github/MP-SPDZ
1
1
Fork 0
mirror of https://github.com/data61/MP-SPDZ.git synced 2026-01-09 13:37:58 -05:00
Code Issues Packages Projects Releases Wiki Activity

Optimized matrix multiplication; data type for quantization.

Browse Source
This commit is contained in:
Marcel Keller
2019-02-14 15:15:37 +11:00
parent 216fbdf1e1
commit b6a18675e8
108 changed files with 2676 additions and 999 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
GC/square64.cpp
View File

@@ -13,7 +13,7 @@ union matrix32x8
__m256i whole;
octet rows[32];
matrix32x8(__m256i x = _mm256_setzero_si256()) : whole(x) {}
matrix32x8(const __m256i& x = _mm256_setzero_si256()) : whole(x) {}
matrix32x8(square64& input, int x, int y)
{
@@ -23,6 +23,7 @@ union matrix32x8
void transpose(square64& output, int x, int y)
{
#ifdef __AVX2__
for (int j = 0; j < 8; j++)
{
int row = _mm256_movemask_epi8(whole);
@@ -31,6 +32,10 @@ union matrix32x8
// _mm_movemask_epi8 uses most significant bit, hence +7-j
output.halfrows[8*x+7-j][y] = row;
}
#else
(void) output, (void) x, (void) y;
throw runtime_error("need to compile with AVX2 support");
#endif
}
};
@@ -51,8 +56,10 @@ case I: \
HIGHS = _mm256_unpackhi_epi##I(A, B); \
break;
void zip(int chunk_size, __m256i& lows, __m256i& highs, __m256i a, __m256i b)
void zip(int chunk_size, __m256i& lows, __m256i& highs,
const __m256i& a, const __m256i& b)
{
#ifdef __AVX2__
switch (chunk_size)
{
ZIP_CASE(8, lows, highs, a, b);
@@ -67,6 +74,10 @@ void zip(int chunk_size, __m256i& lows, __m256i& highs, __m256i a, __m256i b)
default:
throw invalid_argument("not supported");
}
#else
(void) chunk_size, (void) lows, (void) highs, (void) a, (void) b;
throw runtime_error("need to compile with AVX2 support");
#endif
}
void square64::transpose(int n_rows, int n_cols)