Last active
June 24, 2022 21:39
-
-
Save BrettRToomey/dda02ca7c46bd889bb3760c8e9e2f9ac to your computer and use it in GitHub Desktop.
An example showing how to trim an array of integers using a boundary in SIMD
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
// Credit for this algorithm goes to @pervognsen. | |
// I just took his pseudocode and turned it into C | |
// | |
// NOTE: this code is not "production safe" and makes assumptions about | |
// the input (like 'count' being divisible by 4). | |
#include <tmmintrin.h> | |
#include <emmintrin.h> | |
#include <popcntintrin.h> | |
#define SLOT(n) (n*4+0), (n*4+1), (n*4+2), (n*4+3) | |
#define _ 128, 128, 128, 128 | |
// this table maps a 16-bit 'movmskps' mask into a 'pshufb' mask | |
char sse_compact_lut[16][16] = { | |
/* 0b0000 */ { _, _, _, _ }, | |
/* 0b0001 */ { SLOT(0), _, _, _ }, | |
/* 0b0010 */ { SLOT(1), _, _, _ }, | |
/* 0b0011 */ { SLOT(0), SLOT(1), _, _ }, | |
/* 0b0100 */ { SLOT(2), _, _, _ }, | |
/* 0b0101 */ { SLOT(0), SLOT(2), _, _ }, | |
/* 0b0110 */ { SLOT(1), SLOT(2), _, _ }, | |
/* 0b0111 */ { SLOT(0), SLOT(1), SLOT(2), _ }, | |
/* 0b1000 */ { SLOT(3), _, _, _ }, | |
/* 0b1001 */ { SLOT(0), SLOT(3), _, _ }, | |
/* 0b1010 */ { SLOT(1), SLOT(3), _, _ }, | |
/* 0b1011 */ { SLOT(0), SLOT(1), SLOT(3), _ }, | |
/* 0b1100 */ { SLOT(2), SLOT(3), _, _ }, | |
/* 0b1101 */ { SLOT(0), SLOT(2), SLOT(3), _ }, | |
/* 0b1110 */ { SLOT(1), SLOT(2), SLOT(3), _ }, | |
/* 0b1111 */ { SLOT(0), SLOT(1), SLOT(2), SLOT(3) }, | |
}; | |
#undef SLOT | |
#undef _ | |
void trim_using_lower_bound_sse( int *src, int count, int *dst, int bound ) | |
{ | |
int *end = src + count; | |
__m128i b = _mm_set1_epi32(bound); | |
__m128i *lut = (__m128i *) sse_compact_lut; | |
while (src < end) { | |
__m128i nums = _mm_loadu_si128((__m128i_u *) src); | |
int mask = _mm_movemask_ps(_mm_cmplt_epi32(nums, b)); | |
__m128i res = _mm_shuffle_epi8(nums, lut[mask]); | |
_mm_storeu_si128((__m128i_u *) dst, res); | |
dst += _mm_popcnt_u32(mask); | |
src += 4; | |
} | |
} |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment