to_hi128(arg1)));
+}
+
+//The total number of operations is 3.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<1>::ifh(bitblock256_t arg1, bitblock256_t arg2, bitblock256_t arg3)
+{
+        return simd_or(simd_and(arg2, arg1), simd_andc(arg3, arg1));
+}
+
+//The total number of operations is 11.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::ifh(bitblock256_t arg1, bitblock256_t arg2, bitblock256_t arg3)
+{
+        return simd256<(1)>::ifh(simd256<1>::ifh(simd256<2>::himask(), arg1, simd256<2>::srli<(1)>(arg1)), arg2, arg3);
+}
+
+//The total number of operations is 19.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::ifh(bitblock256_t arg1, bitblock256_t arg2, bitblock256_t arg3)
+{
+        return simd256<(2)>::ifh(simd256<1>::ifh(simd256<4>::himask(), arg1, simd256<4>::srli<(2)>(arg1)), arg2, arg3);
+}
+
+//The total number of operations is 8.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::ifh(bitblock256_t arg1, bitblock256_t arg2, bitblock256_t arg3)
+{
+        return simd256<1>::ifh(simd256<8>::gt(simd256<8>::constant<0>(), arg1), arg2, arg3);
+}
+
+//The total number of operations is 8.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::ifh(bitblock256_t arg1, bitblock256_t arg2, bitblock256_t arg3)
+{
+        return simd256<1>::ifh(simd256<16>::gt(simd256<16>::constant<0>(), arg1), arg2, arg3);
+}
+
+//The total number of operations is 8.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::ifh(bitblock256_t arg1, bitblock256_t arg2, bitblock256_t arg3)
+{
+        return simd256<1>::ifh(simd256<32>::gt(simd256<32>::constant<0>(), arg1), arg2, arg3);
+}
+
+//The total number of operations is 1.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::ifh(bitblock256_t arg1, bitblock256_t arg2, bitblock256_t arg3)
+{
+        return (bitblock256_t)_mm256_blendv_pd((__m256d)(arg3), (__m256d)(arg2), (__m256d)(arg1));
+}
+
+//The total number of operations is 12.3333333333
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::ifh(bitblock256_t arg1, bitblock256_t arg2, bitblock256_t arg3)
+{
+        return simd256<(64)>::ifh(simd256<1>::ifh(simd256<128>::himask(), arg1, simd256<128>::srli<(64)>(arg1)), arg2, arg3);
+}
+
+//The total number of operations is 29.8333333333
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::ifh(bitblock256_t arg1, bitblock256_t arg2, bitblock256_t arg3)
+{
+        return simd256<(128)>::ifh(simd256<1>::ifh(simd256<256>::himask(), arg1, simd256<256>::srli<(128)>(arg1)), arg2, arg3);
+}
+
+//The total number of operations is 1.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<1>::sub(bitblock256_t arg1, bitblock256_t arg2)
+{
+        return simd_xor(arg1, arg2);
+}
+
+//The total number of operations is 16.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::sub(bitblock256_t arg1, bitblock256_t arg2)
+{
+        bitblock256_t ans = simd256<(1)>::sub(arg1, arg2);
+        bitblock256_t borrowMask = simd_or(simd_andc(arg2, arg1), simd_and(simd_not(simd_xor(arg1, arg2)), ans));
+        bitblock256_t loMask = simd256<2>::lomask();
+        bitblock256_t borrow = simd256<2>::slli<1>(simd_and(borrowMask, loMask));
+        return simd256<1>::ifh(loMask, ans, simd256<(1)>::sub(ans, borrow));
+}
+
+//The total number of operations is 14.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::sub(bitblock256_t arg1, bitblock256_t arg2)
+{
+        return simd256<1>::ifh(simd256<(8)>::himask(), simd256<(8)>::sub(arg1, simd_and(simd256<(8)>::himask(), arg2)), simd256<(8)>::sub(arg1, arg2));
+}
+
+//The total number of operations is 5.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::sub(bitblock256_t arg1, bitblock256_t arg2)
+{
+        return avx_general_combine256(_mm_sub_epi8(avx_select_hi128(arg1), avx_select_hi128(arg2)), _mm_sub_epi8(avx_select_lo128(arg1), avx_select_lo128(arg2)));
+}
+
+//The total number of operations is 5.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::sub(bitblock256_t arg1, bitblock256_t arg2)
+{
+        return avx_general_combine256(_mm_sub_epi16(avx_select_hi128(arg1), avx_select_hi128(arg2)), _mm_sub_epi16(avx_select_lo128(arg1), avx_select_lo128(arg2)));
+}
+
+//The total number of operations is 5.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::sub(bitblock256_t arg1, bitblock256_t arg2)
+{
+        return avx_general_combine256(_mm_sub_epi32(avx_select_hi128(arg1), avx_select_hi128(arg2)), _mm_sub_epi32(avx_select_lo128(arg1), avx_select_lo128(arg2)));
+}
+
+//The total number of operations is 5.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::sub(bitblock256_t arg1, bitblock256_t arg2)
+{
+        return avx_general_combine256(_mm_sub_epi64(avx_select_hi128(arg1), avx_select_hi128(arg2)), _mm_sub_epi64(avx_select_lo128(arg1), avx_select_lo128(arg2)));
+}
+
+//The total number of operations is 26.3333333333
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::sub(bitblock256_t arg1, bitblock256_t arg2)
+{
+        bitblock256_t partial = simd256<(64)>::sub(arg1, arg2);
+        bitblock256_t borrowMask = simd_or(simd_andc(arg2, arg1), simd_andc(partial, simd_xor(arg1, arg2)));
+        bitblock256_t borrow = simd256<128>::slli<(64)>(simd256<(64)>::srli<(63)>(borrowMask));
+        return simd256<(64)>::sub(partial, borrow);
+}
+
+//The total number of operations is 75.6666666667
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::sub(bitblock256_t arg1, bitblock256_t arg2)
+{
+        bitblock256_t ans = simd256<(128)>::sub(arg1, arg2);
+        bitblock256_t borrowMask = simd_or(simd_andc(arg2, arg1), simd_and(simd_not(simd_xor(arg1, arg2)), ans));
+        bitblock256_t loMask = simd256<256>::lomask();
+        bitblock256_t borrow = simd256<256>::slli<1>(simd_and(borrowMask, loMask));
+        return simd256<1>::ifh(loMask, ans, simd256<(128)>::sub(ans, borrow));
+}
+
+//The total number of operations is 10.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::add_hl(bitblock256_t arg1)
+{
+        return simd256<16>::sub(arg1, simd_and(simd256<2>::lomask(), simd256<16>::srli<1>(arg1)));
+}
+
+//The total number of operations is 11.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::add_hl(bitblock256_t arg1)
+{
+        return simd256<(8)>::add(simd256<4>::srli<(2)>(arg1), simd_and(arg1, simd256<4>::lomask()));
+}
+
+//The total number of operations is 11.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::add_hl(bitblock256_t arg1)
+{
+        return simd256<(16)>::add(simd256<8>::srli<(4)>(arg1), simd_and(arg1, simd256<8>::lomask()));
+}
+
+//The total number of operations is 10.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::add_hl(bitblock256_t arg1)
+{
+        return simd256<(32)>::add(simd256<16>::srli<(8)>(arg1), simd_and(arg1, simd256<16>::lomask()));
+}
+
+//The total number of operations is 10.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::add_hl(bitblock256_t arg1)
+{
+        return simd256<(64)>::add(simd256<32>::srli<(16)>(arg1), simd_and(arg1, simd256<32>::lomask()));
+}
+
+//The total number of operations is 10.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::add_hl(bitblock256_t arg1)
+{
+        return simd256<64>::add(simd256<64>::srli<(32)>(arg1), simd_and(arg1, simd256<64>::lomask()));
+}
+
+//The total number of operations is 35.6666666667
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::add_hl(bitblock256_t arg1)
+{
+        return simd256<128>::add(simd256<128>::srli<(64)>(arg1), simd_and(arg1, simd256<128>::lomask()));
+}
+
+//The total number of operations is 91.1666666667
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::add_hl(bitblock256_t arg1)
+{
+        return simd256<256>::add(simd256<256>::srli<(128)>(arg1), simd_and(arg1, simd256<256>::lomask()));
+}
+
+//The total number of operations is 0
+template <> template <FieldType<1>::T val> IDISA_ALWAYS_INLINE bitblock256_t simd256<1>::constant()
+{
+        return simd256<32>::constant<(-1*val)>();
+}
+
+//The total number of operations is 0
+template <> template <FieldType<2>::T val> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::constant()
+{
+        return ((val < 0) ? simd256<(4)>::constant<((val<<2)|(val^(-4)))>() : simd256<(4)>::constant<((val<<2)|val)>());
+}
+
+//The total number of operations is 0
+template <> template <FieldType<4>::T val> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::constant()
+{
+        return ((val < 0) ? simd256<(8)>::constant<((val<<4)|(val^(-16)))>() : simd256<(8)>::constant<((val<<4)|val)>());
+}
+
+//The total number of operations is 0
+template <> template <FieldType<8>::T val> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::constant()
+{
+        return (bitblock256_t)_mm256_set1_epi8((int32_t)(val));
+}
+
+//The total number of operations is 0
+template <> template <FieldType<16>::T val> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::constant()
+{
+        return (bitblock256_t)_mm256_set1_epi16((int32_t)(val));
+}
+
+//The total number of operations is 0
+template <> template <FieldType<32>::T val> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::constant()
+{
+        return (bitblock256_t)_mm256_set1_epi32((int32_t)(val));
+}
+
+//The total number of operations is 0
+template <> template <FieldType<64>::T val> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::constant()
+{
+        return ((bitblock256_t)(_mm256_set_epi32((int32_t)((val>>32)), (int32_t)(val), (int32_t)((val>>32)), (int32_t)(val), (int32_t)((val>>32)), (int32_t)(val), (int32_t)((val>>32)), (int32_t)(val))));
+}
+
+//The total number of operations is 0
+template <> template <FieldType<128>::T val> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::constant()
+{
+        return ((bitblock256_t)(_mm256_set_epi32((int32_t)(0), (int32_t)(0), (int32_t)((val>>32)), (int32_t)(val), (int32_t)(0), (int32_t)(0), (int32_t)((val>>32)), (int32_t)(val))));
+}
+
+//The total number of operations is 0
+template <> template <FieldType<256>::T val> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::constant()
+{
+        return ((bitblock256_t)(_mm256_set_epi32((int32_t)(0), (int32_t)(0), (int32_t)(0), (int32_t)(0), (int32_t)(0), (int32_t)(0), (int32_t)((val>>32)), (int32_t)(val))));
+}
+
+//The total number of operations is 1.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<1>::min(bitblock256_t arg1, bitblock256_t arg2)
+{
+        return simd_or(arg1, arg2);
+}
+
+//The total number of operations is 25.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::min(bitblock256_t arg1, bitblock256_t arg2)
+{
+        bitblock256_t hiAns = simd256<(1)>::min(arg1, arg2);
+        bitblock256_t loAns = simd256<(1)>::umin(arg1, arg2);
+        bitblock256_t eqMask1 = simd256<2>::srli<(1)>(simd256<(1)>::eq(hiAns, arg1));
+        bitblock256_t eqMask2 = simd256<2>::srli<(1)>(simd256<(1)>::eq(hiAns, arg2));
+        return simd256<1>::ifh(simd256<2>::himask(), hiAns, simd256<1>::ifh(eqMask1, simd256<1>::ifh(eqMask2, loAns, arg1), arg2));
+}
+
+//The total number of operations is 17.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::min(bitblock256_t arg1, bitblock256_t arg2)
+{
+        bitblock256_t high_bit = simd256<4>::constant<(8)>();
+        return simd_xor(simd256<4>::umin(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit)), high_bit);
+}
+
+//The total number of operations is 5.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::min(bitblock256_t arg1, bitblock256_t arg2)
+{
+        return avx_general_combine256(_mm_min_epi8(avx_select_hi128(arg1), avx_select_hi128(arg2)), _mm_min_epi8(avx_select_lo128(arg1), avx_select_lo128(arg2)));
+}
+
+//The total number of operations is 5.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::min(bitblock256_t arg1, bitblock256_t arg2)
+{
+        return avx_general_combine256(_mm_min_epi16(avx_select_hi128(arg1), avx_select_hi128(arg2)), _mm_min_epi16(avx_select_lo128(arg1), avx_select_lo128(arg2)));
+}
+
+//The total number of operations is 5.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::min(bitblock256_t arg1, bitblock256_t arg2)
+{
+        return avx_general_combine256(_mm_min_epi32(avx_select_hi128(arg1), avx_select_hi128(arg2)), _mm_min_epi32(avx_select_lo128(arg1), avx_select_lo128(arg2)));
+}
+
+//The total number of operations is 8.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::min(bitblock256_t arg1, bitblock256_t arg2)
+{
+        return simd256<1>::ifh(simd256<64>::gt(arg1, arg2), arg2, arg1);
+}
+
+//The total number of operations is 54.6666666667
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::min(bitblock256_t arg1, bitblock256_t arg2)
+{
+        bitblock256_t hiAns = simd256<(64)>::min(arg1, arg2);
+        bitblock256_t loAns = simd256<(64)>::umin(arg1, arg2);
+        bitblock256_t eqMask1 = simd256<128>::srli<(64)>(simd256<(64)>::eq(hiAns, arg1));
+        bitblock256_t eqMask2 = simd256<128>::srli<(64)>(simd256<(64)>::eq(hiAns, arg2));
+        return simd256<1>::ifh(simd256<128>::himask(), hiAns, simd256<1>::ifh(eqMask1, simd256<1>::ifh(eqMask2, loAns, arg1), arg2));
+}
+
+//The total number of operations is 186.666666667
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::min(bitblock256_t arg1, bitblock256_t arg2)
+{
+        bitblock256_t hiAns = simd256<(128)>::min(arg1, arg2);
+        bitblock256_t loAns = simd256<(128)>::umin(arg1, arg2);
+        bitblock256_t eqMask1 = simd256<256>::srli<(128)>(simd256<(128)>::eq(hiAns, arg1));
+        bitblock256_t eqMask2 = simd256<256>::srli<(128)>(simd256<(128)>::eq(hiAns, arg2));
+        return simd256<1>::ifh(simd256<256>::himask(), hiAns, simd256<1>::ifh(eqMask1, simd256<1>::ifh(eqMask2, loAns, arg1), arg2));
+}
+
+//The total number of operations is 0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::lomask()
+{
+        return simd256<2>::constant<(1)>();
+}
+
+//The total number of operations is 0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::lomask()
+{
+        return simd256<4>::constant<(3)>();
+}
+
+//The total number of operations is 0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::lomask()
+{
+        return simd256<8>::constant<(15)>();
+}
+
+//The total number of operations is 0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::lomask()
+{
+        return simd256<16>::constant<(255)>();
+}
+
+//The total number of operations is 0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::lomask()
+{
+        return simd256<32>::constant<(65535)>();
+}
+
+//The total number of operations is 0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::lomask()
+{
+        return ((bitblock256_t)(_mm256_set_epi32((int32_t)(0), (int32_t)(-1), (int32_t)(0), (int32_t)(-1), (int32_t)(0), (int32_t)(-1), (int32_t)(0), (int32_t)(-1))));
+}
+
+//The total number of operations is 0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::lomask()
+{
+        return ((bitblock256_t)(_mm256_set_epi32((int32_t)(0), (int32_t)(0), (int32_t)(-1), (int32_t)(-1), (int32_t)(0), (int32_t)(0), (int32_t)(-1), (int32_t)(-1))));
+}
+
+//The total number of operations is 0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::lomask()
+{
+        return ((bitblock256_t)(_mm256_set_epi32((int32_t)(0), (int32_t)(0), (int32_t)(0), (int32_t)(0), (int32_t)(-1), (int32_t)(-1), (int32_t)(-1), (int32_t)(-1))));
+}
+
+//The total number of operations is 1.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<1>::add(bitblock256_t arg1, bitblock256_t arg2)
+{
+        return simd_xor(arg1, arg2);
+}
+
+//The total number of operations is 16.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::add(bitblock256_t arg1, bitblock256_t arg2)
+{
+        bitblock256_t ans = simd256<(1)>::add(arg1, arg2);
+        bitblock256_t carryMask = simd_or(simd_and(arg1, arg2), simd_and(simd_xor(arg1, arg2), simd_not(ans)));
+        bitblock256_t loMask = simd256<2>::lomask();
+        bitblock256_t carry = simd256<2>::slli<1>(simd_and(carryMask, loMask));
