Processor specific optimizations for SSE2 instructions

Collaboration diagram for Processor specific optimizations for SSE2 instructions:

Data Structures
class	bm::sse_empty_guard
	SSE2 reinitialization guard class. More...
Functions
bm::id_t	bm::sse2_bit_count (const __m128i *block, const __m128i *block_end)
BMFORCEINLINE void	bm::sse2_xor_arr_2_mask (__m128i *BMRESTRICT dst, const __m128i *BMRESTRICT src, const __m128i *BMRESTRICT src_end, bm::word_t mask)
	XOR array elements to specified mask dst = *src ^ mask.
BMFORCEINLINE void	bm::sse2_andnot_arr_2_mask (__m128i *BMRESTRICT dst, const __m128i *BMRESTRICT src, const __m128i *BMRESTRICT src_end, bm::word_t mask)
	Inverts array elements and NOT them to specified mask dst = ~*src & mask.
BMFORCEINLINE void	bm::sse2_and_arr (__m128i *BMRESTRICT dst, const __m128i *BMRESTRICT src, const __m128i *BMRESTRICT src_end)
	AND array elements against another array dst &= *src.
BMFORCEINLINE void	bm::sse2_or_arr (__m128i *BMRESTRICT dst, const __m128i *BMRESTRICT src, const __m128i *BMRESTRICT src_end)
	OR array elements against another array dst |= *src.
BMFORCEINLINE void	bm::sse2_xor_arr (__m128i *BMRESTRICT dst, const __m128i *BMRESTRICT src, const __m128i *BMRESTRICT src_end)
	OR array elements against another array dst ^= *src.
BMFORCEINLINE void	bm::sse2_sub_arr (__m128i *BMRESTRICT dst, const __m128i *BMRESTRICT src, const __m128i *BMRESTRICT src_end)
	AND-NOT (SUB) array elements against another array dst &= ~*src.
BMFORCEINLINE void	bm::sse2_set_block (__m128i *BMRESTRICT dst, __m128i *BMRESTRICT dst_end, bm::word_t value)
	SSE2 block memset dst = value.
BMFORCEINLINE void	bm::sse2_copy_block (__m128i *BMRESTRICT dst, const __m128i *BMRESTRICT src, const __m128i *BMRESTRICT src_end)
	SSE2 block copy dst = *src.
BMFORCEINLINE void	bm::sse2_invert_arr (bm::word_t *first, bm::word_t *last)
	Invert array elements dst = ~*dst or dst ^= *dst.

---

Function Documentation

BMFORCEINLINE void bm::sse2_and_arr	(	__m128i *BMRESTRICT	dst,
		const __m128i *BMRESTRICT	src,
		const __m128i *BMRESTRICT	src_end
	)

AND array elements against another array dst &= *src.

Definition at line 126 of file bmsse_util.h.

BMFORCEINLINE void bm::sse2_andnot_arr_2_mask	(	__m128i *BMRESTRICT	dst,
		const __m128i *BMRESTRICT	src,
		const __m128i *BMRESTRICT	src_end,
		bm::word_t	mask
	)

Inverts array elements and NOT them to specified mask dst = ~*src & mask.

Definition at line 98 of file bmsse_util.h.

bm::id_t bm::sse2_bit_count	(	const __m128i *	block,
		const __m128i *	block_end
	)		[inline]

SSE2 optimized bitcounting function implements parallel bitcounting algorithm for SSE2 instruction set.

unsigned CalcBitCount32(unsigned b)
{
    b = (b & 0x55555555) + (b >> 1 & 0x55555555);
    b = (b & 0x33333333) + (b >> 2 & 0x33333333);
    b = (b + (b >> 4)) & 0x0F0F0F0F;
    b = b + (b >> 8);
    b = (b + (b >> 16)) & 0x0000003F;
    return b;
}

Definition at line 64 of file bmsse2.h.

References BM_ALIGN16, and BM_ALIGN16ATTR.

BMFORCEINLINE void bm::sse2_copy_block	(	__m128i *BMRESTRICT	dst,
		const __m128i *BMRESTRICT	src,
		const __m128i *BMRESTRICT	src_end
	)

SSE2 block copy dst = *src.

Definition at line 322 of file bmsse_util.h.

BMFORCEINLINE void bm::sse2_invert_arr	(	bm::word_t *	first,
		bm::word_t *	last
	)

Invert array elements dst = ~*dst or dst ^= *dst.

Definition at line 366 of file bmsse_util.h.

BMFORCEINLINE void bm::sse2_or_arr	(	__m128i *BMRESTRICT	dst,
		const __m128i *BMRESTRICT	src,
		const __m128i *BMRESTRICT	src_end
	)

OR array elements against another array dst |= *src.

Definition at line 167 of file bmsse_util.h.

BMFORCEINLINE void bm::sse2_set_block	(	__m128i *BMRESTRICT	dst,
		__m128i *BMRESTRICT	dst_end,
		bm::word_t	value
	)

SSE2 block memset dst = value.

Definition at line 288 of file bmsse_util.h.

BMFORCEINLINE void bm::sse2_sub_arr	(	__m128i *BMRESTRICT	dst,
		const __m128i *BMRESTRICT	src,
		const __m128i *BMRESTRICT	src_end
	)

AND-NOT (SUB) array elements against another array dst &= ~*src.

Definition at line 248 of file bmsse_util.h.

BMFORCEINLINE void bm::sse2_xor_arr	(	__m128i *BMRESTRICT	dst,
		const __m128i *BMRESTRICT	src,
		const __m128i *BMRESTRICT	src_end
	)

OR array elements against another array dst ^= *src.

Definition at line 207 of file bmsse_util.h.

BMFORCEINLINE void bm::sse2_xor_arr_2_mask	(	__m128i *BMRESTRICT	dst,
		const __m128i *BMRESTRICT	src,
		const __m128i *BMRESTRICT	src_end,
		bm::word_t	mask
	)

XOR array elements to specified mask dst = *src ^ mask.

Definition at line 72 of file bmsse_util.h.