Compute functions for SSE4.2 SIMD instruction set (internal) More...
#include <mmintrin.h>#include <emmintrin.h>#include <smmintrin.h>#include <nmmintrin.h>#include <immintrin.h>#include "bmdef.h"#include "bmutil.h"#include "bmsse_util.h"
Include dependency graph for bmsse4.h:
Go to the source code of this file.
Namespaces | |
| namespace | bm |
Macros | |
| #define | BM_BSF32 bm::bsf_asm32 |
| #define | VECT_XOR_ARR_2_MASK(dst, src, src_end, mask) sse2_xor_arr_2_mask((__m128i*)(dst), (__m128i*)(src), (__m128i*)(src_end), (bm::word_t)mask) |
| #define | VECT_ANDNOT_ARR_2_MASK(dst, src, src_end, mask) sse2_andnot_arr_2_mask((__m128i*)(dst), (__m128i*)(src), (__m128i*)(src_end), (bm::word_t)mask) |
| #define | VECT_BITCOUNT(first, last) sse4_bit_count((__m128i*) (first), (__m128i*) (last)) |
| #define | VECT_BITCOUNT_AND(first, last, mask) sse4_bit_count_op((__m128i*) (first), (__m128i*) (last), (__m128i*) (mask), sse2_and) |
| #define | VECT_BITCOUNT_OR(first, last, mask) sse4_bit_count_op((__m128i*) (first), (__m128i*) (last), (__m128i*) (mask), sse2_or) |
| #define | VECT_BITCOUNT_XOR(first, last, mask) sse4_bit_count_op((__m128i*) (first), (__m128i*) (last), (__m128i*) (mask), sse2_xor) |
| #define | VECT_BITCOUNT_SUB(first, last, mask) sse4_bit_count_op((__m128i*) (first), (__m128i*) (last), (__m128i*) (mask), sse2_sub) |
| #define | VECT_INVERT_BLOCK(first) sse2_invert_block((__m128i*)first); |
| #define | VECT_AND_BLOCK(dst, src) sse4_and_block((__m128i*) dst, (__m128i*) (src)) |
| #define | VECT_AND_DIGEST(dst, src) sse4_and_digest((__m128i*) dst, (const __m128i*) (src)) |
| #define | VECT_AND_OR_DIGEST_2WAY(dst, src1, src2) sse4_and_or_digest_2way((__m128i*) dst, (const __m128i*) (src1), (const __m128i*) (src2)) |
| #define | VECT_AND_DIGEST_5WAY(dst, src1, src2, src3, src4) sse4_and_digest_5way((__m128i*) dst, (const __m128i*) (src1), (const __m128i*) (src2), (const __m128i*) (src3), (const __m128i*) (src4)) |
| #define | VECT_AND_DIGEST_2WAY(dst, src1, src2) sse4_and_digest_2way((__m128i*) dst, (const __m128i*) (src1), (const __m128i*) (src2)) |
| #define | VECT_OR_BLOCK(dst, src) sse2_or_block((__m128i*) dst, (__m128i*) (src)) |
| #define | VECT_OR_BLOCK_2WAY(dst, src1, src2) sse2_or_block_2way((__m128i*) (dst), (const __m128i*) (src1), (const __m128i*) (src2)) |
| #define | VECT_OR_BLOCK_3WAY(dst, src1, src2) sse2_or_block_3way((__m128i*) (dst), (const __m128i*) (src1), (const __m128i*) (src2)) |
| #define | VECT_OR_BLOCK_5WAY(dst, src1, src2, src3, src4) sse2_or_block_5way((__m128i*) (dst), (__m128i*) (src1), (__m128i*) (src2), (__m128i*) (src3), (__m128i*) (src4)) |
| #define | VECT_SUB_BLOCK(dst, src) sse2_sub_block((__m128i*) dst, (const __m128i*) (src)) |
| #define | VECT_SUB_DIGEST(dst, src) sse4_sub_digest((__m128i*) dst, (const __m128i*) (src)) |