+        return simd256<1>::ifh(loMask, ans, simd256<(1)>::add(ans, carry));
+}
+
+//The total number of operations is 14.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::add(bitblock256_t arg1, bitblock256_t arg2)
+{
+        return simd256<1>::ifh(simd256<(8)>::himask(), simd256<(8)>::add(arg1, simd_and(simd256<(8)>::himask(), arg2)), simd256<(8)>::add(arg1, arg2));
+}
+
+//The total number of operations is 5.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::add(bitblock256_t arg1, bitblock256_t arg2)
+{
+        return avx_general_combine256(_mm_add_epi8(avx_select_hi128(arg1), avx_select_hi128(arg2)), _mm_add_epi8(avx_select_lo128(arg1), avx_select_lo128(arg2)));
+}
+
+//The total number of operations is 5.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::add(bitblock256_t arg1, bitblock256_t arg2)
+{
+        return avx_general_combine256(_mm_add_epi16(avx_select_hi128(arg1), avx_select_hi128(arg2)), _mm_add_epi16(avx_select_lo128(arg1), avx_select_lo128(arg2)));
+}
+
+//The total number of operations is 5.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::add(bitblock256_t arg1, bitblock256_t arg2)
+{
+        return avx_general_combine256(_mm_add_epi32(avx_select_hi128(arg1), avx_select_hi128(arg2)), _mm_add_epi32(avx_select_lo128(arg1), avx_select_lo128(arg2)));
+}
+
+//The total number of operations is 5.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::add(bitblock256_t arg1, bitblock256_t arg2)
+{
+        return avx_general_combine256(_mm_add_epi64(avx_select_hi128(arg1), avx_select_hi128(arg2)), _mm_add_epi64(avx_select_lo128(arg1), avx_select_lo128(arg2)));
+}
+
+//The total number of operations is 26.3333333333
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::add(bitblock256_t arg1, bitblock256_t arg2)
+{
+        bitblock256_t partial = simd256<(64)>::add(arg1, arg2);
+        bitblock256_t carryMask = simd_or(simd_and(arg1, arg2), simd_andc(simd_xor(arg1, arg2), partial));
+        bitblock256_t carry = simd256<128>::slli<(64)>(simd256<(64)>::srli<(63)>(carryMask));
+        return simd256<(64)>::add(partial, carry);
+}
+
+//The total number of operations is 75.6666666667
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::add(bitblock256_t arg1, bitblock256_t arg2)
+{
+        bitblock256_t ans = simd256<(128)>::add(arg1, arg2);
+        bitblock256_t carryMask = simd_or(simd_and(arg1, arg2), simd_and(simd_xor(arg1, arg2), simd_not(ans)));
+        bitblock256_t loMask = simd256<256>::lomask();
+        bitblock256_t carry = simd256<256>::slli<1>(simd_and(carryMask, loMask));
+        return simd256<1>::ifh(loMask, ans, simd256<(128)>::add(ans, carry));
+}
+
+//The total number of operations is 1.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<1>::umin(bitblock256_t arg1, bitblock256_t arg2)
+{
+        return simd_and(arg1, arg2);
+}
+
+//The total number of operations is 24.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::umin(bitblock256_t arg1, bitblock256_t arg2)
+{
+        bitblock256_t tmpAns = simd256<(1)>::umin(arg1, arg2);
+        bitblock256_t eqMask1 = simd256<2>::srli<(1)>(simd256<(1)>::eq(tmpAns, arg1));
+        bitblock256_t eqMask2 = simd256<2>::srli<(1)>(simd256<(1)>::eq(tmpAns, arg2));
+        return simd256<1>::ifh(simd256<2>::himask(), tmpAns, simd256<1>::ifh(eqMask1, simd256<1>::ifh(eqMask2, tmpAns, arg1), arg2));
+}
+
+//The total number of operations is 14.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::umin(bitblock256_t arg1, bitblock256_t arg2)
+{
+        return simd_or(simd_and(simd256<(8)>::himask(), simd256<(8)>::umin(arg1, arg2)), simd256<(8)>::umin(simd_and(simd256<(8)>::lomask(), arg1), simd_and(simd256<(8)>::lomask(), arg2)));
+}
+
+//The total number of operations is 5.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::umin(bitblock256_t arg1, bitblock256_t arg2)
+{
+        return avx_general_combine256(_mm_min_epu8(avx_select_hi128(arg1), avx_select_hi128(arg2)), _mm_min_epu8(avx_select_lo128(arg1), avx_select_lo128(arg2)));
+}
+
+//The total number of operations is 5.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::umin(bitblock256_t arg1, bitblock256_t arg2)
+{
+        return avx_general_combine256(_mm_min_epu16(avx_select_hi128(arg1), avx_select_hi128(arg2)), _mm_min_epu16(avx_select_lo128(arg1), avx_select_lo128(arg2)));
+}
+
+//The total number of operations is 5.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::umin(bitblock256_t arg1, bitblock256_t arg2)
+{
+        return avx_general_combine256(_mm_min_epu32(avx_select_hi128(arg1), avx_select_hi128(arg2)), _mm_min_epu32(avx_select_lo128(arg1), avx_select_lo128(arg2)));
+}
+
+//The total number of operations is 11.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::umin(bitblock256_t arg1, bitblock256_t arg2)
+{
+        bitblock256_t high_bit = simd256<64>::constant<(9223372036854775808ULL)>();
+        return simd_xor(simd256<64>::min(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit)), high_bit);
+}
+
+//The total number of operations is 46.6666666667
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::umin(bitblock256_t arg1, bitblock256_t arg2)
+{
+        bitblock256_t tmpAns = simd256<(64)>::umin(arg1, arg2);
+        bitblock256_t eqMask1 = simd256<128>::srli<(64)>(simd256<(64)>::eq(tmpAns, arg1));
+        bitblock256_t eqMask2 = simd256<128>::srli<(64)>(simd256<(64)>::eq(tmpAns, arg2));
+        return simd256<1>::ifh(simd256<128>::himask(), tmpAns, simd256<1>::ifh(eqMask1, simd256<1>::ifh(eqMask2, tmpAns, arg1), arg2));
+}
+
+//The total number of operations is 132.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::umin(bitblock256_t arg1, bitblock256_t arg2)
+{
+        bitblock256_t tmpAns = simd256<(128)>::umin(arg1, arg2);
+        bitblock256_t eqMask1 = simd256<256>::srli<(128)>(simd256<(128)>::eq(tmpAns, arg1));
+        bitblock256_t eqMask2 = simd256<256>::srli<(128)>(simd256<(128)>::eq(tmpAns, arg2));
+        return simd256<1>::ifh(simd256<256>::himask(), tmpAns, simd256<1>::ifh(eqMask1, simd256<1>::ifh(eqMask2, tmpAns, arg1), arg2));
+}
+
+//The total number of operations is 19.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::abs(bitblock256_t arg1)
+{
+        return simd256<1>::ifh(simd256<2>::himask(), simd_and(arg1, simd256<256>::slli<1>(simd_not(arg1))), arg1);
+}
+
+//The total number of operations is 39.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::abs(bitblock256_t arg1)
+{
+        bitblock256_t gtMask = simd256<4>::gt(arg1, simd256<4>::constant<0>());
+        return simd256<1>::ifh(gtMask, arg1, simd256<4>::sub(gtMask, arg1));
+}
+
+//The total number of operations is 4.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::abs(bitblock256_t arg1)
+{
+        return avx_general_combine256(_mm_abs_epi8(avx_select_hi128(arg1)), _mm_abs_epi8(avx_select_lo128(arg1)));
+}
+
+//The total number of operations is 4.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::abs(bitblock256_t arg1)
+{
+        return avx_general_combine256(_mm_abs_epi16(avx_select_hi128(arg1)), _mm_abs_epi16(avx_select_lo128(arg1)));
+}
+
+//The total number of operations is 4.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::abs(bitblock256_t arg1)
+{
+        return avx_general_combine256(_mm_abs_epi32(avx_select_hi128(arg1)), _mm_abs_epi32(avx_select_lo128(arg1)));
+}
+
+//The total number of operations is 13.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::abs(bitblock256_t arg1)
+{
+        bitblock256_t gtMask = simd256<64>::gt(arg1, simd256<64>::constant<0>());
+        return simd256<1>::ifh(gtMask, arg1, simd256<64>::sub(gtMask, arg1));
+}
+
+//The total number of operations is 69.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::abs(bitblock256_t arg1)
+{
+        bitblock256_t eqMask = simd256<128>::eq(simd256<1>::ifh(simd256<128>::himask(), simd256<(64)>::abs(arg1), arg1), arg1);
+        return simd256<1>::ifh(eqMask, arg1, simd256<128>::sub(eqMask, arg1));
+}
+
+//The total number of operations is 204.833333333
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::abs(bitblock256_t arg1)
+{
+        bitblock256_t eqMask = simd256<256>::eq(simd256<1>::ifh(simd256<256>::himask(), simd256<(128)>::abs(arg1), arg1), arg1);
+        return simd256<1>::ifh(eqMask, arg1, simd256<256>::sub(eqMask, arg1));
+}
+
+//The total number of operations is 2.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<1>::eq(bitblock256_t arg1, bitblock256_t arg2)
+{
+        return simd_not(simd_xor(arg1, arg2));
+}
+
+//The total number of operations is 14.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::eq(bitblock256_t arg1, bitblock256_t arg2)
+{
+        bitblock256_t tmpAns = simd256<(1)>::eq(arg1, arg2);
+        bitblock256_t loMask = simd_and(tmpAns, simd256<2>::srli<(1)>(tmpAns));
+        bitblock256_t hiMask = simd256<2>::slli<(1)>(loMask);
+        return simd_or(loMask, hiMask);
+}
+
+//The total number of operations is 17.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::eq(bitblock256_t arg1, bitblock256_t arg2)
+{
+        return simd_or(simd_and(simd256<(8)>::himask(), simd256<(8)>::eq(simd_and(simd256<(8)>::himask(), arg1), simd_and(simd256<(8)>::himask(), arg2))), simd_and(simd256<(8)>::lomask(), simd256<(8)>::eq(simd_and(simd256<(8)>::lomask(), arg1), simd_and(simd256<(8)>::lomask(), arg2))));
+}
+
+//The total number of operations is 5.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::eq(bitblock256_t arg1, bitblock256_t arg2)
+{
+        return avx_general_combine256(_mm_cmpeq_epi8(avx_select_hi128(arg1), avx_select_hi128(arg2)), _mm_cmpeq_epi8(avx_select_lo128(arg1), avx_select_lo128(arg2)));
+}
+
+//The total number of operations is 5.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::eq(bitblock256_t arg1, bitblock256_t arg2)
+{
+        return avx_general_combine256(_mm_cmpeq_epi16(avx_select_hi128(arg1), avx_select_hi128(arg2)), _mm_cmpeq_epi16(avx_select_lo128(arg1), avx_select_lo128(arg2)));
+}
+
+//The total number of operations is 5.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::eq(bitblock256_t arg1, bitblock256_t arg2)
+{
+        return avx_general_combine256(_mm_cmpeq_epi32(avx_select_hi128(arg1), avx_select_hi128(arg2)), _mm_cmpeq_epi32(avx_select_lo128(arg1), avx_select_lo128(arg2)));
+}
+
+//The total number of operations is 5.0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::eq(bitblock256_t arg1, bitblock256_t arg2)
+{
+        return avx_general_combine256(_mm_cmpeq_epi64(avx_select_hi128(arg1), avx_select_hi128(arg2)), _mm_cmpeq_epi64(avx_select_lo128(arg1), avx_select_lo128(arg2)));
+}
+
+//The total number of operations is 23.6666666667
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::eq(bitblock256_t arg1, bitblock256_t arg2)
+{
+        bitblock256_t tmpAns = simd256<(64)>::eq(arg1, arg2);
+        bitblock256_t loMask = simd_and(tmpAns, simd256<128>::srli<(64)>(tmpAns));
+        bitblock256_t hiMask = simd256<128>::slli<(64)>(loMask);
+        return simd_or(loMask, hiMask);
+}
+
+//The total number of operations is 54.1666666667
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::eq(bitblock256_t arg1, bitblock256_t arg2)
+{
+        bitblock256_t tmpAns = simd256<(128)>::eq(arg1, arg2);
+        bitblock256_t loMask = simd_and(tmpAns, simd256<256>::srli<(128)>(tmpAns));
+        bitblock256_t hiMask = simd256<256>::slli<(128)>(loMask);
+        return simd_or(loMask, hiMask);
+}
+
+//The total number of operations is 7.0
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::srai(bitblock256_t arg1)
+{
+        return ((sh == 0) ? arg1 : simd_or(simd_and(simd256<2>::himask(), arg1), simd256<2>::srli<1>(arg1)));
+}
+
+//The total number of operations is 17.5
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::srai(bitblock256_t arg1)
+{
+        return simd_or(simd_and(simd256<4>::himask(), simd256<(2)>::srai<((sh < (2)) ? sh : (2))>(arg1)), ((sh <= (2)) ? simd256<4>::srli<sh>(arg1) : simd256<(2)>::srai<(sh-(2))>(simd256<4>::srli<(2)>(arg1))));
+}
+
+//The total number of operations is 12.0
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::srai(bitblock256_t arg1)
+{
+        bitblock256_t tmp = simd256<8>::srli<((sh >= 8) ? (7) : ((sh < 0) ? 0 : sh))>(arg1);
+        return simd_or(tmp, simd256<8>::sub(simd256<8>::constant<0>(), simd_and(simd256<8>::constant<(1<<((8-((sh >= 8) ? (7) : ((sh < 0) ? 0 : sh)))-1))>(), tmp)));
+}
+
+//The total number of operations is 4.0
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::srai(bitblock256_t arg1)
+{
+        return avx_general_combine256(_mm_srai_epi16(avx_select_hi128(arg1), (int32_t)(sh)), _mm_srai_epi16(avx_select_lo128(arg1), (int32_t)(sh)));
+}
+
+//The total number of operations is 4.0
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::srai(bitblock256_t arg1)
+{
+        return avx_general_combine256(_mm_srai_epi32(avx_select_hi128(arg1), (int32_t)(sh)), _mm_srai_epi32(avx_select_lo128(arg1), (int32_t)(sh)));
+}
+
+//The total number of operations is 12.0
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::srai(bitblock256_t arg1)
+{
+        return simd_or(simd_and(simd256<64>::himask(), simd256<(32)>::srai<((sh < (32)) ? sh : (32))>(arg1)), ((sh <= (32)) ? simd256<64>::srli<sh>(arg1) : simd256<(32)>::srai<(sh-(32))>(simd256<64>::srli<(32)>(arg1))));
+}
+
+//The total number of operations is 28.3333333333
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::srai(bitblock256_t arg1)
+{
+        return simd_or(simd_and(simd256<128>::himask(), simd256<(64)>::srai<((sh < (64)) ? sh : (64))>(arg1)), ((sh <= (64)) ? simd256<128>::srli<sh>(arg1) : simd256<(64)>::srai<(sh-(64))>(simd256<128>::srli<(64)>(arg1))));
+}
+
+//The total number of operations is 59.0
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::srai(bitblock256_t arg1)
+{
+        return simd_or(simd_and(simd256<256>::himask(), simd256<(128)>::srai<((sh < (128)) ? sh : (128))>(arg1)), ((sh <= (128)) ? simd256<256>::srli<sh>(arg1) : simd256<(128)>::srai<(sh-(128))>(simd256<256>::srli<(128)>(arg1))));
+}
+
+//The total number of operations is 0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::himask()
+{
+        return simd256<2>::constant<(2)>();
+}
+
+//The total number of operations is 0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::himask()
+{
+        return simd256<4>::constant<(12)>();
+}
+
+//The total number of operations is 0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::himask()
+{
+        return simd256<8>::constant<(240)>();
+}
+
+//The total number of operations is 0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::himask()
+{
+        return simd256<16>::constant<(65280)>();
+}
+
+//The total number of operations is 0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::himask()
+{
+        return simd256<32>::constant<-65536>();
+}
+
+//The total number of operations is 0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::himask()
+{
+        return ((bitblock256_t)(_mm256_set_epi32((int32_t)(-1), (int32_t)(0), (int32_t)(-1), (int32_t)(0), (int32_t)(-1), (int32_t)(0), (int32_t)(-1), (int32_t)(0))));
+}
+
+//The total number of operations is 0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::himask()
+{
+        return ((bitblock256_t)(_mm256_set_epi32((int32_t)(-1), (int32_t)(-1), (int32_t)(0), (int32_t)(0), (int32_t)(-1), (int32_t)(-1), (int32_t)(0), (int32_t)(0))));
+}
+
+//The total number of operations is 0
+template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::himask()
+{
+        return ((bitblock256_t)(_mm256_set_epi32((int32_t)(-1), (int32_t)(-1), (int32_t)(-1), (int32_t)(-1), (int32_t)(0), (int32_t)(0), (int32_t)(0), (int32_t)(0))));
+}
+
 //The total number of operations is 1.0
 template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<1>::lt(bitblock256_t arg1, bitblock256_t arg2)
@@ -1029 +2193 @@
 }
 
