Changeset 3462 for trunk/lib

Timestamp:

Sep 11, 2013, 5:04:25 PM (6 years ago)

Author:

linmengl

Message:

update mvmd_srli, mvmd_slli, bitblock_popcount, have a little bit better performance now

File:

• trunk/lib/idisa_cpp/idisa_avx2.cpp (modified) (16 diffs)

trunk/lib/idisa_cpp/idisa_avx2.cpp

@@ -1153 +1153 @@
 }
 
-//The total number of operations is 103.166666667
+//The total number of operations is 62.1666666667
 template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::ctz(bitblock256_t arg1)
 {
@@ -1177 +1177 @@
 }
 
-//The total number of operations is 17.75
+//The total number of operations is 6.875
 template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::sll(bitblock256_t arg1, bitblock256_t shift_mask)
 {
@@ -1415 +1415 @@
 }
 
-//The total number of operations is 59.0
+//The total number of operations is 18.0
 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));
+        return _mm256_castsi128_si256(_mm_cvtsi64_si128((int64_t)(bitblock256::popcount(arg1))));
 }
 
@@ -1750 +1749 @@
 }
 
-//The total number of operations is 18.5
+//The total number of operations is 6.875
 template <> IDISA_ALWAYS_INLINE bitblock256_t simd256<256>::srl(bitblock256_t arg1, bitblock256_t shift_mask)
 {
@@ -2506 +2505 @@
 //The total number of operations is 1.0
 template <> IDISA_ALWAYS_INLINE FieldType<256/8>::T hsimd256<8>::signmask(bitblock256_t arg1)
-{       
+{
         return _mm256_movemask_epi8(arg1);
 }
@@ -3091 +3090 @@
 }
 
-//The total number of operations is 20.5
+//The total number of operations is 5.5
 template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<64>::dsrli(bitblock256_t arg1, bitblock256_t arg2)
 {
@@ -3097 +3096 @@
 }
 
-//The total number of operations is 20.5
+//The total number of operations is 5.5
 template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<128>::dsrli(bitblock256_t arg1, bitblock256_t arg2)
 {
@@ -3103 +3102 @@
 }
 
-//The total number of operations is 20.5
+//The total number of operations is 5.5
 template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<256>::dsrli(bitblock256_t arg1, bitblock256_t arg2)
 {
@@ -3361 +3360 @@
 }
 
-//The total number of operations is 10.0
+//The total number of operations is 2.25
 template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<64>::srli(bitblock256_t arg1)
 {
-        return simd256<256>::srli<(sh*64)>(arg1);
-}
-
-//The total number of operations is 10.0
+        return ((sh == 3) ? simd_and(_mm256_set_epi64x((int64_t)(0), (int64_t)(0), (int64_t)(0), (int64_t)(-1)), _mm256_permute4x64_epi64(arg1, (int32_t)(3))) : ((sh == 2) ? simd_and(_mm256_set_epi64x((int64_t)(0), (int64_t)(0), (int64_t)(-1), (int64_t)(-1)), _mm256_permute4x64_epi64(arg1, (int32_t)(14))) : ((sh == 1) ? simd_and(_mm256_set_epi64x((int64_t)(0), (int64_t)(-1), (int64_t)(-1), (int64_t)(-1)), _mm256_permute4x64_epi64(arg1, (int32_t)(57))) : ((sh == 0) ? arg1 : simd256<32>::constant<0>()))));
+}
+
+//The total number of operations is 2.25
 template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<128>::srli(bitblock256_t arg1)
 {
-        return simd256<256>::srli<(sh*128)>(arg1);
-}
-
-//The total number of operations is 10.0
+        return mvmd256<(64)>::srli<(sh*2)>(arg1);
+}
+
+//The total number of operations is 2.25
 template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<256>::srli(bitblock256_t arg1)
 {
-        return simd256<256>::srli<(sh*256)>(arg1);
+        return mvmd256<(128)>::srli<(sh*2)>(arg1);
 }
 
@@ -3457 +3456 @@
 }
 
-//The total number of operations is 20.5
+//The total number of operations is 5.5
 template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<64>::dslli(bitblock256_t arg1, bitblock256_t arg2)
 {
@@ -3463 +3462 @@
 }
 
-//The total number of operations is 20.5
+//The total number of operations is 5.5
 template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<128>::dslli(bitblock256_t arg1, bitblock256_t arg2)
 {
@@ -3469 +3468 @@
 }
 
-//The total number of operations is 20.5
+//The total number of operations is 5.5
 template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<256>::dslli(bitblock256_t arg1, bitblock256_t arg2)
 {
@@ -3505 +3504 @@
 }
 
-//The total number of operations is 9.5
+//The total number of operations is 2.25
 template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<64>::slli(bitblock256_t arg1)
 {
-        return simd256<256>::slli<(sh*64)>(arg1);
-}
-
-//The total number of operations is 9.5
+        return ((sh == 1) ? simd_and(_mm256_set_epi64x((int64_t)(-1), (int64_t)(-1), (int64_t)(-1), (int64_t)(0)), _mm256_permute4x64_epi64(arg1, (int32_t)((144)))) : ((sh == 2) ? simd_and(_mm256_set_epi64x((int64_t)(-1), (int64_t)(-1), (int64_t)(0), (int64_t)(0)), _mm256_permute4x64_epi64(arg1, (int32_t)(64))) : ((sh == 3) ? simd_and(_mm256_set_epi64x((int64_t)(-1), (int64_t)(0), (int64_t)(0), (int64_t)(0)), _mm256_permute4x64_epi64(arg1, (int32_t)(0))) : ((sh == 0) ? arg1 : simd256<32>::constant<0>()))));
+}
+
+//The total number of operations is 2.25
 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 9.5
+        return mvmd256<(64)>::slli<(sh*2)>(arg1);
+}
+
+//The total number of operations is 2.25
 template <> template <uint16_t sh> IDISA_ALWAYS_INLINE bitblock256_t mvmd256<256>::slli(bitblock256_t arg1)
 {
-        return simd256<256>::slli<(sh*256)>(arg1);
+        return mvmd256<(128)>::slli<(sh*2)>(arg1);
 }
 
@@ -3559 +3558 @@
 }
 
-//The total number of operations is 17.75
+//The total number of operations is 6.875
 IDISA_ALWAYS_INLINE bitblock256_t bitblock256::sll(bitblock256_t arg1, bitblock256_t arg2)
 {
@@ -3577 +3576 @@
 }
 
-//The total number of operations is 62.0
+//The total number of operations is 16.0
 IDISA_ALWAYS_INLINE uint16_t bitblock256::popcount(bitblock256_t arg1)
 {
-        return mvmd256<64>::extract<0>(simd256<256>::popcount(arg1));
+        return (((__builtin_popcountll((uint64_t)(mvmd256<64>::extract<0>(arg1)))+__builtin_popcountll((uint64_t)(mvmd256<64>::extract<1>(arg1))))+__builtin_popcountll((uint64_t)(mvmd256<64>::extract<2>(arg1))))+__builtin_popcountll((uint64_t)(mvmd256<64>::extract<3>(arg1))));
 }
 
@@ -3589 +3588 @@
 }
 
-//The total number of operations is 18.5
+//The total number of operations is 6.875
 IDISA_ALWAYS_INLINE bitblock256_t bitblock256::srl(bitblock256_t arg1, bitblock256_t arg2)
 {