| #define | VECT_SUB_DIGEST_2WAY(dst, src1, src2) sse4_sub_digest_2way((__m128i*) dst, (const __m128i*) (src1), (const __m128i*) (src2)) |
| #define | VECT_XOR_BLOCK(dst, src) sse2_xor_block((__m128i*) dst, (__m128i*) (src)) |
| #define | VECT_XOR_BLOCK_2WAY(dst, src1, src2) sse2_xor_block_2way((__m128i*) (dst), (const __m128i*) (src1), (const __m128i*) (src2)) |
| #define | VECT_COPY_BLOCK(dst, src) sse2_copy_block((__m128i*) dst, (__m128i*) (src)) |
| #define | VECT_COPY_BLOCK_UNALIGN(dst, src) sse2_copy_block_unalign((__m128i*) dst, (__m128i*) (src)) |
| #define | VECT_STREAM_BLOCK(dst, src) sse2_stream_block((__m128i*) dst, (__m128i*) (src)) |
| #define | VECT_STREAM_BLOCK_UNALIGN(dst, src) sse2_stream_block_unalign((__m128i*) dst, (__m128i*) (src)) |
| #define | VECT_SET_BLOCK(dst, value) sse2_set_block((__m128i*) dst, value) |
| #define | VECT_IS_ZERO_BLOCK(dst) sse4_is_all_zero((__m128i*) dst) |
| #define | VECT_IS_ONE_BLOCK(dst) sse4_is_all_one((__m128i*) dst) |
| #define | VECT_IS_DIGEST_ZERO(start) sse4_is_digest_zero((__m128i*)start) |
| #define | VECT_BLOCK_SET_DIGEST(dst, val) sse4_block_set_digest((__m128i*)dst, val) |
| #define | VECT_LOWER_BOUND_SCAN_U32(arr, target, from, to) sse2_lower_bound_scan_u32(arr, target, from, to) |
| #define | VECT_SHIFT_L1(b, acc, co) sse42_shift_l1((__m128i*)b, acc, co) |
| #define | VECT_SHIFT_R1(b, acc, co) sse42_shift_r1((__m128i*)b, acc, co) |
| #define | VECT_SHIFT_R1_AND(b, co, m, digest) sse42_shift_r1_and((__m128i*)b, co, (__m128i*)m, digest) |
| #define | VECT_ARR_BLOCK_LOOKUP(idx, size, nb, start) sse42_idx_arr_block_lookup(idx, size, nb, start) |
| #define | VECT_SET_BLOCK_BITS(block, idx, start, stop) sse42_set_block_bits(block, idx, start, stop) |
| #define | VECT_BLOCK_CHANGE(block, size) sse42_bit_block_calc_change((__m128i*)block, size) |
| #define | VECT_BLOCK_XOR_CHANGE(block, xor_block, size, gc, bc) sse42_bit_block_calc_xor_change((__m128i*)block, (__m128i*)xor_block, size, gc, bc) |
| #define | VECT_BLOCK_CHANGE_BC(block, gc, bc) sse42_bit_block_calc_change_bc((__m128i*)block, gc, bc) |
| #define | VECT_BIT_FIND_FIRST(src, pos) sse42_bit_find_first((__m128i*) src, pos) |
| #define | VECT_BIT_FIND_DIFF(src1, src2, pos) sse42_bit_find_first_diff((__m128i*) src1, (__m128i*) (src2), pos) |
| #define | VECT_BIT_BLOCK_XOR(t, src, src_xor, d) sse42_bit_block_xor(t, src, src_xor, d) |
| #define | VECT_BIT_BLOCK_XOR_2WAY(t, src_xor, d) sse42_bit_block_xor_2way(t, src_xor, d) |
| #define | VECT_GAP_BFIND(buf, pos, is_set) sse42_gap_bfind(buf, pos, is_set) |
Functions | |
| bm::id_t | bm::sse4_bit_count (const __m128i *block, const __m128i *block_end) BMNOEXCEPT |
| More... | |
| BMFORCEINLINE unsigned | bm::op_xor (unsigned a, unsigned b) BMNOEXCEPT |
| BMFORCEINLINE unsigned | bm::op_or (unsigned a, unsigned b) BMNOEXCEPT |
| BMFORCEINLINE unsigned | bm::op_and (unsigned a, unsigned b) BMNOEXCEPT |
| template<class Func > | |
| bm::id_t | bm::sse4_bit_count_op (const __m128i *BMRESTRICT block, const __m128i *BMRESTRICT block_end, const __m128i *BMRESTRICT mask_block, Func sse2_func) BMNOEXCEPT |
| More... | |
| bool | bm::sse4_is_all_zero (const __m128i *BMRESTRICT block) BMNOEXCEPT |
| check if block is all zero bits More... | |
| BMFORCEINLINE bool | bm::sse4_is_digest_zero (const __m128i *BMRESTRICT block) BMNOEXCEPT |
| check if digest stride is all zero bits More... | |
| BMFORCEINLINE void | bm::sse4_block_set_digest (__m128i *dst, unsigned value) BMNOEXCEPT |
| set digest stride to 0xFF.. or 0x0 value More... | |
| unsigned | bm::sse4_and_block (__m128i *BMRESTRICT dst, const __m128i *BMRESTRICT src) BMNOEXCEPT |
| AND blocks2 dst &= *src. More... | |
| BMFORCEINLINE bool | bm::sse4_and_digest (__m128i *BMRESTRICT dst, const __m128i *BMRESTRICT src) BMNOEXCEPT |
| AND block digest stride dst &= *src. More... | |
| BMFORCEINLINE bool | bm::sse4_and_digest_2way (__m128i *BMRESTRICT dst, const __m128i *BMRESTRICT src1, const __m128i *BMRESTRICT src2) BMNOEXCEPT |
| AND block digest stride dst = *src1 & src2. More... | |
| bool | bm::sse4_and_or_digest_2way (__m128i *BMRESTRICT dst, const __m128i *BMRESTRICT src1, const __m128i *BMRESTRICT src2) BMNOEXCEPT |
| AND-OR block digest stride dst |= *src1 & src2. More... | |
| bool | bm::sse4_and_digest_5way (__m128i *BMRESTRICT dst, const __m128i *BMRESTRICT src1, const __m128i *BMRESTRICT src2, const __m128i *BMRESTRICT src3, const __m128i *BMRESTRICT src4) BMNOEXCEPT |
| AND block digest stride. More... | |
| BMFORCEINLINE bool | bm::sse4_sub_digest (__m128i *BMRESTRICT dst, const __m128i *BMRESTRICT src) BMNOEXCEPT |
| SUB (AND NOT) block digest stride dst &= ~*src. More... | |
| BMFORCEINLINE bool | bm::sse4_sub_digest_2way (__m128i *BMRESTRICT dst, const __m128i *BMRESTRICT src1, const __m128i *BMRESTRICT src2) BMNOEXCEPT |
| 2-operand SUB (AND NOT) block digest stride dst = src1 & ~*src2 More... | |
| bool | bm::sse4_is_all_one (const __m128i *BMRESTRICT block) BMNOEXCEPT |
| check if block is all ONE bits More... | |
| BMFORCEINLINE bool | bm::sse42_test_all_one_wave (const void *ptr) BMNOEXCEPT |
| check if SSE wave is all oxFFFF...FFF More... | |
| BMFORCEINLINE bool | bm::sse42_test_all_zero_wave (const void *ptr) BMNOEXCEPT |
| check if wave of pointers is all NULL More... | |
| BMFORCEINLINE bool | bm::sse42_test_all_zero_wave2 (const void *ptr0, const void *ptr1) BMNOEXCEPT |
| check if 2 waves of pointers are all NULL More... | |
| BMFORCEINLINE bool | bm::sse42_test_all_eq_wave2 (const void *ptr0, const void *ptr1) BMNOEXCEPT |
| check if wave of 2 pointers are the same (null or FULL) More... | |
| unsigned | bm::sse42_bit_block_calc_change (const __m128i *BMRESTRICT block, unsigned size) BMNOEXCEPT |
| More... | |