-//The total number of operations is 5.0
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::srli(bitblock256_t arg1)
-{
-        return simd_and(simd256<32>::srli<sh>(arg1), simd256<2>::constant<((3)>>sh)>());
-}
-
-//The total number of operations is 5.0
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::srli(bitblock256_t arg1)
-{
-        return simd_and(simd256<32>::srli<sh>(arg1), simd256<4>::constant<((15)>>sh)>());
-}
-
-//The total number of operations is 5.0
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::srli(bitblock256_t arg1)
-{
-        return simd_and(simd256<32>::srli<sh>(arg1), simd256<8>::constant<((255)>>sh)>());
-}
-
-//The total number of operations is 4.0
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::srli(bitblock256_t arg1)
-{
-        return avx_general_combine256(_mm_srli_epi16(avx_select_hi128(arg1), (int32_t)(sh)), _mm_srli_epi16(avx_select_lo128(arg1), (int32_t)(sh)));
-}
-
-//The total number of operations is 4.0
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::srli(bitblock256_t arg1)
-{
-        return avx_general_combine256(_mm_srli_epi32(avx_select_hi128(arg1), (int32_t)(sh)), _mm_srli_epi32(avx_select_lo128(arg1), (int32_t)(sh)));
-}
-
-//The total number of operations is 4.0
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::srli(bitblock256_t arg1)
-{
-        return avx_general_combine256(_mm_srli_epi64(avx_select_hi128(arg1), (int32_t)(sh)), _mm_srli_epi64(avx_select_lo128(arg1), (int32_t)(sh)));
-}
-
-//The total number of operations is 8.33333333333
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::srli(bitblock256_t arg1)
-{
-        return (((sh%8) == 0) ? avx_byte_shift_right(arg1, (sh/8)) : ((sh >= 64) ? simd256<64>::srli<(sh&63)>(avx_byte_shift_right(arg1, 8)) : simd_or(simd256<64>::srli<sh>(arg1), avx_byte_shift_right(simd256<64>::slli<((128-sh)&63)>(arg1), 8))));
-}
-
-//The total number of operations is 14.5
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::srli(bitblock256_t arg1)
-{
-        return ((sh < 128) ? simd_or(simd256<128>::srli<sh>(arg1), simd256<128>::slli<((256-sh)&127)>(((bitblock256_t)(_mm256_castsi128_si256(avx_select_hi128(arg1)))))) : simd256<128>::srli<(sh&127)>(avx_move_hi128_to_lo128(arg1)));
-}
-
-//The total number of operations is 1.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<1>::ctz(bitblock256_t arg1)
-{
-        return simd_not(arg1);
-}
-
-//The total number of operations is 27.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::ctz(bitblock256_t arg1)
-{
-        return simd256<2>::popcount(simd_andc(simd256<2>::sub(arg1, simd256<2>::constant<1>()), arg1));
-}
-
-//The total number of operations is 36.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::ctz(bitblock256_t arg1)
-{
-        return simd256<4>::popcount(simd_andc(simd256<4>::sub(arg1, simd256<4>::constant<1>()), arg1));
-}
-
-//The total number of operations is 38.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::ctz(bitblock256_t arg1)
-{
-        return simd256<8>::popcount(simd_andc(simd256<8>::sub(arg1, simd256<8>::constant<1>()), arg1));
-}
-
-//The total number of operations is 48.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::ctz(bitblock256_t arg1)
-{
-        return simd256<16>::popcount(simd_andc(simd256<16>::sub(arg1, simd256<16>::constant<1>()), arg1));
-}
-
-//The total number of operations is 58.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::ctz(bitblock256_t arg1)
-{
-        return simd256<32>::popcount(simd_andc(simd256<32>::sub(arg1, simd256<32>::constant<1>()), arg1));
-}
-
-//The total number of operations is 44.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::ctz(bitblock256_t arg1)
-{
-        return simd256<64>::popcount(simd_andc(simd256<64>::sub(arg1, simd256<64>::constant<1>()), arg1));
-}
-
-//The total number of operations is 101.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::ctz(bitblock256_t arg1)
-{
-        return simd256<128>::popcount(simd_andc(simd256<128>::sub(arg1, simd256<128>::constant<1>()), arg1));
-}
-
-//The total number of operations is 192.166666667
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::ctz(bitblock256_t arg1)
-{
-        return simd256<256>::popcount(simd_andc(simd256<256>::sub(arg1, simd256<256>::constant<1>()), arg1));
-}
-
-//The total number of operations is 1.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<1>::ugt(bitblock256_t arg1, bitblock256_t arg2)
-{
-        return simd_andc(arg1, arg2);
-}
-
-//The total number of operations is 23.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::ugt(bitblock256_t arg1, bitblock256_t arg2)
-{
-        bitblock256_t tmpAns = simd256<(1)>::ugt(arg1, arg2);
-        bitblock256_t mask = simd_and(tmpAns, simd256<2>::srli<(1)>(simd256<(1)>::eq(arg1, arg2)));
-        mask = simd_or(mask, simd256<2>::slli<(1)>(mask));
-        return simd_or(simd256<2>::srai<(1)>(tmpAns), mask);
-}
-
-//The total number of operations is 20.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::ugt(bitblock256_t arg1, bitblock256_t arg2)
-{
-        return simd256<1>::ifh(simd256<(8)>::himask(), simd256<(8)>::ugt(simd_and(simd256<(8)>::himask(), arg1), arg2), simd256<(8)>::ugt(simd_andc(arg1, simd256<(8)>::himask()), simd_andc(arg2, simd256<(8)>::himask())));
-}
-
-//The total number of operations is 7.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::ugt(bitblock256_t arg1, bitblock256_t arg2)
-{
-        bitblock256_t high_bit = simd256<8>::constant<(128)>();
-        return simd256<8>::gt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
-}
-
-//The total number of operations is 7.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::ugt(bitblock256_t arg1, bitblock256_t arg2)
-{
-        bitblock256_t high_bit = simd256<16>::constant<(32768)>();
-        return simd256<16>::gt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
-}
-
-//The total number of operations is 7.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::ugt(bitblock256_t arg1, bitblock256_t arg2)
-{
-        bitblock256_t high_bit = simd256<32>::constant<(2147483648ULL)>();
-        return simd256<32>::gt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
-}
-
-//The total number of operations is 7.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::ugt(bitblock256_t arg1, bitblock256_t arg2)
-{
-        bitblock256_t high_bit = simd256<64>::constant<(9223372036854775808ULL)>();
-        return simd256<64>::gt(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit));
-}
-
-//The total number of operations is 60.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::ugt(bitblock256_t arg1, bitblock256_t arg2)
-{
-        bitblock256_t tmpAns = simd256<(64)>::ugt(arg1, arg2);
-        bitblock256_t mask = simd_and(tmpAns, simd256<128>::srli<(64)>(simd256<(64)>::eq(arg1, arg2)));
-        mask = simd_or(mask, simd256<128>::slli<(64)>(mask));
-        return simd_or(simd256<128>::srai<(64)>(tmpAns), mask);
-}
-
-//The total number of operations is 174.166666667
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::ugt(bitblock256_t arg1, bitblock256_t arg2)
-{
-        bitblock256_t tmpAns = simd256<(128)>::ugt(arg1, arg2);
-        bitblock256_t mask = simd_and(tmpAns, simd256<256>::srli<(128)>(simd256<(128)>::eq(arg1, arg2)));
-        mask = simd_or(mask, simd256<256>::slli<(128)>(mask));
-        return simd_or(simd256<256>::srai<(128)>(tmpAns), mask);
-}
-
-//The total number of operations is 7.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::xor_hl(bitblock256_t arg1)
-{
-        return simd_xor(simd256<2>::srli<(1)>(arg1), simd_and(arg1, simd256<2>::lomask()));
-}
-
-//The total number of operations is 7.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::xor_hl(bitblock256_t arg1)
-{
-        return simd_xor(simd256<4>::srli<(2)>(arg1), simd_and(arg1, simd256<4>::lomask()));
-}
-
-//The total number of operations is 7.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::xor_hl(bitblock256_t arg1)
-{
-        return simd_xor(simd256<8>::srli<(4)>(arg1), simd_and(arg1, simd256<8>::lomask()));
-}
-
-//The total number of operations is 6.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::xor_hl(bitblock256_t arg1)
-{
-        return simd_xor(simd256<16>::srli<(8)>(arg1), simd_and(arg1, simd256<16>::lomask()));
-}
-
-//The total number of operations is 6.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::xor_hl(bitblock256_t arg1)
-{
-        return simd_xor(simd256<32>::srli<(16)>(arg1), simd_and(arg1, simd256<32>::lomask()));
-}
-
-//The total number of operations is 6.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::xor_hl(bitblock256_t arg1)
-{
-        return simd_xor(simd256<64>::srli<(32)>(arg1), simd_and(arg1, simd256<64>::lomask()));
-}
-
-//The total number of operations is 10.3333333333
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::xor_hl(bitblock256_t arg1)
-{
-        return simd_xor(simd256<128>::srli<(64)>(arg1), simd_and(arg1, simd256<128>::lomask()));
-}
-
-//The total number of operations is 16.5
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::xor_hl(bitblock256_t arg1)
-{
-        return simd_xor(simd256<256>::srli<(128)>(arg1), simd_and(arg1, simd256<256>::lomask()));
-}
-
-//The total number of operations is 0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<1>::popcount(bitblock256_t arg1)
-{
-        return arg1;
-}
-
-//The total number of operations is 10.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::popcount(bitblock256_t arg1)
-{
-        return simd256<2>::add_hl(simd256<(1)>::popcount(arg1));
-}
-
-//The total number of operations is 21.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::popcount(bitblock256_t arg1)
-{
-        return simd256<4>::add_hl(simd256<(2)>::popcount(arg1));
-}
-
-//The total number of operations is 32.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::popcount(bitblock256_t arg1)
-{
-        return simd256<8>::add_hl(simd256<(4)>::popcount(arg1));
-}
-
-//The total number of operations is 42.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::popcount(bitblock256_t arg1)
-{
-        return simd256<16>::add_hl(simd256<(8)>::popcount(arg1));
-}
-
-//The total number of operations is 52.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::popcount(bitblock256_t arg1)
-{
-        return simd256<32>::add_hl(simd256<(16)>::popcount(arg1));
-}
-
-//The total number of operations is 38.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::popcount(bitblock256_t arg1)
-{
-        bitblock256_t tmpAns = simd256<8>::popcount(arg1);
-        return avx_general_combine256(_mm_sad_epu8(avx_select_hi128(tmpAns), _mm_set1_epi32((int32_t)(0))), _mm_sad_epu8(avx_select_lo128(tmpAns), _mm_set1_epi32((int32_t)(0))));
-}
-
-//The total number of operations is 73.6666666667
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::popcount(bitblock256_t arg1)
-{
-        return simd256<128>::add_hl(simd256<(64)>::popcount(arg1));
-}
-
-//The total number of operations is 115.5
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::popcount(bitblock256_t arg1)
-{
-        bitblock256_t tmpAns = simd256<(128)>::popcount(arg1);
-        return simd256<(128)>::add(simd_and(tmpAns, simd256<256>::lomask()), simd256<256>::srli<(128)>(tmpAns));
-}
-
-//The total number of operations is 16.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::neg(bitblock256_t arg1)
-{
-        return simd256<2>::sub(simd256<2>::constant<0>(), arg1);
-}
-
-//The total number of operations is 14.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::neg(bitblock256_t arg1)
-{
-        return simd256<4>::sub(simd256<4>::constant<0>(), arg1);
-}
-
-//The total number of operations is 5.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::neg(bitblock256_t arg1)
-{
-        return simd256<8>::sub(simd256<8>::constant<0>(), arg1);
-}
-
-//The total number of operations is 5.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::neg(bitblock256_t arg1)
-{
-        return simd256<16>::sub(simd256<16>::constant<0>(), arg1);
-}
-
-//The total number of operations is 5.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::neg(bitblock256_t arg1)
-{
-        return simd256<32>::sub(simd256<32>::constant<0>(), arg1);
-}
-
-//The total number of operations is 5.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::neg(bitblock256_t arg1)
-{
-        return simd256<64>::sub(simd256<64>::constant<0>(), arg1);
-}
-
-//The total number of operations is 26.3333333333
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::neg(bitblock256_t arg1)
-{
-        return simd256<128>::sub(simd256<128>::constant<0>(), arg1);
-}
-
-//The total number of operations is 75.6666666667
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::neg(bitblock256_t arg1)
-{
-        return simd256<256>::sub(simd256<256>::constant<0>(), arg1);
-}
-
-//The total number of operations is 5.0
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::slli(bitblock256_t arg1)
-{
-        return simd_and(simd256<32>::slli<sh>(arg1), simd256<2>::constant<(((3)<<sh)&(3))>());
-}
-
-//The total number of operations is 5.0
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::slli(bitblock256_t arg1)
-{
-        return simd_and(simd256<32>::slli<sh>(arg1), simd256<4>::constant<(((15)<<sh)&(15))>());
-}
-
-//The total number of operations is 5.0
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::slli(bitblock256_t arg1)
-{
-        return simd_and(simd256<32>::slli<sh>(arg1), simd256<8>::constant<(((255)<<sh)&(255))>());
-}
-
-//The total number of operations is 4.0
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::slli(bitblock256_t arg1)
-{
-        return avx_general_combine256(_mm_slli_epi16(avx_select_hi128(arg1), (int32_t)(sh)), _mm_slli_epi16(avx_select_lo128(arg1), (int32_t)(sh)));
-}
-
-//The total number of operations is 4.0
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::slli(bitblock256_t arg1)
-{
-        return avx_general_combine256(_mm_slli_epi32(avx_select_hi128(arg1), (int32_t)(sh)), _mm_slli_epi32(avx_select_lo128(arg1), (int32_t)(sh)));
-}
-
-//The total number of operations is 4.0
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::slli(bitblock256_t arg1)
-{
-        return avx_general_combine256(_mm_slli_epi64(avx_select_hi128(arg1), (int32_t)(sh)), _mm_slli_epi64(avx_select_lo128(arg1), (int32_t)(sh)));
-}
-
-//The total number of operations is 8.33333333333
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::slli(bitblock256_t arg1)
-{
-        return (((sh%8) == 0) ? avx_byte_shift_left(arg1, (sh/8)) : ((sh >= 64) ? simd256<64>::slli<(sh&63)>(avx_byte_shift_left(arg1, 8)) : simd_or(simd256<64>::slli<sh>(arg1), avx_byte_shift_left(simd256<64>::srli<((128-sh)&63)>(arg1), 8))));
-}
-
-//The total number of operations is 14.0
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::slli(bitblock256_t arg1)
-{
-        return ((sh < 128) ? simd_or(simd256<128>::slli<sh>(arg1), avx_move_lo128_to_hi128(simd256<128>::srli<((256-sh)&127)>(arg1))) : simd256<128>::slli<(sh&127)>(avx_move_lo128_to_hi128(arg1)));
-}
-
-//The total number of operations is 3.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<1>::ifh(bitblock256_t arg1, bitblock256_t arg2, bitblock256_t arg3)
-{
-        return simd_or(simd_and(arg2, arg1), simd_andc(arg3, arg1));
-}
-
-//The total number of operations is 11.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::ifh(bitblock256_t arg1, bitblock256_t arg2, bitblock256_t arg3)
-{
-        return simd256<(1)>::ifh(simd256<1>::ifh(simd256<2>::himask(), arg1, simd256<2>::srli<(1)>(arg1)), arg2, arg3);
-}
-
-//The total number of operations is 19.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::ifh(bitblock256_t arg1, bitblock256_t arg2, bitblock256_t arg3)
-{
-        return simd256<(2)>::ifh(simd256<1>::ifh(simd256<4>::himask(), arg1, simd256<4>::srli<(2)>(arg1)), arg2, arg3);
-}
-
-//The total number of operations is 8.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::ifh(bitblock256_t arg1, bitblock256_t arg2, bitblock256_t arg3)
-{
-        return simd256<1>::ifh(simd256<8>::gt(simd256<8>::constant<0>(), arg1), arg2, arg3);
-}
-
-//The total number of operations is 8.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::ifh(bitblock256_t arg1, bitblock256_t arg2, bitblock256_t arg3)
-{
-        return simd256<1>::ifh(simd256<16>::gt(simd256<16>::constant<0>(), arg1), arg2, arg3);
-}
-
-//The total number of operations is 8.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::ifh(bitblock256_t arg1, bitblock256_t arg2, bitblock256_t arg3)
-{
-        return simd256<1>::ifh(simd256<32>::gt(simd256<32>::constant<0>(), arg1), arg2, arg3);
-}
-
-//The total number of operations is 1.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::ifh(bitblock256_t arg1, bitblock256_t arg2, bitblock256_t arg3)
-{
-        return (bitblock256_t)_mm256_blendv_pd((__m256d)(arg3), (__m256d)(arg2), (__m256d)(arg1));
-}
-
-//The total number of operations is 12.3333333333
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::ifh(bitblock256_t arg1, bitblock256_t arg2, bitblock256_t arg3)
-{
-        return simd256<(64)>::ifh(simd256<1>::ifh(simd256<128>::himask(), arg1, simd256<128>::srli<(64)>(arg1)), arg2, arg3);
-}
-
-//The total number of operations is 29.8333333333
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::ifh(bitblock256_t arg1, bitblock256_t arg2, bitblock256_t arg3)
-{
-        return simd256<(128)>::ifh(simd256<1>::ifh(simd256<256>::himask(), arg1, simd256<256>::srli<(128)>(arg1)), arg2, arg3);
-}
-
-//The total number of operations is 1.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<1>::sub(bitblock256_t arg1, bitblock256_t arg2)
-{
-        return simd_xor(arg1, arg2);
-}
-
-//The total number of operations is 16.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::sub(bitblock256_t arg1, bitblock256_t arg2)
-{
-        bitblock256_t ans = simd256<(1)>::sub(arg1, arg2);
-        bitblock256_t borrowMask = simd_or(simd_andc(arg2, arg1), simd_and(simd_not(simd_xor(arg1, arg2)), ans));
-        bitblock256_t loMask = simd256<2>::lomask();
-        bitblock256_t borrow = simd256<2>::slli<1>(simd_and(borrowMask, loMask));
-        return simd256<1>::ifh(loMask, ans, simd256<(1)>::sub(ans, borrow));
-}
-
-//The total number of operations is 14.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::sub(bitblock256_t arg1, bitblock256_t arg2)
-{
-        return simd256<1>::ifh(simd256<(8)>::himask(), simd256<(8)>::sub(arg1, simd_and(simd256<(8)>::himask(), arg2)), simd256<(8)>::sub(arg1, arg2));
-}
-
-//The total number of operations is 5.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::sub(bitblock256_t arg1, bitblock256_t arg2)
-{
-        return avx_general_combine256(_mm_sub_epi8(avx_select_hi128(arg1), avx_select_hi128(arg2)), _mm_sub_epi8(avx_select_lo128(arg1), avx_select_lo128(arg2)));
-}
-
-//The total number of operations is 5.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::sub(bitblock256_t arg1, bitblock256_t arg2)
-{
-        return avx_general_combine256(_mm_sub_epi16(avx_select_hi128(arg1), avx_select_hi128(arg2)), _mm_sub_epi16(avx_select_lo128(arg1), avx_select_lo128(arg2)));
-}
-
-//The total number of operations is 5.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::sub(bitblock256_t arg1, bitblock256_t arg2)
-{
-        return avx_general_combine256(_mm_sub_epi32(avx_select_hi128(arg1), avx_select_hi128(arg2)), _mm_sub_epi32(avx_select_lo128(arg1), avx_select_lo128(arg2)));
-}
-
-//The total number of operations is 5.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::sub(bitblock256_t arg1, bitblock256_t arg2)
-{
-        return avx_general_combine256(_mm_sub_epi64(avx_select_hi128(arg1), avx_select_hi128(arg2)), _mm_sub_epi64(avx_select_lo128(arg1), avx_select_lo128(arg2)));
-}
-
-//The total number of operations is 26.3333333333
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::sub(bitblock256_t arg1, bitblock256_t arg2)
-{
-        bitblock256_t partial = simd256<(64)>::sub(arg1, arg2);
-        bitblock256_t borrowMask = simd_or(simd_andc(arg2, arg1), simd_andc(partial, simd_xor(arg1, arg2)));
-        bitblock256_t borrow = simd256<128>::slli<(64)>(simd256<(64)>::srli<(63)>(borrowMask));
-        return simd256<(64)>::sub(partial, borrow);
-}
-
-//The total number of operations is 75.6666666667
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::sub(bitblock256_t arg1, bitblock256_t arg2)
-{
-        bitblock256_t ans = simd256<(128)>::sub(arg1, arg2);
-        bitblock256_t borrowMask = simd_or(simd_andc(arg2, arg1), simd_and(simd_not(simd_xor(arg1, arg2)), ans));
-        bitblock256_t loMask = simd256<256>::lomask();
-        bitblock256_t borrow = simd256<256>::slli<1>(simd_and(borrowMask, loMask));
-        return simd256<1>::ifh(loMask, ans, simd256<(128)>::sub(ans, borrow));
-}
-
-//The total number of operations is 10.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::add_hl(bitblock256_t arg1)
-{
-        return simd256<16>::sub(arg1, simd_and(simd256<2>::lomask(), simd256<16>::srli<1>(arg1)));
-}
-
-//The total number of operations is 11.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::add_hl(bitblock256_t arg1)
-{
-        return simd256<(8)>::add(simd256<4>::srli<(2)>(arg1), simd_and(arg1, simd256<4>::lomask()));
-}
-
-//The total number of operations is 11.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::add_hl(bitblock256_t arg1)
-{
-        return simd256<(16)>::add(simd256<8>::srli<(4)>(arg1), simd_and(arg1, simd256<8>::lomask()));
-}
-
-//The total number of operations is 10.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::add_hl(bitblock256_t arg1)
-{
-        return simd256<(32)>::add(simd256<16>::srli<(8)>(arg1), simd_and(arg1, simd256<16>::lomask()));
-}
-
-//The total number of operations is 10.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::add_hl(bitblock256_t arg1)
-{
-        return simd256<(64)>::add(simd256<32>::srli<(16)>(arg1), simd_and(arg1, simd256<32>::lomask()));
-}
-
-//The total number of operations is 10.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::add_hl(bitblock256_t arg1)
-{
-        return simd256<64>::add(simd256<64>::srli<(32)>(arg1), simd_and(arg1, simd256<64>::lomask()));
-}
-
-//The total number of operations is 35.6666666667
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::add_hl(bitblock256_t arg1)
-{
-        return simd256<128>::add(simd256<128>::srli<(64)>(arg1), simd_and(arg1, simd256<128>::lomask()));
-}
-
-//The total number of operations is 91.1666666667
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::add_hl(bitblock256_t arg1)
-{
-        return simd256<256>::add(simd256<256>::srli<(128)>(arg1), simd_and(arg1, simd256<256>::lomask()));
-}
-
-//The total number of operations is 0
-template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock256_t simd256<1>::constant()
-{
-        return simd256<32>::constant<(-1*val)>();
-}
-
-//The total number of operations is 0
-template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::constant()
-{
-        return simd256<(4)>::constant<((val<<2)|(val&(3)))>();
-}
-
-//The total number of operations is 0
