github_mirrors/vectorscan
3
0
Fork 1
mirror of https://github.com/VectorCamp/vectorscan.git synced 2025-06-28 16:41:01 +03:00
Code Issues Packages Projects Releases Wiki Activity

comparemask_type is u64a on Arm, use single load_mask

Browse Source
This commit is contained in:
Konstantinos Margaritis 2023-12-18 20:23:07 +00:00
parent 5814d3298f
commit db3b0e9474
3 changed files with 35 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

17
src/hwlm/noodle_engine_simd.hpp
View File

@ -86,15 +86,21 @@ hwlm_error_t scanSingleMain(const struct noodTable *n, const u8 *buf,
DEBUG_PRINTF("d - d0: %ld \n", d - d0); DEBUG_PRINTF("d - d0: %ld \n", d - d0);
#if defined(HAVE_MASKED_LOADS) #if defined(HAVE_MASKED_LOADS)
uint8_t l = d - d0; uint8_t l = d - d0;
typename SuperVector<S>::comparemask_type mask = ~SuperVector<S>::single_load_mask(l); typename SuperVector<S>::comparemask_type mask = ~SuperVector<S>::load_mask(l);
SuperVector<S> chars = SuperVector<S>::loadu_maskz(d0, mask) & caseMask; SuperVector<S> chars = SuperVector<S>::loadu_maskz(d0, mask) & caseMask;
typename SuperVector<S>::comparemask_type z = mask1.eqmask(chars); typename SuperVector<S>::comparemask_type z = mask1.eqmask(chars);
DEBUG_PRINTF("mask: %08llx\n", mask); DEBUG_PRINTF("mask: %08llx\n", mask);
hwlm_error_t rv = single_zscan<S>(n, d0, buf, z, len, cbi); hwlm_error_t rv = single_zscan<S>(n, d0, buf, z, len, cbi);
#else #else
uint8_t l = d0 + S - d; uint8_t l = d0 + S - d;
DEBUG_PRINTF("l: %d \n", l);
SuperVector<S> chars = SuperVector<S>::loadu_maskz(d, l) & caseMask; SuperVector<S> chars = SuperVector<S>::loadu_maskz(d, l) & caseMask;
chars.print8("chars");
typename SuperVector<S>::comparemask_type z = mask1.eqmask(chars); typename SuperVector<S>::comparemask_type z = mask1.eqmask(chars);
DEBUG_PRINTF("z: %08llx\n", (u64a) z);
z = SuperVector<S>::iteration_mask(z);
DEBUG_PRINTF("z: %08llx\n", (u64a) z);
hwlm_error_t rv = single_zscan<S>(n, d, buf, z, len, cbi); hwlm_error_t rv = single_zscan<S>(n, d, buf, z, len, cbi);
#endif #endif
chars.print32("chars"); chars.print32("chars");
@ -125,6 +131,8 @@ hwlm_error_t scanSingleMain(const struct noodTable *n, const u8 *buf,
uint8_t l = buf_end - d; uint8_t l = buf_end - d;
SuperVector<S> chars = SuperVector<S>::loadu_maskz(d, l) & caseMask; SuperVector<S> chars = SuperVector<S>::loadu_maskz(d, l) & caseMask;
typename SuperVector<S>::comparemask_type z = mask1.eqmask(chars); typename SuperVector<S>::comparemask_type z = mask1.eqmask(chars);
z = SuperVector<S>::iteration_mask(z);
hwlm_error_t rv = single_zscan<S>(n, d, buf, z, len, cbi); hwlm_error_t rv = single_zscan<S>(n, d, buf, z, len, cbi);
RETURN_IF_TERMINATED(rv); RETURN_IF_TERMINATED(rv);
} }
@ -160,12 +168,12 @@ hwlm_error_t scanDoubleMain(const struct noodTable *n, const u8 *buf,
const u8 *d0 = ROUNDDOWN_PTR(d, S); const u8 *d0 = ROUNDDOWN_PTR(d, S);
#if defined(HAVE_MASKED_LOADS) #if defined(HAVE_MASKED_LOADS)
uint8_t l = d - d0; uint8_t l = d - d0;
typename SuperVector<S>::comparemask_type mask = ~SuperVector<S>::double_load_mask(l); typename SuperVector<S>::comparemask_type mask = ~SuperVector<S>::load_mask(l);
SuperVector<S> chars = SuperVector<S>::loadu_maskz(d0, mask) & caseMask; SuperVector<S> chars = SuperVector<S>::loadu_maskz(d0, mask) & caseMask;
typename SuperVector<S>::comparemask_type z1 = mask1.eqmask(chars); typename SuperVector<S>::comparemask_type z1 = mask1.eqmask(chars);
typename SuperVector<S>::comparemask_type z2 = mask2.eqmask(chars); typename SuperVector<S>::comparemask_type z2 = mask2.eqmask(chars);
typename SuperVector<S>::comparemask_type z = (z1 << SuperVector<S>::mask_width()) & z2; typename SuperVector<S>::comparemask_type z = (z1 << SuperVector<S>::mask_width()) & z2;
DEBUG_PRINTF("z: %0llx\n", z); z = SuperVector<S>::iteration_mask(z);
lastz1 = z1 >> (S - 1); lastz1 = z1 >> (S - 1);
DEBUG_PRINTF("mask: %08llx\n", mask); DEBUG_PRINTF("mask: %08llx\n", mask);
@ -176,8 +184,9 @@ hwlm_error_t scanDoubleMain(const struct noodTable *n, const u8 *buf,
chars.print8("chars"); chars.print8("chars");
typename SuperVector<S>::comparemask_type z1 = mask1.eqmask(chars); typename SuperVector<S>::comparemask_type z1 = mask1.eqmask(chars);
typename SuperVector<S>::comparemask_type z2 = mask2.eqmask(chars); typename SuperVector<S>::comparemask_type z2 = mask2.eqmask(chars);
typename SuperVector<S>::comparemask_type z = (z1 << SuperVector<S>::mask_width()) & z2; typename SuperVector<S>::comparemask_type z = (z1 << SuperVector<S>::mask_width()) & z2;
z = SuperVector<S>::iteration_mask(z);
hwlm_error_t rv = double_zscan<S>(n, d, buf, z, len, cbi); hwlm_error_t rv = double_zscan<S>(n, d, buf, z, len, cbi);
lastz1 = z1 >> (l - 1); lastz1 = z1 >> (l - 1);
#endif #endif

