git_clone/thorvg
1
0
Fork 0
mirror of https://github.com/thorvg/thorvg.git synced 2025-06-14 12:04:29 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 5

sw_engine avx: the not aligned memory separated in the rasterRGBA32 func

Browse source 
The aligned and not aligned memory parts handled separately.
This gives around 10% of the performance increase (tested on i7-8700 CPU)
This commit is contained in:
Mira Grudzinska 2021-08-27 01:50:00 +02:00 • committed by Hermet Park
parent 3ded3e359d
commit dafc229a75
1 changed files with 19 additions and 11 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

30
src/lib/sw_engine/tvgSwRasterAvx.h
View file

@ -25,6 +25,7 @@
#include <immintrin.h> #include <immintrin.h>
#define N_32BITS_IN_128REG 4 #define N_32BITS_IN_128REG 4
#define N_32BITS_IN_256REG 8
static inline __m128i ALPHA_BLEND(__m128i c, __m128i a) static inline __m128i ALPHA_BLEND(__m128i c, __m128i a)
{ {
@ -63,24 +64,31 @@ static inline __m128i ALPHA_BLEND(__m128i c, __m128i a)
static inline void avxRasterRGBA32(uint32_t *dst, uint32_t val, uint32_t offset, int32_t len) static inline void avxRasterRGBA32(uint32_t *dst, uint32_t val, uint32_t offset, int32_t len)
{ {
//1. calculate how many iterations we need to cover the length //1. set the beginning of the array
uint32_t iterations = len / 8;
uint32_t avxFilled = iterations * 8;
//2. set the beginning of the array
dst += offset; dst += offset;
__m256i_u* avxDst = (__m256i_u*) dst;
//3. fill the octets //2. fill the not aligned memory (for a 256-bit register a 32-bytes alignment is required)
uint32_t notAligned = ((uintptr_t)dst & 0x1f) / 4;
if (notAligned) {
notAligned = (N_32BITS_IN_256REG - notAligned > (uint32_t)len ? (uint32_t)len : N_32BITS_IN_256REG - notAligned);
for (uint32_t x = 0; x < notAligned; ++x) *dst++ = val;
}
//3. calculate how many iterations we need to cover the length
uint32_t iterations = (len - notAligned) / N_32BITS_IN_256REG;
uint32_t avxFilled = iterations * N_32BITS_IN_256REG;
//4. fill the octets
__m256i avxVal = _mm256_set1_epi32(val);
auto avxDst = (__m256i*)dst;
for (uint32_t i = 0; i < iterations; ++i) { for (uint32_t i = 0; i < iterations; ++i) {
*avxDst = _mm256_set1_epi32(val); *avxDst = avxVal;
avxDst++; avxDst++;
} }
//4. fill leftovers (in the first step we have to set the pointer to the place where the avx job is done) //4. fill leftovers (in the first step set the pointer to the place where the avx job is done)
int32_t leftovers = len - avxFilled; int32_t leftovers = len - notAligned - avxFilled;
dst += avxFilled; dst += avxFilled;
while (leftovers--) *dst++ = val; while (leftovers--) *dst++ = val;
} }