-template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::constant()
-{
-        return simd256<(8)>::constant<((val<<4)|(val&(15)))>();
-}
-
-//The total number of operations is 0
-template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::constant()
-{
-        return (bitblock256_t)_mm256_set1_epi8((int32_t)(val));
-}
-
-//The total number of operations is 0
-template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::constant()
-{
-        return (bitblock256_t)_mm256_set1_epi16((int32_t)(val));
-}
-
-//The total number of operations is 0
-template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::constant()
-{
-        return (bitblock256_t)_mm256_set1_epi32((int32_t)(val));
-}
-
-//The total number of operations is 0
-template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::constant()
-{
-        return ((bitblock256_t)(_mm256_set_epi32((int32_t)((val>>32)), (int32_t)(val), (int32_t)((val>>32)), (int32_t)(val), (int32_t)((val>>32)), (int32_t)(val), (int32_t)((val>>32)), (int32_t)(val))));
-}
-
-//The total number of operations is 0
-template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::constant()
-{
-        return ((bitblock256_t)(_mm256_set_epi32((int32_t)(0), (int32_t)(0), (int32_t)((val>>32)), (int32_t)(val), (int32_t)(0), (int32_t)(0), (int32_t)((val>>32)), (int32_t)(val))));
-}
-
-//The total number of operations is 0
-template <> template <uint64_t val> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::constant()
-{
-        return ((bitblock256_t)(_mm256_set_epi32((int32_t)(0), (int32_t)(0), (int32_t)(0), (int32_t)(0), (int32_t)(0), (int32_t)(0), (int32_t)((val>>32)), (int32_t)(val))));
-}
-
-//The total number of operations is 1.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<1>::min(bitblock256_t arg1, bitblock256_t arg2)
-{
-        return simd_or(arg1, arg2);
-}
-
-//The total number of operations is 25.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::min(bitblock256_t arg1, bitblock256_t arg2)
-{
-        bitblock256_t hiAns = simd256<(1)>::min(arg1, arg2);
-        bitblock256_t loAns = simd256<(1)>::umin(arg1, arg2);
-        bitblock256_t eqMask1 = simd256<2>::srli<(1)>(simd256<(1)>::eq(hiAns, arg1));
-        bitblock256_t eqMask2 = simd256<2>::srli<(1)>(simd256<(1)>::eq(hiAns, arg2));
-        return simd256<1>::ifh(simd256<2>::himask(), hiAns, simd256<1>::ifh(eqMask1, simd256<1>::ifh(eqMask2, loAns, arg1), arg2));
-}
-
-//The total number of operations is 17.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::min(bitblock256_t arg1, bitblock256_t arg2)
-{
-        bitblock256_t high_bit = simd256<4>::constant<(8)>();
-        return simd_xor(simd256<4>::umin(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit)), high_bit);
-}
-
-//The total number of operations is 5.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::min(bitblock256_t arg1, bitblock256_t arg2)
-{
-        return avx_general_combine256(_mm_min_epi8(avx_select_hi128(arg1), avx_select_hi128(arg2)), _mm_min_epi8(avx_select_lo128(arg1), avx_select_lo128(arg2)));
-}
-
-//The total number of operations is 5.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::min(bitblock256_t arg1, bitblock256_t arg2)
-{
-        return avx_general_combine256(_mm_min_epi16(avx_select_hi128(arg1), avx_select_hi128(arg2)), _mm_min_epi16(avx_select_lo128(arg1), avx_select_lo128(arg2)));
-}
-
-//The total number of operations is 5.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::min(bitblock256_t arg1, bitblock256_t arg2)
-{
-        return avx_general_combine256(_mm_min_epi32(avx_select_hi128(arg1), avx_select_hi128(arg2)), _mm_min_epi32(avx_select_lo128(arg1), avx_select_lo128(arg2)));
-}
-
-//The total number of operations is 8.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::min(bitblock256_t arg1, bitblock256_t arg2)
-{
-        return simd256<1>::ifh(simd256<64>::gt(arg1, arg2), arg2, arg1);
-}
-
-//The total number of operations is 54.6666666667
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::min(bitblock256_t arg1, bitblock256_t arg2)
-{
-        bitblock256_t hiAns = simd256<(64)>::min(arg1, arg2);
-        bitblock256_t loAns = simd256<(64)>::umin(arg1, arg2);
-        bitblock256_t eqMask1 = simd256<128>::srli<(64)>(simd256<(64)>::eq(hiAns, arg1));
-        bitblock256_t eqMask2 = simd256<128>::srli<(64)>(simd256<(64)>::eq(hiAns, arg2));
-        return simd256<1>::ifh(simd256<128>::himask(), hiAns, simd256<1>::ifh(eqMask1, simd256<1>::ifh(eqMask2, loAns, arg1), arg2));
-}
-
-//The total number of operations is 186.666666667
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::min(bitblock256_t arg1, bitblock256_t arg2)
-{
-        bitblock256_t hiAns = simd256<(128)>::min(arg1, arg2);
-        bitblock256_t loAns = simd256<(128)>::umin(arg1, arg2);
-        bitblock256_t eqMask1 = simd256<256>::srli<(128)>(simd256<(128)>::eq(hiAns, arg1));
-        bitblock256_t eqMask2 = simd256<256>::srli<(128)>(simd256<(128)>::eq(hiAns, arg2));
-        return simd256<1>::ifh(simd256<256>::himask(), hiAns, simd256<1>::ifh(eqMask1, simd256<1>::ifh(eqMask2, loAns, arg1), arg2));
-}
-
-//The total number of operations is 0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::lomask()
-{
-        return simd256<2>::constant<(1)>();
-}
-
-//The total number of operations is 0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::lomask()
-{
-        return simd256<4>::constant<(3)>();
-}
-
-//The total number of operations is 0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::lomask()
-{
-        return simd256<8>::constant<(15)>();
-}
-
-//The total number of operations is 0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::lomask()
-{
-        return simd256<16>::constant<(255)>();
-}
-
-//The total number of operations is 0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::lomask()
-{
-        return simd256<32>::constant<(65535)>();
-}
-
-//The total number of operations is 0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::lomask()
-{
-        return ((bitblock256_t)(_mm256_set_epi32((int32_t)(0), (int32_t)(-1), (int32_t)(0), (int32_t)(-1), (int32_t)(0), (int32_t)(-1), (int32_t)(0), (int32_t)(-1))));
-}
-
-//The total number of operations is 0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::lomask()
-{
-        return ((bitblock256_t)(_mm256_set_epi32((int32_t)(0), (int32_t)(0), (int32_t)(-1), (int32_t)(-1), (int32_t)(0), (int32_t)(0), (int32_t)(-1), (int32_t)(-1))));
-}
-
-//The total number of operations is 0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::lomask()
-{
-        return ((bitblock256_t)(_mm256_set_epi32((int32_t)(0), (int32_t)(0), (int32_t)(0), (int32_t)(0), (int32_t)(-1), (int32_t)(-1), (int32_t)(-1), (int32_t)(-1))));
-}
-
-//The total number of operations is 1.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<1>::umin(bitblock256_t arg1, bitblock256_t arg2)
-{
-        return simd_and(arg1, arg2);
-}
-
-//The total number of operations is 24.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::umin(bitblock256_t arg1, bitblock256_t arg2)
-{
-        bitblock256_t tmpAns = simd256<(1)>::umin(arg1, arg2);
-        bitblock256_t eqMask1 = simd256<2>::srli<(1)>(simd256<(1)>::eq(tmpAns, arg1));
-        bitblock256_t eqMask2 = simd256<2>::srli<(1)>(simd256<(1)>::eq(tmpAns, arg2));
-        return simd256<1>::ifh(simd256<2>::himask(), tmpAns, simd256<1>::ifh(eqMask1, simd256<1>::ifh(eqMask2, tmpAns, arg1), arg2));
-}
-
-//The total number of operations is 14.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::umin(bitblock256_t arg1, bitblock256_t arg2)
-{
-        return simd_or(simd_and(simd256<(8)>::himask(), simd256<(8)>::umin(arg1, arg2)), simd256<(8)>::umin(simd_and(simd256<(8)>::lomask(), arg1), simd_and(simd256<(8)>::lomask(), arg2)));
-}
-
-//The total number of operations is 5.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::umin(bitblock256_t arg1, bitblock256_t arg2)
-{
-        return avx_general_combine256(_mm_min_epu8(avx_select_hi128(arg1), avx_select_hi128(arg2)), _mm_min_epu8(avx_select_lo128(arg1), avx_select_lo128(arg2)));
-}
-
-//The total number of operations is 5.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::umin(bitblock256_t arg1, bitblock256_t arg2)
-{
-        return avx_general_combine256(_mm_min_epu16(avx_select_hi128(arg1), avx_select_hi128(arg2)), _mm_min_epu16(avx_select_lo128(arg1), avx_select_lo128(arg2)));
-}
-
-//The total number of operations is 5.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::umin(bitblock256_t arg1, bitblock256_t arg2)
-{
-        return avx_general_combine256(_mm_min_epu32(avx_select_hi128(arg1), avx_select_hi128(arg2)), _mm_min_epu32(avx_select_lo128(arg1), avx_select_lo128(arg2)));
-}
-
-//The total number of operations is 11.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::umin(bitblock256_t arg1, bitblock256_t arg2)
-{
-        bitblock256_t high_bit = simd256<64>::constant<(9223372036854775808ULL)>();
-        return simd_xor(simd256<64>::min(simd_xor(arg1, high_bit), simd_xor(arg2, high_bit)), high_bit);
-}
-
-//The total number of operations is 46.6666666667
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::umin(bitblock256_t arg1, bitblock256_t arg2)
-{
-        bitblock256_t tmpAns = simd256<(64)>::umin(arg1, arg2);
-        bitblock256_t eqMask1 = simd256<128>::srli<(64)>(simd256<(64)>::eq(tmpAns, arg1));
-        bitblock256_t eqMask2 = simd256<128>::srli<(64)>(simd256<(64)>::eq(tmpAns, arg2));
-        return simd256<1>::ifh(simd256<128>::himask(), tmpAns, simd256<1>::ifh(eqMask1, simd256<1>::ifh(eqMask2, tmpAns, arg1), arg2));
-}
-
-//The total number of operations is 132.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::umin(bitblock256_t arg1, bitblock256_t arg2)
-{
-        bitblock256_t tmpAns = simd256<(128)>::umin(arg1, arg2);
-        bitblock256_t eqMask1 = simd256<256>::srli<(128)>(simd256<(128)>::eq(tmpAns, arg1));
-        bitblock256_t eqMask2 = simd256<256>::srli<(128)>(simd256<(128)>::eq(tmpAns, arg2));
-        return simd256<1>::ifh(simd256<256>::himask(), tmpAns, simd256<1>::ifh(eqMask1, simd256<1>::ifh(eqMask2, tmpAns, arg1), arg2));
-}
-
-//The total number of operations is 19.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::abs(bitblock256_t arg1)
-{
-        return simd256<1>::ifh(simd256<2>::himask(), simd_and(arg1, simd256<256>::slli<1>(simd_not(arg1))), arg1);
-}
-
-//The total number of operations is 39.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::abs(bitblock256_t arg1)
-{
-        bitblock256_t gtMask = simd256<4>::gt(arg1, simd256<4>::constant<0>());
-        return simd256<1>::ifh(gtMask, arg1, simd256<4>::sub(gtMask, arg1));
-}
-
-//The total number of operations is 4.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::abs(bitblock256_t arg1)
-{
-        return avx_general_combine256(_mm_abs_epi8(avx_select_hi128(arg1)), _mm_abs_epi8(avx_select_lo128(arg1)));
-}
-
-//The total number of operations is 4.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::abs(bitblock256_t arg1)
-{
-        return avx_general_combine256(_mm_abs_epi16(avx_select_hi128(arg1)), _mm_abs_epi16(avx_select_lo128(arg1)));
-}
-
-//The total number of operations is 4.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::abs(bitblock256_t arg1)
-{
-        return avx_general_combine256(_mm_abs_epi32(avx_select_hi128(arg1)), _mm_abs_epi32(avx_select_lo128(arg1)));
-}
-
-//The total number of operations is 13.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::abs(bitblock256_t arg1)
-{
-        bitblock256_t gtMask = simd256<64>::gt(arg1, simd256<64>::constant<0>());
-        return simd256<1>::ifh(gtMask, arg1, simd256<64>::sub(gtMask, arg1));
-}
-
-//The total number of operations is 69.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::abs(bitblock256_t arg1)
-{
-        bitblock256_t eqMask = simd256<128>::eq(simd256<1>::ifh(simd256<128>::himask(), simd256<(64)>::abs(arg1), arg1), arg1);
-        return simd256<1>::ifh(eqMask, arg1, simd256<128>::sub(eqMask, arg1));
-}
-
-//The total number of operations is 204.833333333
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::abs(bitblock256_t arg1)
-{
-        bitblock256_t eqMask = simd256<256>::eq(simd256<1>::ifh(simd256<256>::himask(), simd256<(128)>::abs(arg1), arg1), arg1);
-        return simd256<1>::ifh(eqMask, arg1, simd256<256>::sub(eqMask, arg1));
-}
-
-//The total number of operations is 2.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<1>::eq(bitblock256_t arg1, bitblock256_t arg2)
-{
-        return simd_not(simd_xor(arg1, arg2));
-}
-
-//The total number of operations is 14.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::eq(bitblock256_t arg1, bitblock256_t arg2)
-{
-        bitblock256_t tmpAns = simd256<(1)>::eq(arg1, arg2);
-        bitblock256_t loMask = simd_and(tmpAns, simd256<2>::srli<(1)>(tmpAns));
-        bitblock256_t hiMask = simd256<2>::slli<(1)>(loMask);
-        return simd_or(loMask, hiMask);
-}
-
-//The total number of operations is 17.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::eq(bitblock256_t arg1, bitblock256_t arg2)
-{
-        return simd_or(simd_and(simd256<(8)>::himask(), simd256<(8)>::eq(simd_and(simd256<(8)>::himask(), arg1), simd_and(simd256<(8)>::himask(), arg2))), simd_and(simd256<(8)>::lomask(), simd256<(8)>::eq(simd_and(simd256<(8)>::lomask(), arg1), simd_and(simd256<(8)>::lomask(), arg2))));
-}
-
-//The total number of operations is 5.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::eq(bitblock256_t arg1, bitblock256_t arg2)
-{
-        return avx_general_combine256(_mm_cmpeq_epi8(avx_select_hi128(arg1), avx_select_hi128(arg2)), _mm_cmpeq_epi8(avx_select_lo128(arg1), avx_select_lo128(arg2)));
-}
-
-//The total number of operations is 5.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::eq(bitblock256_t arg1, bitblock256_t arg2)
-{
-        return avx_general_combine256(_mm_cmpeq_epi16(avx_select_hi128(arg1), avx_select_hi128(arg2)), _mm_cmpeq_epi16(avx_select_lo128(arg1), avx_select_lo128(arg2)));
-}
-
-//The total number of operations is 5.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::eq(bitblock256_t arg1, bitblock256_t arg2)
-{
-        return avx_general_combine256(_mm_cmpeq_epi32(avx_select_hi128(arg1), avx_select_hi128(arg2)), _mm_cmpeq_epi32(avx_select_lo128(arg1), avx_select_lo128(arg2)));
-}
-
-//The total number of operations is 5.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::eq(bitblock256_t arg1, bitblock256_t arg2)
-{
-        return avx_general_combine256(_mm_cmpeq_epi64(avx_select_hi128(arg1), avx_select_hi128(arg2)), _mm_cmpeq_epi64(avx_select_lo128(arg1), avx_select_lo128(arg2)));
-}
-
-//The total number of operations is 23.6666666667
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::eq(bitblock256_t arg1, bitblock256_t arg2)
-{
-        bitblock256_t tmpAns = simd256<(64)>::eq(arg1, arg2);
-        bitblock256_t loMask = simd_and(tmpAns, simd256<128>::srli<(64)>(tmpAns));
-        bitblock256_t hiMask = simd256<128>::slli<(64)>(loMask);
-        return simd_or(loMask, hiMask);
-}
-
-//The total number of operations is 54.1666666667
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::eq(bitblock256_t arg1, bitblock256_t arg2)
-{
-        bitblock256_t tmpAns = simd256<(128)>::eq(arg1, arg2);
-        bitblock256_t loMask = simd_and(tmpAns, simd256<256>::srli<(128)>(tmpAns));
-        bitblock256_t hiMask = simd256<256>::slli<(128)>(loMask);
-        return simd_or(loMask, hiMask);
-}
-
-//The total number of operations is 7.0
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::srai(bitblock256_t arg1)
-{
-        return ((sh == 0) ? arg1 : simd_or(simd_and(simd256<2>::himask(), arg1), simd256<2>::srli<1>(arg1)));
-}
-
-//The total number of operations is 17.5
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::srai(bitblock256_t arg1)
-{
-        return simd_or(simd_and(simd256<4>::himask(), simd256<(2)>::srai<((sh < (2)) ? sh : (2))>(arg1)), ((sh <= (2)) ? simd256<4>::srli<sh>(arg1) : simd256<(2)>::srai<(sh-(2))>(simd256<4>::srli<(2)>(arg1))));
-}
-
-//The total number of operations is 12.0
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::srai(bitblock256_t arg1)
-{
-        bitblock256_t tmp = simd256<8>::srli<((sh >= 8) ? (7) : ((sh < 0) ? 0 : sh))>(arg1);
-        return simd_or(tmp, simd256<8>::sub(simd256<8>::constant<0>(), simd_and(simd256<8>::constant<(1<<((8-((sh >= 8) ? (7) : ((sh < 0) ? 0 : sh)))-1))>(), tmp)));
-}
-
-//The total number of operations is 4.0
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::srai(bitblock256_t arg1)
-{
-        return avx_general_combine256(_mm_srai_epi16(avx_select_hi128(arg1), (int32_t)(sh)), _mm_srai_epi16(avx_select_lo128(arg1), (int32_t)(sh)));
-}
-
-//The total number of operations is 4.0
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::srai(bitblock256_t arg1)
-{
-        return avx_general_combine256(_mm_srai_epi32(avx_select_hi128(arg1), (int32_t)(sh)), _mm_srai_epi32(avx_select_lo128(arg1), (int32_t)(sh)));
-}
-
-//The total number of operations is 12.0
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::srai(bitblock256_t arg1)
-{
-        return simd_or(simd_and(simd256<64>::himask(), simd256<(32)>::srai<((sh < (32)) ? sh : (32))>(arg1)), ((sh <= (32)) ? simd256<64>::srli<sh>(arg1) : simd256<(32)>::srai<(sh-(32))>(simd256<64>::srli<(32)>(arg1))));
-}
-
-//The total number of operations is 28.3333333333
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::srai(bitblock256_t arg1)
-{
-        return simd_or(simd_and(simd256<128>::himask(), simd256<(64)>::srai<((sh < (64)) ? sh : (64))>(arg1)), ((sh <= (64)) ? simd256<128>::srli<sh>(arg1) : simd256<(64)>::srai<(sh-(64))>(simd256<128>::srli<(64)>(arg1))));
-}
-
-//The total number of operations is 59.0
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::srai(bitblock256_t arg1)
-{
-        return simd_or(simd_and(simd256<256>::himask(), simd256<(128)>::srai<((sh < (128)) ? sh : (128))>(arg1)), ((sh <= (128)) ? simd256<256>::srli<sh>(arg1) : simd256<(128)>::srai<(sh-(128))>(simd256<256>::srli<(128)>(arg1))));
-}
-
-//The total number of operations is 0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::himask()
-{
-        return simd256<2>::constant<(2)>();
-}
-
-//The total number of operations is 0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::himask()
-{
-        return simd256<4>::constant<(12)>();
-}
-
-//The total number of operations is 0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::himask()
-{
-        return simd256<8>::constant<(240)>();
-}
-
-//The total number of operations is 0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::himask()
-{
-        return simd256<16>::constant<(65280)>();
-}
-
-//The total number of operations is 0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::himask()
-{
-        return simd256<32>::constant<-65536>();
-}
-
-//The total number of operations is 0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::himask()
-{
-        return ((bitblock256_t)(_mm256_set_epi32((int32_t)(-1), (int32_t)(0), (int32_t)(-1), (int32_t)(0), (int32_t)(-1), (int32_t)(0), (int32_t)(-1), (int32_t)(0))));
-}
-
-//The total number of operations is 0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::himask()
-{
-        return ((bitblock256_t)(_mm256_set_epi32((int32_t)(-1), (int32_t)(-1), (int32_t)(0), (int32_t)(0), (int32_t)(-1), (int32_t)(-1), (int32_t)(0), (int32_t)(0))));
-}
-
-//The total number of operations is 0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::himask()
-{
-        return ((bitblock256_t)(_mm256_set_epi32((int32_t)(-1), (int32_t)(-1), (int32_t)(-1), (int32_t)(-1), (int32_t)(0), (int32_t)(0), (int32_t)(0), (int32_t)(0))));
-}
-
-//The total number of operations is 1.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<1>::add(bitblock256_t arg1, bitblock256_t arg2)
-{
-        return simd_xor(arg1, arg2);
-}
-
-//The total number of operations is 16.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<2>::add(bitblock256_t arg1, bitblock256_t arg2)
-{
-        bitblock256_t ans = simd256<(1)>::add(arg1, arg2);
-        bitblock256_t carryMask = simd_or(simd_and(arg1, arg2), simd_and(simd_xor(arg1, arg2), simd_not(ans)));
-        bitblock256_t loMask = simd256<2>::lomask();
-        bitblock256_t carry = simd256<2>::slli<1>(simd_and(carryMask, loMask));
-        return simd256<1>::ifh(loMask, ans, simd256<(1)>::add(ans, carry));
-}
-
-//The total number of operations is 14.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<4>::add(bitblock256_t arg1, bitblock256_t arg2)
-{
-        return simd256<1>::ifh(simd256<(8)>::himask(), simd256<(8)>::add(arg1, simd_and(simd256<(8)>::himask(), arg2)), simd256<(8)>::add(arg1, arg2));
-}
-
-//The total number of operations is 5.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<8>::add(bitblock256_t arg1, bitblock256_t arg2)
-{
-        return avx_general_combine256(_mm_add_epi8(avx_select_hi128(arg1), avx_select_hi128(arg2)), _mm_add_epi8(avx_select_lo128(arg1), avx_select_lo128(arg2)));
-}
-
-//The total number of operations is 5.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<16>::add(bitblock256_t arg1, bitblock256_t arg2)
-{
-        return avx_general_combine256(_mm_add_epi16(avx_select_hi128(arg1), avx_select_hi128(arg2)), _mm_add_epi16(avx_select_lo128(arg1), avx_select_lo128(arg2)));
-}
-
-//The total number of operations is 5.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<32>::add(bitblock256_t arg1, bitblock256_t arg2)
-{
-        return avx_general_combine256(_mm_add_epi32(avx_select_hi128(arg1), avx_select_hi128(arg2)), _mm_add_epi32(avx_select_lo128(arg1), avx_select_lo128(arg2)));
-}
-
-//The total number of operations is 5.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<64>::add(bitblock256_t arg1, bitblock256_t arg2)
-{
-        return avx_general_combine256(_mm_add_epi64(avx_select_hi128(arg1), avx_select_hi128(arg2)), _mm_add_epi64(avx_select_lo128(arg1), avx_select_lo128(arg2)));
-}
-
-//The total number of operations is 26.3333333333
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<128>::add(bitblock256_t arg1, bitblock256_t arg2)
-{
-        bitblock256_t partial = simd256<(64)>::add(arg1, arg2);
-        bitblock256_t carryMask = simd_or(simd_and(arg1, arg2), simd_andc(simd_xor(arg1, arg2), partial));
-        bitblock256_t carry = simd256<128>::slli<(64)>(simd256<(64)>::srli<(63)>(carryMask));
-        return simd256<(64)>::add(partial, carry);
-}
-
-//The total number of operations is 75.6666666667
-template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::add(bitblock256_t arg1, bitblock256_t arg2)
-{
-        bitblock256_t ans = simd256<(128)>::add(arg1, arg2);
-        bitblock256_t carryMask = simd_or(simd_and(arg1, arg2), simd_and(simd_xor(arg1, arg2), simd_not(ans)));
-        bitblock256_t loMask = simd256<256>::lomask();
-        bitblock256_t carry = simd256<256>::slli<1>(simd_and(carryMask, loMask));
-        return simd256<1>::ifh(loMask, ans, simd256<(128)>::add(ans, carry));
-}
-
 //The total number of operations is 1.0
 template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<1>::umax(bitblock256_t arg1, bitblock256_t arg2)
@@ -2296 +2414 @@
 
 //The total number of operations is 3.0
-template <> IDISA_ALWAYS_INLINE uint64_t hsimd256<8>::signmask(bitblock256_t arg1)
+template <> IDISA_ALWAYS_INLINE FieldType<256/8>::T hsimd256<8>::signmask(bitblock256_t arg1)
 {
         return ((((uint64_t)(_mm_movemask_epi8(((__m128i)(avx_select_hi128(arg1))))))<<16)|((uint64_t)(_mm_movemask_epi8(((__m128i)(avx_select_lo128(arg1)))))));
@@ -2302 +2420 @@
 
 //The total number of operations is 8.0
-template <> IDISA_ALWAYS_INLINE uint64_t hsimd256<16>::signmask(bitblock256_t arg1)
+template <> IDISA_ALWAYS_INLINE FieldType<256/16>::T hsimd256<16>::signmask(bitblock256_t arg1)
 {
         return hsimd256<(8)>::signmask(hsimd256<16>::packss(simd256<16>::constant<0>(), arg1));
@@ -2308 +2426 @@
 
 //The total number of operations is 13.0
-template <> IDISA_ALWAYS_INLINE uint64_t hsimd256<32>::signmask(bitblock256_t arg1)
+template <> IDISA_ALWAYS_INLINE FieldType<256/32>::T hsimd256<32>::signmask(bitblock256_t arg1)
 {
         return hsimd256<(16)>::signmask(hsimd256<32>::packss(simd256<32>::constant<0>(), arg1));
@@ -2314 +2432 @@
 