19
src/util/supervector/arch/arm/impl.cpp
View File

@ -525,11 +525,26 @@ really_inline SuperVector<16> SuperVector<16>::load(void const *ptr)
template <> template <>
really_inline SuperVector<16> SuperVector<16>::loadu_maskz(void const *ptr, uint8_t const len) really_inline SuperVector<16> SuperVector<16>::loadu_maskz(void const *ptr, uint8_t const len)
{ {
SuperVector mask = Ones_vshr(16 -len); SuperVector mask = Ones_vshr(16 - len);
SuperVector<16> v = loadu(ptr); SuperVector v = loadu(ptr);
return mask & v; return mask & v;
} }
template <>
really_inline SuperVector<16> SuperVector<16>::loadu_maskz(void const *ptr, typename base_type::comparemask_type const mask)
{
DEBUG_PRINTF("mask = %08llx\n", mask);
SuperVector v = loadu(ptr);
(void)mask;
return v; // FIXME: & mask
}
template<>
really_inline typename SuperVector<16>::comparemask_type SuperVector<16>::findLSB(typename SuperVector<16>::comparemask_type &z)
{
return findAndClearLSB_64(&z) >> 2;
}
template<> template<>
really_inline SuperVector<16> SuperVector<16>::alignr(SuperVector<16> &other, int8_t offset) really_inline SuperVector<16> SuperVector<16>::alignr(SuperVector<16> &other, int8_t offset)
{ {

7
src/util/supervector/supervector.hpp
View File

@ -130,7 +130,11 @@ struct BaseVector<16>
static constexpr bool is_valid = true; static constexpr bool is_valid = true;
static constexpr u16 size = 16; static constexpr u16 size = 16;
using type = m128; using type = m128;
#if defined(ARCH_ARM32) || defined(ARCH_AARCH64)
using comparemask_type = u64a;
#else
using comparemask_type = u32; using comparemask_type = u32;
#endif
static constexpr bool has_previous = false; static constexpr bool has_previous = false;
using previous_type = u64a; using previous_type = u64a;
static constexpr u16 previous_size = 8; static constexpr u16 previous_size = 8;
@ -229,8 +233,7 @@ public:
static typename base_type::comparemask_type static typename base_type::comparemask_type
iteration_mask(typename base_type::comparemask_type mask); iteration_mask(typename base_type::comparemask_type mask);
static typename base_type::comparemask_type single_load_mask(uint8_t const len) { return (((1ULL) << (len)) - 1ULL); } static typename base_type::comparemask_type load_mask(uint8_t const len) { return (((1ULL) << (len)) - 1ULL); }
static typename base_type::comparemask_type double_load_mask(uint8_t const len) { return (((1ULL) << (len)) - 1ULL); }
static typename base_type::comparemask_type findLSB(typename base_type::comparemask_type &z); static typename base_type::comparemask_type findLSB(typename base_type::comparemask_type &z);
static SuperVector loadu(void const *ptr); static SuperVector loadu(void const *ptr);
static SuperVector load(void const *ptr); static SuperVector load(void const *ptr);