fix failing corner case, add pshufb_maskz()

This commit is contained in:
Konstantinos Margaritis
2021-07-23 18:55:56 +03:00
parent e35b88f2c8
commit 603bc14cdd
4 changed files with 68 additions and 2 deletions

View File

@@ -348,6 +348,13 @@ really_inline SuperVector<16> SuperVector<16>::pshufb(SuperVector<16> b)
return {vqtbl1q_s8((int8x16_t)u.v128[0], (uint8x16_t)btranslated)};
}
template<>
really_inline SuperVector<16> SuperVector<16>::pshufb_maskz(SuperVector<16> b, uint8_t const len)
{
SuperVector<16> mask = Ones().rshift128_var(16 -len);
return mask & pshufb(b);
}
#ifdef HS_OPTIMIZE
template<>
really_inline SuperVector<16> SuperVector<16>::lshift64(uint8_t const N)

View File

@@ -312,6 +312,13 @@ really_inline SuperVector<16> SuperVector<16>::pshufb(SuperVector<16> b)
return {_mm_shuffle_epi8(u.v128[0], b.u.v128[0])};
}
template<>
really_inline SuperVector<16> SuperVector<16>::pshufb_maskz(SuperVector<16> b, uint8_t const len)
{
SuperVector<16> mask = Ones().rshift128_var(16 -len);
return mask & pshufb(b);
}
#ifdef HS_OPTIMIZE
template<>
really_inline SuperVector<16> SuperVector<16>::lshift64(uint8_t const N)
@@ -733,6 +740,13 @@ really_inline SuperVector<32> SuperVector<32>::pshufb(SuperVector<32> b)
return {_mm256_shuffle_epi8(u.v256[0], b.u.v256[0])};
}
template<>
really_inline SuperVector<32> SuperVector<32>::pshufb_maskz(SuperVector<32> b, uint8_t const len)
{
SuperVector<32> mask = Ones().rshift128_var(32 -len);
return mask & pshufb(b);
}
#ifdef HS_OPTIMIZE
template<>
really_inline SuperVector<32> SuperVector<32>::lshift64(uint8_t const N)
@@ -1176,6 +1190,13 @@ really_inline SuperVector<64> SuperVector<64>::pshufb(SuperVector<64> b)
return {_mm512_shuffle_epi8(u.v512[0], b.u.v512[0])};
}
template<>
really_inline SuperVector<64> SuperVector<64>::pshufb_maskz(SuperVector<64> b, uint8_t const len)
{
u64a mask = (~0ULL) >> (64 - len);
DEBUG_PRINTF("mask = %016llx\n", mask);
return {_mm512_maskz_shuffle_epi8(mask, u.v512[0], b.u.v512[0])};
}
#ifdef HS_OPTIMIZE
template<>

View File

@@ -216,6 +216,7 @@ public:
SuperVector alignr(SuperVector &other, int8_t offset);
SuperVector pshufb(SuperVector b);
SuperVector pshufb_maskz(SuperVector b, uint8_t const len);
SuperVector lshift64(uint8_t const N);
SuperVector rshift64(uint8_t const N);
SuperVector lshift128(uint8_t const N);