 //The total number of operations is 104.0
-template <> IDISA_ALWAYS_INLINE uint64_t hsimd256<64>::signmask(bitblock256_t arg1)
+template <> IDISA_ALWAYS_INLINE FieldType<256/64>::T hsimd256<64>::signmask(bitblock256_t arg1)
 {
         return hsimd256<(32)>::signmask(hsimd256<64>::packh(simd256<64>::constant<0>(), arg1));
@@ -2320 +2438 @@
 
 //The total number of operations is 248.666666667
-template <> IDISA_ALWAYS_INLINE uint64_t hsimd256<128>::signmask(bitblock256_t arg1)
+template <> IDISA_ALWAYS_INLINE FieldType<256/128>::T hsimd256<128>::signmask(bitblock256_t arg1)
 {
         return hsimd256<(64)>::signmask(hsimd256<128>::packh(simd256<128>::constant<0>(), arg1));
@@ -2326 +2444 @@
 
 //The total number of operations is 266.166666667
-template <> IDISA_ALWAYS_INLINE uint64_t hsimd256<256>::signmask(bitblock256_t arg1)
+template <> IDISA_ALWAYS_INLINE FieldType<256/256>::T hsimd256<256>::signmask(bitblock256_t arg1)
 {
         return hsimd256<(128)>::signmask(hsimd256<256>::packh(simd256<256>::constant<0>(), arg1));
@@ -2849 +2967 @@
 
 //The total number of operations is 29.5
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<2>::dsrli(bitblock256_t arg1, bitblock256_t arg2)
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<2>::dsrli(bitblock256_t arg1, bitblock256_t arg2)
 {
         return simd_or(mvmd256<2>::srli<sh>(arg1), mvmd256<2>::slli<((128)-sh)>(arg2));
@@ -2855 +2973 @@
 
 //The total number of operations is 29.5
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<4>::dsrli(bitblock256_t arg1, bitblock256_t arg2)
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<4>::dsrli(bitblock256_t arg1, bitblock256_t arg2)
 {
         return simd_or(mvmd256<4>::srli<sh>(arg1), mvmd256<4>::slli<((64)-sh)>(arg2));
@@ -2861 +2979 @@
 
 //The total number of operations is 29.5
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<8>::dsrli(bitblock256_t arg1, bitblock256_t arg2)
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<8>::dsrli(bitblock256_t arg1, bitblock256_t arg2)
 {
         return simd_or(mvmd256<8>::srli<sh>(arg1), mvmd256<8>::slli<((32)-sh)>(arg2));
@@ -2867 +2985 @@
 
 //The total number of operations is 29.5
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<16>::dsrli(bitblock256_t arg1, bitblock256_t arg2)
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<16>::dsrli(bitblock256_t arg1, bitblock256_t arg2)
 {
         return simd_or(mvmd256<16>::srli<sh>(arg1), mvmd256<16>::slli<((16)-sh)>(arg2));
@@ -2873 +2991 @@
 
 //The total number of operations is 29.5
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<32>::dsrli(bitblock256_t arg1, bitblock256_t arg2)
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<32>::dsrli(bitblock256_t arg1, bitblock256_t arg2)
 {
         return simd_or(mvmd256<32>::srli<sh>(arg1), mvmd256<32>::slli<((8)-sh)>(arg2));
@@ -2879 +2997 @@
 
 //The total number of operations is 29.5
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<64>::dsrli(bitblock256_t arg1, bitblock256_t arg2)
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<64>::dsrli(bitblock256_t arg1, bitblock256_t arg2)
 {
         return simd_or(mvmd256<64>::srli<sh>(arg1), mvmd256<64>::slli<((4)-sh)>(arg2));
@@ -2885 +3003 @@
 
 //The total number of operations is 29.5
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<128>::dsrli(bitblock256_t arg1, bitblock256_t arg2)
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<128>::dsrli(bitblock256_t arg1, bitblock256_t arg2)
 {
         return simd_or(mvmd256<128>::srli<sh>(arg1), mvmd256<128>::slli<((2)-sh)>(arg2));
@@ -2891 +3009 @@
 
 //The total number of operations is 29.5
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<256>::dsrli(bitblock256_t arg1, bitblock256_t arg2)
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<256>::dsrli(bitblock256_t arg1, bitblock256_t arg2)
 {
         return simd_or(mvmd256<256>::srli<sh>(arg1), mvmd256<256>::slli<((1)-sh)>(arg2));
@@ -2897 +3015 @@
 
 //The total number of operations is 1.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<1>::fill(uint64_t val1)
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<1>::fill(FieldType<1>::T val1)
 {
         return mvmd256<32>::fill((-1*val1));
@@ -2903 +3021 @@
 
 //The total number of operations is 1.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<2>::fill(uint64_t val1)
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<2>::fill(FieldType<2>::T val1)
 {
         return mvmd256<(4)>::fill(((val1<<2)|val1));
@@ -2909 +3027 @@
 
 //The total number of operations is 1.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<4>::fill(uint64_t val1)
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<4>::fill(FieldType<4>::T val1)
 {
         return mvmd256<(8)>::fill(((val1<<4)|val1));
@@ -2915 +3033 @@
 
 //The total number of operations is 1.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<8>::fill(uint64_t val1)
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<8>::fill(FieldType<8>::T val1)
 {
         return (bitblock256_t)_mm256_set1_epi8((int32_t)(val1));
@@ -2921 +3039 @@
 
 //The total number of operations is 1.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<16>::fill(uint64_t val1)
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<16>::fill(FieldType<16>::T val1)
 {
         return (bitblock256_t)_mm256_set1_epi16((int32_t)(val1));
@@ -2927 +3045 @@
 
 //The total number of operations is 1.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<32>::fill(uint64_t val1)
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<32>::fill(FieldType<32>::T val1)
 {
         return (bitblock256_t)_mm256_set1_epi32((int32_t)(val1));
@@ -2933 +3051 @@
 
 //The total number of operations is 5.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<64>::fill(uint64_t val1)
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<64>::fill(FieldType<64>::T val1)
 {
         return mvmd256<(32)>::fill2((val1>>(32)), (val1&((4294967296ULL)-1)));
@@ -2939 +3057 @@
 
 //The total number of operations is 13.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<128>::fill(uint64_t val1)
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<128>::fill(FieldType<128>::T val1)
 {
         return mvmd256<(64)>::fill2(0, val1);
@@ -2945 +3063 @@
 
 //The total number of operations is 29.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<256>::fill(uint64_t val1)
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<256>::fill(FieldType<256>::T val1)
 {
         return mvmd256<(128)>::fill2(0, val1);
@@ -2951 +3069 @@
 
 //The total number of operations is 1.5
-template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd256<1>::extract(bitblock256_t arg1)
+template <> template <uint8_t pos> IDISA_ALWAYS_INLINE FieldType<1>::T mvmd256<1>::extract(bitblock256_t arg1)
 {
         return (((pos%2) == 0) ? (mvmd256<(2)>::extract<(pos/2)>(arg1)&(1)) : (mvmd256<(2)>::extract<(pos/2)>(arg1)>>1));
@@ -2957 +3075 @@
 
 //The total number of operations is 1.5
-template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd256<2>::extract(bitblock256_t arg1)
+template <> template <uint8_t pos> IDISA_ALWAYS_INLINE FieldType<2>::T mvmd256<2>::extract(bitblock256_t arg1)
 {
         return (((pos%2) == 0) ? (mvmd256<(4)>::extract<(pos/2)>(arg1)&(3)) : (mvmd256<(4)>::extract<(pos/2)>(arg1)>>2));
@@ -2963 +3081 @@
 
 //The total number of operations is 1.5
-template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd256<4>::extract(bitblock256_t arg1)
+template <> template <uint8_t pos> IDISA_ALWAYS_INLINE FieldType<4>::T mvmd256<4>::extract(bitblock256_t arg1)
 {
         return (((pos%2) == 0) ? (mvmd256<(8)>::extract<(pos/2)>(arg1)&(15)) : (mvmd256<(8)>::extract<(pos/2)>(arg1)>>4));
@@ -2969 +3087 @@
 
 //The total number of operations is 1.5
-template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd256<8>::extract(bitblock256_t arg1)
+template <> template <uint8_t pos> IDISA_ALWAYS_INLINE FieldType<8>::T mvmd256<8>::extract(bitblock256_t arg1)
 {
         return (((pos%2) == 0) ? (mvmd256<(16)>::extract<(pos/2)>(arg1)&(255)) : (mvmd256<(16)>::extract<(pos/2)>(arg1)>>8));
@@ -2975 +3093 @@
 
 //The total number of operations is 1.5
-template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd256<16>::extract(bitblock256_t arg1)
+template <> template <uint8_t pos> IDISA_ALWAYS_INLINE FieldType<16>::T mvmd256<16>::extract(bitblock256_t arg1)
 {
         return ((pos < 8) ? (65535&_mm_extract_epi16(avx_select_lo128(arg1), (int32_t)(pos))) : (65535&_mm_extract_epi16(avx_select_hi128(arg1), (int32_t)((pos-8)))));
@@ -2981 +3099 @@
 
 //The total number of operations is 1.5
-template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd256<32>::extract(bitblock256_t arg1)
+template <> template <uint8_t pos> IDISA_ALWAYS_INLINE FieldType<32>::T mvmd256<32>::extract(bitblock256_t arg1)
 {
         return ((pos < 4) ? (((uint64_t)(((4294967296ULL)-1)))&_mm_extract_epi32(avx_select_lo128(arg1), (int32_t)(pos))) : (((uint64_t)(((4294967296ULL)-1)))&_mm_extract_epi32(avx_select_hi128(arg1), (int32_t)((pos-4)))));
@@ -2987 +3105 @@
 