| void | bm::sse42_bit_block_calc_xor_change (const __m128i *BMRESTRICT block, const __m128i *BMRESTRICT xor_block, unsigned size, unsigned *BMRESTRICT gc, unsigned *BMRESTRICT bc) BMNOEXCEPT |
| More... | |
| void | bm::sse42_bit_block_calc_change_bc (const __m128i *BMRESTRICT block, unsigned *gc, unsigned *bc) BMNOEXCEPT |
| More... | |
| bool | bm::sse42_bit_find_first_diff (const __m128i *BMRESTRICT block1, const __m128i *BMRESTRICT block2, unsigned *pos) BMNOEXCEPT |
| Find first bit which is different between two bit-blocks. More... | |
| bool | bm::sse42_bit_find_first (const __m128i *BMRESTRICT block, unsigned *pos) BMNOEXCEPT |
| Find first non-zero bit. More... | |
| unsigned | bm::sse4_gap_find (const bm::gap_word_t *BMRESTRICT pbuf, const bm::gap_word_t pos, const unsigned size) BMNOEXCEPT |
| More... | |
| unsigned | bm::sse42_gap_bfind (const unsigned short *BMRESTRICT buf, unsigned pos, unsigned *BMRESTRICT is_set) BMNOEXCEPT |
| Hybrid binary search, starts as binary, then switches to linear scan. More... | |
| unsigned | bm::sse42_gap_test (const unsigned short *BMRESTRICT buf, unsigned pos) BMNOEXCEPT |
| Hybrid binary search, starts as binary, then switches to scan. More... | |
| int | bm::sse42_cmpge_u32 (__m128i vect4, unsigned value) BMNOEXCEPT |
| Experimental (test) function to do SIMD vector search (lower bound) in sorted, growing array. More... | |
| unsigned | bm::sse42_idx_arr_block_lookup (const unsigned *idx, unsigned size, unsigned nb, unsigned start) BMNOEXCEPT |
| More... | |
| void | bm::sse42_set_block_bits (bm::word_t *BMRESTRICT block, const unsigned *BMRESTRICT idx, unsigned start, unsigned stop) BMNOEXCEPT |
| More... | |
| void | bm::sse4_bit_block_gather_scatter (unsigned *BMRESTRICT arr, const unsigned *BMRESTRICT blk, const unsigned *BMRESTRICT idx, unsigned size, unsigned start, unsigned bit_idx) BMNOEXCEPT |
| More... | |
| bool | bm::sse42_shift_l1 (__m128i *block, unsigned *empty_acc, unsigned co1) BMNOEXCEPT |
| block shift left by 1 More... | |
| bool | bm::sse42_shift_r1 (__m128i *block, unsigned *empty_acc, unsigned co1) BMNOEXCEPT |
| block shift right by 1 More... | |
| bool | bm::sse42_shift_r1_and (__m128i *block, bm::word_t co1, const __m128i *BMRESTRICT mask_block, bm::id64_t *digest) BMNOEXCEPT |
| block shift right by 1 plus AND More... | |
| void | bm::sse42_bit_block_xor (bm::word_t *target_block, const bm::word_t *block, const bm::word_t *xor_block, bm::id64_t digest) BMNOEXCEPT |
| Build partial XOR product of 2 bit-blocks using digest mask. More... | |
| void | bm::sse42_bit_block_xor_2way (bm::word_t *target_block, const bm::word_t *xor_block, bm::id64_t digest) BMNOEXCEPT |
| Build partial XOR product of 2 bit-blocks using digest mask. More... | |
Detailed Description
Compute functions for SSE4.2 SIMD instruction set (internal)
Aside from SSE4.2 it also compiles in WASM SIMD mode for 128-bit SIMD portable target.
Definition in file bmsse4.h.
