Mirrored_Repos/LookingGlass
1
0
Fork 0
mirror of https://github.com/gnif/LookingGlass.git synced 2024-11-10 00:28:20 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

[common] rects: fix avx implementation for unaligned accesses

Browse Source
This commit is contained in:
Geoffrey McRae 2023-11-19 17:15:44 +11:00
parent 4911e129f8
commit 660b4b8ec8
1 changed files with 15 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

24
common/src/rects.c
View File

@ -328,29 +328,35 @@ static void rectCopyUnaligned_avx(
src += ystart * srcPitch + dx; src += ystart * srcPitch + dx;
dst += ystart * dstPitch + dx; dst += ystart * dstPitch + dx;
const int nvec = width / sizeof(__m256i); const int nvec = width / sizeof(__m256i);
const int rem = width % sizeof(__m256i); const int rem = width % sizeof(__m256i);
const int align = (uintptr_t)dst & 31;
for (int i = ystart; i < yend; ++i) for (int i = ystart; i < yend; ++i)
{ {
const __m256i *restrict s = (__m256i*)src; // copy the unaligned bytes
__m256i *restrict d = (__m256i*)dst; for(int col = align; col > 0; --col)
dst[col] = src[col];
const __m256i *restrict s = (__m256i*)(src + align);
__m256i *restrict d = (__m256i*)ALIGN_TO((uintptr_t)dst, 32);
int vec; int vec;
for(vec = nvec; vec > 3; vec -= 4) for(vec = nvec; vec > 3; vec -= 4)
{ {
_mm256_stream_si256(d + 0, _mm256_load_si256(s + 0)); _mm256_stream_si256(d + 0, _mm256_loadu_si256(s + 0));
_mm256_stream_si256(d + 1, _mm256_load_si256(s + 1)); _mm256_stream_si256(d + 1, _mm256_loadu_si256(s + 1));
_mm256_stream_si256(d + 2, _mm256_load_si256(s + 2)); _mm256_stream_si256(d + 2, _mm256_loadu_si256(s + 2));
_mm256_stream_si256(d + 3, _mm256_load_si256(s + 3)); _mm256_stream_si256(d + 3, _mm256_loadu_si256(s + 3));
s += 4; s += 4;
d += 4; d += 4;
} }
for(; vec > 0; --vec, ++d, ++s) for(; vec > 0; --vec, ++d, ++s)
_mm256_stream_si256(d, _mm256_load_si256(s)); _mm256_stream_si256(d, _mm256_loadu_si256(s));
// copy any remaining bytes
for(int col = width - rem; col < width; ++col) for(int col = width - rem; col < width; ++col)
dst[col] = src[col]; dst[col] = src[col];