 //The total number of operations is 3.0
-template <> template <uint64_t pos> IDISA_ALWAYS_INLINE uint64_t mvmd256<64>::extract(bitblock256_t arg1)
-{
-        return ((mvmd256<(32)>::extract<((2*pos)+1)>(arg1)<<(32))|mvmd256<(32)>::extract<(2*pos)>(arg1));
+template <> template <uint8_t pos> IDISA_ALWAYS_INLINE FieldType<64>::T mvmd256<64>::extract(bitblock256_t arg1)
+{
+        return ((((uint64_t)(mvmd256<(32)>::extract<((2*pos)+1)>(arg1)))<<(32))|mvmd256<(32)>::extract<(2*pos)>(arg1));
 }
 
 //The total number of operations is 23.5
-template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<1>::splat(bitblock256_t arg1)
-{
-        bitblock256_t tmpArg = (((pos%2) == 0) ? simd256<(2)>::slli<1>(arg1) : simd256<(2)>::srli<1>(arg1));
-        bitblock256_t arg11 = (((pos%2) == 0) ? simd_and(simd256<(2)>::lomask(), arg1) : simd_and(simd256<(2)>::himask(), arg1));
-        return mvmd256<(2)>::splat<(pos/2)>(simd_or(tmpArg, arg11));
+template <> template <uint16_t pos> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<1>::splat(bitblock256_t arg1)
+{
+        return mvmd256<(2)>::splat<(pos/2)>(simd_or((((pos%2) == 0) ? simd256<(2)>::slli<1>(arg1) : simd256<(2)>::srli<1>(arg1)), (((pos%2) == 0) ? simd_and(simd256<(2)>::lomask(), arg1) : simd_and(simd256<(2)>::himask(), arg1))));
 }
 
 //The total number of operations is 16.5
-template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<2>::splat(bitblock256_t arg1)
-{
-        bitblock256_t tmpArg = (((pos%2) == 0) ? simd256<(4)>::slli<2>(arg1) : simd256<(4)>::srli<2>(arg1));
-        bitblock256_t arg11 = (((pos%2) == 0) ? simd_and(simd256<(4)>::lomask(), arg1) : simd_and(simd256<(4)>::himask(), arg1));
-        return mvmd256<(4)>::splat<(pos/2)>(simd_or(tmpArg, arg11));
+template <> template <uint16_t pos> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<2>::splat(bitblock256_t arg1)
+{
+        return mvmd256<(4)>::splat<(pos/2)>(simd_or((((pos%2) == 0) ? simd256<(4)>::slli<2>(arg1) : simd256<(4)>::srli<2>(arg1)), (((pos%2) == 0) ? simd_and(simd256<(4)>::lomask(), arg1) : simd_and(simd256<(4)>::himask(), arg1))));
 }
 
 //The total number of operations is 9.5
-template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<4>::splat(bitblock256_t arg1)
-{
-        bitblock256_t tmpArg = (((pos%2) == 0) ? simd256<(8)>::slli<4>(arg1) : simd256<(8)>::srli<4>(arg1));
-        bitblock256_t arg11 = (((pos%2) == 0) ? simd_and(simd256<(8)>::lomask(), arg1) : simd_and(simd256<(8)>::himask(), arg1));
-        return mvmd256<(8)>::splat<(pos/2)>(simd_or(tmpArg, arg11));
+template <> template <uint16_t pos> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<4>::splat(bitblock256_t arg1)
+{
+        return mvmd256<(8)>::splat<(pos/2)>(simd_or((((pos%2) == 0) ? simd256<(8)>::slli<4>(arg1) : simd256<(8)>::srli<4>(arg1)), (((pos%2) == 0) ? simd_and(simd256<(8)>::lomask(), arg1) : simd_and(simd256<(8)>::himask(), arg1))));
 }
 
 //The total number of operations is 2.5
-template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<8>::splat(bitblock256_t arg1)
+template <> template <uint16_t pos> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<8>::splat(bitblock256_t arg1)
 {
         return ((pos < 16) ? mvmd256<8>::fill(_mm_extract_epi8(avx_select_lo128(arg1), (int32_t)(pos))) : mvmd256<8>::fill(_mm_extract_epi8(avx_select_hi128(arg1), (int32_t)((pos-16)))));
@@ -3023 +3135 @@
 
 //The total number of operations is 2.5
-template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<16>::splat(bitblock256_t arg1)
+template <> template <uint16_t pos> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<16>::splat(bitblock256_t arg1)
 {
         return ((pos < 8) ? mvmd256<16>::fill(_mm_extract_epi16(avx_select_lo128(arg1), (int32_t)(pos))) : mvmd256<16>::fill(_mm_extract_epi16(avx_select_hi128(arg1), (int32_t)((pos-8)))));
@@ -3029 +3141 @@
 
 //The total number of operations is 2.5
-template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<32>::splat(bitblock256_t arg1)
+template <> template <uint16_t pos> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<32>::splat(bitblock256_t arg1)
 {
         return ((pos < 4) ? mvmd256<32>::fill(_mm_extract_epi32(avx_select_lo128(arg1), (int32_t)(pos))) : mvmd256<32>::fill(_mm_extract_epi32(avx_select_hi128(arg1), (int32_t)((pos-4)))));
@@ -3035 +3147 @@
 
 //The total number of operations is 8.0
-template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<64>::splat(bitblock256_t arg1)
+template <> template <uint16_t pos> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<64>::splat(bitblock256_t arg1)
 {
         return simd256<1>::ifh(simd256<64>::himask(), mvmd256<(32)>::splat<((2*pos)+1)>(arg1), mvmd256<(32)>::splat<(2*pos)>(arg1));
@@ -3041 +3153 @@
 
 //The total number of operations is 19.0
-template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<128>::splat(bitblock256_t arg1)
+template <> template <uint16_t pos> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<128>::splat(bitblock256_t arg1)
 {
         return simd256<1>::ifh(simd256<128>::himask(), mvmd256<(64)>::splat<((2*pos)+1)>(arg1), mvmd256<(64)>::splat<(2*pos)>(arg1));
@@ -3047 +3159 @@
 
 //The total number of operations is 41.0
-template <> template <uint64_t pos> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<256>::splat(bitblock256_t arg1)
+template <> template <uint16_t pos> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<256>::splat(bitblock256_t arg1)
 {
         return simd256<1>::ifh(simd256<256>::himask(), mvmd256<(128)>::splat<((2*pos)+1)>(arg1), mvmd256<(128)>::splat<(2*pos)>(arg1));
@@ -3053 +3165 @@
 
 //The total number of operations is 15.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<1>::fill16(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8, uint64_t val9, uint64_t val10, uint64_t val11, uint64_t val12, uint64_t val13, uint64_t val14, uint64_t val15, uint64_t val16)
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<1>::fill16(FieldType<1>::T val1, FieldType<1>::T val2, FieldType<1>::T val3, FieldType<1>::T val4, FieldType<1>::T val5, FieldType<1>::T val6, FieldType<1>::T val7, FieldType<1>::T val8, FieldType<1>::T val9, FieldType<1>::T val10, FieldType<1>::T val11, FieldType<1>::T val12, FieldType<1>::T val13, FieldType<1>::T val14, FieldType<1>::T val15, FieldType<1>::T val16)
 {
         return simd_or(mvmd256<(2)>::fill16((val1<<1), (val3<<1), (val5<<1), (val7<<1), (val9<<1), (val11<<1), (val13<<1), (val15<<1), (val1<<1), (val3<<1), (val5<<1), (val7<<1), (val9<<1), (val11<<1), (val13<<1), (val15<<1)), mvmd256<(2)>::fill16((val2&(1)), (val4&(1)), (val6&(1)), (val8&(1)), (val10&(1)), (val12&(1)), (val14&(1)), (val16&(1)), (val2&(1)), (val4&(1)), (val6&(1)), (val8&(1)), (val10&(1)), (val12&(1)), (val14&(1)), (val16&(1))));
@@ -3059 +3171 @@
 
 //The total number of operations is 7.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<2>::fill16(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8, uint64_t val9, uint64_t val10, uint64_t val11, uint64_t val12, uint64_t val13, uint64_t val14, uint64_t val15, uint64_t val16)
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<2>::fill16(FieldType<2>::T val1, FieldType<2>::T val2, FieldType<2>::T val3, FieldType<2>::T val4, FieldType<2>::T val5, FieldType<2>::T val6, FieldType<2>::T val7, FieldType<2>::T val8, FieldType<2>::T val9, FieldType<2>::T val10, FieldType<2>::T val11, FieldType<2>::T val12, FieldType<2>::T val13, FieldType<2>::T val14, FieldType<2>::T val15, FieldType<2>::T val16)
 {
         return simd_or(mvmd256<(4)>::fill16((val1<<2), (val3<<2), (val5<<2), (val7<<2), (val9<<2), (val11<<2), (val13<<2), (val15<<2), (val1<<2), (val3<<2), (val5<<2), (val7<<2), (val9<<2), (val11<<2), (val13<<2), (val15<<2)), mvmd256<(4)>::fill16((val2&(3)), (val4&(3)), (val6&(3)), (val8&(3)), (val10&(3)), (val12&(3)), (val14&(3)), (val16&(3)), (val2&(3)), (val4&(3)), (val6&(3)), (val8&(3)), (val10&(3)), (val12&(3)), (val14&(3)), (val16&(3))));
@@ -3065 +3177 @@
 
 //The total number of operations is 3.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<4>::fill16(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8, uint64_t val9, uint64_t val10, uint64_t val11, uint64_t val12, uint64_t val13, uint64_t val14, uint64_t val15, uint64_t val16)
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<4>::fill16(FieldType<4>::T val1, FieldType<4>::T val2, FieldType<4>::T val3, FieldType<4>::T val4, FieldType<4>::T val5, FieldType<4>::T val6, FieldType<4>::T val7, FieldType<4>::T val8, FieldType<4>::T val9, FieldType<4>::T val10, FieldType<4>::T val11, FieldType<4>::T val12, FieldType<4>::T val13, FieldType<4>::T val14, FieldType<4>::T val15, FieldType<4>::T val16)
 {
         return simd_or(mvmd256<(8)>::fill16((val1<<4), (val3<<4), (val5<<4), (val7<<4), (val9<<4), (val11<<4), (val13<<4), (val15<<4), (val1<<4), (val3<<4), (val5<<4), (val7<<4), (val9<<4), (val11<<4), (val13<<4), (val15<<4)), mvmd256<(8)>::fill16((val2&(15)), (val4&(15)), (val6&(15)), (val8&(15)), (val10&(15)), (val12&(15)), (val14&(15)), (val16&(15)), (val2&(15)), (val4&(15)), (val6&(15)), (val8&(15)), (val10&(15)), (val12&(15)), (val14&(15)), (val16&(15))));
@@ -3071 +3183 @@
 
 //The total number of operations is 1.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<8>::fill16(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8, uint64_t val9, uint64_t val10, uint64_t val11, uint64_t val12, uint64_t val13, uint64_t val14, uint64_t val15, uint64_t val16)
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<8>::fill16(FieldType<8>::T val1, FieldType<8>::T val2, FieldType<8>::T val3, FieldType<8>::T val4, FieldType<8>::T val5, FieldType<8>::T val6, FieldType<8>::T val7, FieldType<8>::T val8, FieldType<8>::T val9, FieldType<8>::T val10, FieldType<8>::T val11, FieldType<8>::T val12, FieldType<8>::T val13, FieldType<8>::T val14, FieldType<8>::T val15, FieldType<8>::T val16)
 {
         return (bitblock256_t)_mm256_set_epi8((int32_t)(val1), (int32_t)(val2), (int32_t)(val3), (int32_t)(val4), (int32_t)(val5), (int32_t)(val6), (int32_t)(val7), (int32_t)(val8), (int32_t)(val9), (int32_t)(val10), (int32_t)(val11), (int32_t)(val12), (int32_t)(val13), (int32_t)(val14), (int32_t)(val15), (int32_t)(val16), (int32_t)(val1), (int32_t)(val2), (int32_t)(val3), (int32_t)(val4), (int32_t)(val5), (int32_t)(val6), (int32_t)(val7), (int32_t)(val8), (int32_t)(val9), (int32_t)(val10), (int32_t)(val11), (int32_t)(val12), (int32_t)(val13), (int32_t)(val14), (int32_t)(val15), (int32_t)(val16));
@@ -3077 +3189 @@
 
 //The total number of operations is 5.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<16>::fill16(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8, uint64_t val9, uint64_t val10, uint64_t val11, uint64_t val12, uint64_t val13, uint64_t val14, uint64_t val15, uint64_t val16)
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<16>::fill16(FieldType<16>::T val1, FieldType<16>::T val2, FieldType<16>::T val3, FieldType<16>::T val4, FieldType<16>::T val5, FieldType<16>::T val6, FieldType<16>::T val7, FieldType<16>::T val8, FieldType<16>::T val9, FieldType<16>::T val10, FieldType<16>::T val11, FieldType<16>::T val12, FieldType<16>::T val13, FieldType<16>::T val14, FieldType<16>::T val15, FieldType<16>::T val16)
 {
         return simd256<1>::ifh(simd256<(256)>::himask(), mvmd256<16>::fill8(val1, val2, val3, val4, val5, val6, val7, val8), mvmd256<16>::fill8(val9, val10, val11, val12, val13, val14, val15, val16));
@@ -3083 +3195 @@
 
 //The total number of operations is 5.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<1>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4)
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<1>::fill4(FieldType<1>::T val1, FieldType<1>::T val2, FieldType<1>::T val3, FieldType<1>::T val4)
 {
         return simd256<1>::ifh(simd256<(4)>::himask(), mvmd256<1>::fill2(val1, val2), mvmd256<1>::fill2(val3, val4));
@@ -3089 +3201 @@
 
 //The total number of operations is 5.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<2>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4)
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<2>::fill4(FieldType<2>::T val1, FieldType<2>::T val2, FieldType<2>::T val3, FieldType<2>::T val4)
 {
         return simd256<1>::ifh(simd256<(8)>::himask(), mvmd256<2>::fill2(val1, val2), mvmd256<2>::fill2(val3, val4));
@@ -3095 +3207 @@
 
 //The total number of operations is 5.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<4>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4)
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<4>::fill4(FieldType<4>::T val1, FieldType<4>::T val2, FieldType<4>::T val3, FieldType<4>::T val4)
 {
         return simd256<1>::ifh(simd256<(16)>::himask(), mvmd256<4>::fill2(val1, val2), mvmd256<4>::fill2(val3, val4));
@@ -3101 +3213 @@
 
 //The total number of operations is 5.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<8>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4)
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<8>::fill4(FieldType<8>::T val1, FieldType<8>::T val2, FieldType<8>::T val3, FieldType<8>::T val4)
 {
         return simd256<1>::ifh(simd256<(32)>::himask(), mvmd256<8>::fill2(val1, val2), mvmd256<8>::fill2(val3, val4));
@@ -3107 +3219 @@
 
 //The total number of operations is 3.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<16>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4)
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<16>::fill4(FieldType<16>::T val1, FieldType<16>::T val2, FieldType<16>::T val3, FieldType<16>::T val4)
 {
         return simd_or(mvmd256<(32)>::fill4((val1<<16), (val3<<16), (val1<<16), (val3<<16)), mvmd256<(32)>::fill4((val2&(65535)), (val4&(65535)), (val2&(65535)), (val4&(65535))));
@@ -3113 +3225 @@
 
 //The total number of operations is 1.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<32>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4)
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<32>::fill4(FieldType<32>::T val1, FieldType<32>::T val2, FieldType<32>::T val3, FieldType<32>::T val4)
 {
         return (bitblock256_t)_mm256_set_epi32((int32_t)(val1), (int32_t)(val2), (int32_t)(val3), (int32_t)(val4), (int32_t)(val1), (int32_t)(val2), (int32_t)(val3), (int32_t)(val4));
@@ -3119 +3231 @@
 
 //The total number of operations is 29.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<64>::fill4(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4)
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<64>::fill4(FieldType<64>::T val1, FieldType<64>::T val2, FieldType<64>::T val3, FieldType<64>::T val4)
 {
         return simd256<1>::ifh(simd256<(256)>::himask(), mvmd256<64>::fill2(val1, val2), mvmd256<64>::fill2(val3, val4));
@@ -3125 +3237 @@
 
 //The total number of operations is 14.5
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<2>::srli(bitblock256_t arg1)
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<2>::srli(bitblock256_t arg1)
 {
         return simd256<256>::srli<(sh*2)>(arg1);
@@ -3131 +3243 @@
 
 //The total number of operations is 14.5
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<4>::srli(bitblock256_t arg1)
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<4>::srli(bitblock256_t arg1)
 {
         return simd256<256>::srli<(sh*4)>(arg1);
@@ -3137 +3249 @@
 
 //The total number of operations is 14.5
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<8>::srli(bitblock256_t arg1)
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<8>::srli(bitblock256_t arg1)
 {
         return simd256<256>::srli<(sh*8)>(arg1);
@@ -3143 +3255 @@
 
 //The total number of operations is 14.5
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<16>::srli(bitblock256_t arg1)
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<16>::srli(bitblock256_t arg1)
 {
         return simd256<256>::srli<(sh*16)>(arg1);
@@ -3149 +3261 @@
 
 //The total number of operations is 14.5
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<32>::srli(bitblock256_t arg1)
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<32>::srli(bitblock256_t arg1)
 {
         return simd256<256>::srli<(sh*32)>(arg1);
@@ -3155 +3267 @@
 
 //The total number of operations is 14.5
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<64>::srli(bitblock256_t arg1)
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<64>::srli(bitblock256_t arg1)
 {
         return simd256<256>::srli<(sh*64)>(arg1);
@@ -3161 +3273 @@
 
 //The total number of operations is 14.5
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<128>::srli(bitblock256_t arg1)
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<128>::srli(bitblock256_t arg1)
 {
         return simd256<256>::srli<(sh*128)>(arg1);
@@ -3167 +3279 @@
 
 //The total number of operations is 14.5
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<256>::srli(bitblock256_t arg1)
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<256>::srli(bitblock256_t arg1)
 {
         return simd256<256>::srli<(sh*256)>(arg1);
@@ -3173 +3285 @@
 
 //The total number of operations is 1.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<1>::fill2(uint64_t val1, uint64_t val2)
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<1>::fill2(FieldType<1>::T val1, FieldType<1>::T val2)
 {
         return mvmd256<(2)>::fill(((val1<<1)|(val2&(1))));
@@ -3179 +3291 @@
 
 //The total number of operations is 1.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<2>::fill2(uint64_t val1, uint64_t val2)
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<2>::fill2(FieldType<2>::T val1, FieldType<2>::T val2)
 {
         return mvmd256<(4)>::fill(((val1<<2)|(val2&(3))));
@@ -3185 +3297 @@
 
 //The total number of operations is 1.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<4>::fill2(uint64_t val1, uint64_t val2)
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<4>::fill2(FieldType<4>::T val1, FieldType<4>::T val2)
 {
         return mvmd256<(8)>::fill(((val1<<4)|(val2&(15))));
@@ -3191 +3303 @@
 
 //The total number of operations is 1.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<8>::fill2(uint64_t val1, uint64_t val2)
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<8>::fill2(FieldType<8>::T val1, FieldType<8>::T val2)
 {
         return mvmd256<(16)>::fill(((val1<<8)|(val2&(255))));
@@ -3197 +3309 @@
 
 //The total number of operations is 1.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<16>::fill2(uint64_t val1, uint64_t val2)
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<16>::fill2(FieldType<16>::T val1, FieldType<16>::T val2)
 {
         return mvmd256<(32)>::fill(((val1<<16)|(val2&(65535))));
@@ -3203 +3315 @@
 
 //The total number of operations is 5.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<32>::fill2(uint64_t val1, uint64_t val2)
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<32>::fill2(FieldType<32>::T val1, FieldType<32>::T val2)
 {
         return simd256<1>::ifh(simd256<(64)>::himask(), mvmd256<32>::fill(val1), mvmd256<32>::fill(val2));
@@ -3209 +3321 @@
 
 //The total number of operations is 13.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<64>::fill2(uint64_t val1, uint64_t val2)
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<64>::fill2(FieldType<64>::T val1, FieldType<64>::T val2)
 {
         return simd256<1>::ifh(simd256<(128)>::himask(), mvmd256<64>::fill(val1), mvmd256<64>::fill(val2));
@@ -3215 +3327 @@
 
 //The total number of operations is 29.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<128>::fill2(uint64_t val1, uint64_t val2)
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<128>::fill2(FieldType<128>::T val1, FieldType<128>::T val2)
 {
         return simd256<1>::ifh(simd256<(256)>::himask(), mvmd256<128>::fill(val1), mvmd256<128>::fill(val2));
@@ -3221 +3333 @@
 
 //The total number of operations is 29.5
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<2>::dslli(bitblock256_t arg1, bitblock256_t arg2)
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<2>::dslli(bitblock256_t arg1, bitblock256_t arg2)
 {
         return simd_or(mvmd256<2>::slli<sh>(arg1), mvmd256<2>::srli<((128)-sh)>(arg2));
@@ -3227 +3339 @@
 
 //The total number of operations is 29.5
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<4>::dslli(bitblock256_t arg1, bitblock256_t arg2)
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<4>::dslli(bitblock256_t arg1, bitblock256_t arg2)
 {
         return simd_or(mvmd256<4>::slli<sh>(arg1), mvmd256<4>::srli<((64)-sh)>(arg2));
@@ -3233 +3345 @@
 
 //The total number of operations is 29.5
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<8>::dslli(bitblock256_t arg1, bitblock256_t arg2)
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<8>::dslli(bitblock256_t arg1, bitblock256_t arg2)
 {
         return simd_or(mvmd256<8>::slli<sh>(arg1), mvmd256<8>::srli<((32)-sh)>(arg2));
@@ -3239 +3351 @@
 
 //The total number of operations is 29.5
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<16>::dslli(bitblock256_t arg1, bitblock256_t arg2)
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<16>::dslli(bitblock256_t arg1, bitblock256_t arg2)
 {
         return simd_or(mvmd256<16>::slli<sh>(arg1), mvmd256<16>::srli<((16)-sh)>(arg2));
@@ -3245 +3357 @@
 
 //The total number of operations is 29.5
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<32>::dslli(bitblock256_t arg1, bitblock256_t arg2)
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<32>::dslli(bitblock256_t arg1, bitblock256_t arg2)
 {
         return simd_or(mvmd256<32>::slli<sh>(arg1), mvmd256<32>::srli<((8)-sh)>(arg2));
@@ -3251 +3363 @@
 
 //The total number of operations is 29.5
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<64>::dslli(bitblock256_t arg1, bitblock256_t arg2)
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<64>::dslli(bitblock256_t arg1, bitblock256_t arg2)
 {
         return simd_or(mvmd256<64>::slli<sh>(arg1), mvmd256<64>::srli<((4)-sh)>(arg2));
@@ -3257 +3369 @@
 
 //The total number of operations is 29.5
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<128>::dslli(bitblock256_t arg1, bitblock256_t arg2)
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<128>::dslli(bitblock256_t arg1, bitblock256_t arg2)
 {
         return simd_or(mvmd256<128>::slli<sh>(arg1), mvmd256<128>::srli<((2)-sh)>(arg2));
@@ -3263 +3375 @@
 
 //The total number of operations is 29.5
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<256>::dslli(bitblock256_t arg1, bitblock256_t arg2)
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<256>::dslli(bitblock256_t arg1, bitblock256_t arg2)
 {
         return simd_or(mvmd256<256>::slli<sh>(arg1), mvmd256<256>::srli<((1)-sh)>(arg2));
@@ -3269 +3381 @@
 
 //The total number of operations is 14.0
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<2>::slli(bitblock256_t arg1)
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<2>::slli(bitblock256_t arg1)
 {
         return simd256<256>::slli<(sh*2)>(arg1);
@@ -3275 +3387 @@
 
 //The total number of operations is 14.0
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<4>::slli(bitblock256_t arg1)
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<4>::slli(bitblock256_t arg1)
 {
         return simd256<256>::slli<(sh*4)>(arg1);
@@ -3281 +3393 @@
 
 //The total number of operations is 14.0
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<8>::slli(bitblock256_t arg1)
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<8>::slli(bitblock256_t arg1)
 {
         return simd256<256>::slli<(sh*8)>(arg1);
@@ -3287 +3399 @@
 
 //The total number of operations is 14.0
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<16>::slli(bitblock256_t arg1)
-{
-        return mvmd256<(8)>::slli<(sh*2)>(arg1);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<16>::slli(bitblock256_t arg1)
+{
+        return simd256<256>::slli<(sh*16)>(arg1);
 }
 
 //The total number of operations is 14.0
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<32>::slli(bitblock256_t arg1)
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<32>::slli(bitblock256_t arg1)
 {
         return simd256<256>::slli<(sh*32)>(arg1);
@@ -3299 +3411 @@
 
 //The total number of operations is 14.0
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<64>::slli(bitblock256_t arg1)
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<64>::slli(bitblock256_t arg1)
 {
         return simd256<256>::slli<(sh*64)>(arg1);
@@ -3305 +3417 @@
 
 //The total number of operations is 14.0
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<128>::slli(bitblock256_t arg1)
-{
-        return mvmd256<(64)>::slli<(sh*2)>(arg1);
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<128>::slli(bitblock256_t arg1)
+{
+        return simd256<256>::slli<(sh*128)>(arg1);
 }
 
 //The total number of operations is 14.0
-template <> template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<256>::slli(bitblock256_t arg1)
+template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<256>::slli(bitblock256_t arg1)
 {
         return simd256<256>::slli<(sh*256)>(arg1);
@@ -3317 +3429 @@
 
 //The total number of operations is 13.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<1>::fill8(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8)
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<1>::fill8(FieldType<1>::T val1, FieldType<1>::T val2, FieldType<1>::T val3, FieldType<1>::T val4, FieldType<1>::T val5, FieldType<1>::T val6, FieldType<1>::T val7, FieldType<1>::T val8)
 {
         return simd256<1>::ifh(simd256<(8)>::himask(), mvmd256<1>::fill4(val1, val2, val3, val4), mvmd256<1>::fill4(val5, val6, val7, val8));
@@ -3323 +3435 @@
 
 //The total number of operations is 13.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<2>::fill8(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8)
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<2>::fill8(FieldType<2>::T val1, FieldType<2>::T val2, FieldType<2>::T val3, FieldType<2>::T val4, FieldType<2>::T val5, FieldType<2>::T val6, FieldType<2>::T val7, FieldType<2>::T val8)
 {
         return simd256<1>::ifh(simd256<(16)>::himask(), mvmd256<2>::fill4(val1, val2, val3, val4), mvmd256<2>::fill4(val5, val6, val7, val8));
@@ -3329 +3441 @@
 
 //The total number of operations is 7.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<4>::fill8(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8)
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<4>::fill8(FieldType<4>::T val1, FieldType<4>::T val2, FieldType<4>::T val3, FieldType<4>::T val4, FieldType<4>::T val5, FieldType<4>::T val6, FieldType<4>::T val7, FieldType<4>::T val8)
 {
         return simd_or(mvmd256<(8)>::fill8((val1<<4), (val3<<4), (val5<<4), (val7<<4), (val1<<4), (val3<<4), (val5<<4), (val7<<4)), mvmd256<(8)>::fill8((val2&(15)), (val4&(15)), (val6&(15)), (val8&(15)), (val2&(15)), (val4&(15)), (val6&(15)), (val8&(15))));
@@ -3335 +3447 @@
 
 //The total number of operations is 3.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<8>::fill8(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8)
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<8>::fill8(FieldType<8>::T val1, FieldType<8>::T val2, FieldType<8>::T val3, FieldType<8>::T val4, FieldType<8>::T val5, FieldType<8>::T val6, FieldType<8>::T val7, FieldType<8>::T val8)
 {
         return simd_or(mvmd256<(16)>::fill8((val1<<8), (val3<<8), (val5<<8), (val7<<8), (val1<<8), (val3<<8), (val5<<8), (val7<<8)), mvmd256<(16)>::fill8((val2&(255)), (val4&(255)), (val6&(255)), (val8&(255)), (val2&(255)), (val4&(255)), (val6&(255)), (val8&(255))));
@@ -3341 +3453 @@
 
 //The total number of operations is 1.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<16>::fill8(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8)
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<16>::fill8(FieldType<16>::T val1, FieldType<16>::T val2, FieldType<16>::T val3, FieldType<16>::T val4, FieldType<16>::T val5, FieldType<16>::T val6, FieldType<16>::T val7, FieldType<16>::T val8)
 {
         return (bitblock256_t)_mm256_set_epi16((int32_t)(val1), (int32_t)(val2), (int32_t)(val3), (int32_t)(val4), (int32_t)(val5), (int32_t)(val6), (int32_t)(val7), (int32_t)(val8), (int32_t)(val1), (int32_t)(val2), (int32_t)(val3), (int32_t)(val4), (int32_t)(val5), (int32_t)(val6), (int32_t)(val7), (int32_t)(val8));
@@ -3347 +3459 @@
 
 //The total number of operations is 5.0
-template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<32>::fill8(uint64_t val1, uint64_t val2, uint64_t val3, uint64_t val4, uint64_t val5, uint64_t val6, uint64_t val7, uint64_t val8)
+template <> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<32>::fill8(FieldType<32>::T val1, FieldType<32>::T val2, FieldType<32>::T val3, FieldType<32>::T val4, FieldType<32>::T val5, FieldType<32>::T val6, FieldType<32>::T val7, FieldType<32>::T val8)
 {
         return simd256<1>::ifh(simd256<(256)>::himask(), mvmd256<32>::fill4(val1, val2, val3, val4), mvmd256<32>::fill4(val5, val6, val7, val8));
@@ -3359 +3471 @@
 
 //The total number of operations is 14.5
-template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t bitblock256::srli(bitblock256_t arg1)
+template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t bitblock256::srli(bitblock256_t arg1)
 {
         return simd256<256>::srli<sh>(arg1);
@@ -3371 +3483 @@
 
 //The total number of operations is 118.5
-IDISA_ALWAYS_INLINE uint64_t bitblock256::popcount(bitblock256_t arg1)
+IDISA_ALWAYS_INLINE uint16_t bitblock256::popcount(bitblock256_t arg1)
 {
         return mvmd256<64>::extract<0>(simd256<256>::popcount(arg1));
@@ -3383 +3495 @@
 
 //The total number of operations is 14.0
-template <uint64_t sh> IDISA_ALWAYS_INLINE bitblock256_t bitblock256::slli(bitblock256_t arg1)
+template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t bitblock256::slli(bitblock256_t arg1)
 {
         return simd256<256>::slli<sh>(arg1);

trunk/lib/idisa_cpp/idisa_sse2.cpp

@@ -15 +15 @@
 typedef __m128i bitblock128_t;
 
+#ifndef FIELD_TYPE
+#define FIELD_TYPE
 template <uint32_t fw> struct FieldType {
    typedef int T;  //default for FieldType::T is int
@@ -27 +29 @@
 template <> struct FieldType<64> {typedef uint64_t T;};
 template <> struct FieldType<128> {typedef uint64_t T;};
+#endif
 
 typedef FieldType<1>::T fw1_t;