ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-04-30 19:31:48 +00:00
Code Issues Packages Projects Releases Wiki Activity

iqk_mul_mat: fix q8_0

Browse Source 
I was happily using _mm256_packs_epi32() to pack the
q8_0 x q8_0 dot products back to int16_t, and getting useful
results. But theoretically this can overflow, so it is
better to use _mm256_unpacklo_ and _mm256_unpackhi_ to combine
the 4 dot products using int32_t additions. This is (almost)
as fast, unlike _mm256_hadd_epi32(), which seems excessively
slow on the Ryzen-7950X.
This commit is contained in:
Kawrakow
2024-06-08 13:47:02 +03:00
parent 7f91151c2e
commit 5812618409
1 changed files with 42 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files

56
iqk_mul_mat.cpp
View File

@@ -1746,19 +1746,41 @@ struct UnsignedDot {
return helper.dot(x, y); return helper.dot(x, y);
} }
}; };
template <typename Q8, typename Q8x4, typename Dot> struct Sum4 {
template <typename Q8, typename Q8x4, typename Dot, bool can_pack = true> struct Sum4 {
Dot dot; Dot dot;
inline __m256i compute(const __m256i * qx, const Q8 * y) const { inline __m256i compute(const __m256i * qx, const Q8 * y) const {
const Q8x4 * y4 = (const Q8x4 *)y; const Q8x4 * y4 = (const Q8x4 *)y;
const __m256i p0 = dot.compute(qx[0], _mm256_loadu_si256((const __m256i *)y4->qs+0)); const __m256i p0 = dot.compute(qx[0], _mm256_loadu_si256((const __m256i *)y4->qs+0)); // 8x block 0
const __m256i p1 = dot.compute(qx[1], _mm256_loadu_si256((const __m256i *)y4->qs+1)); const __m256i p1 = dot.compute(qx[1], _mm256_loadu_si256((const __m256i *)y4->qs+1)); // 8x block 1
const __m256i p2 = dot.compute(qx[2], _mm256_loadu_si256((const __m256i *)y4->qs+2)); const __m256i p2 = dot.compute(qx[2], _mm256_loadu_si256((const __m256i *)y4->qs+2)); // 8x block 2
const __m256i p3 = dot.compute(qx[3], _mm256_loadu_si256((const __m256i *)y4->qs+3)); const __m256i p3 = dot.compute(qx[3], _mm256_loadu_si256((const __m256i *)y4->qs+3)); // 8x block 3
const __m256i p01 = _mm256_madd_epi16(dot.helper.m1, _mm256_packs_epi32(p0, p1)); // 0,0, 1,1, 0,0, 1,1 if constexpr (can_pack) {
const __m256i p23 = _mm256_madd_epi16(dot.helper.m1, _mm256_packs_epi32(p2, p3)); // 2,2, 3,3, 2,2, 3,3 const __m256i p01 = _mm256_madd_epi16(dot.helper.m1, _mm256_packs_epi32(p0, p1)); // 0,0, 1,1, 0,0, 1,1
return _mm256_madd_epi16(dot.helper.m1, _mm256_packs_epi32(p01, p23)); // 0,1,2,3, 0,1,2,3 const __m256i p23 = _mm256_madd_epi16(dot.helper.m1, _mm256_packs_epi32(p2, p3)); // 2,2, 3,3, 2,2, 3,3
return _mm256_madd_epi16(dot.helper.m1, _mm256_packs_epi32(p01, p23)); // 0,1,2,3, 0,1,2,3
} else {
// Note to myself: this is much faster than using _mm256_hadd_epi32()
auto p01 = _mm256_add_epi32(_mm256_unpacklo_epi32(p0, p1), _mm256_unpackhi_epi32(p0, p1)); // 0,1, 0,1, 0,1, 0,1
auto p23 = _mm256_add_epi32(_mm256_unpacklo_epi32(p2, p3), _mm256_unpackhi_epi32(p2, p3)); // 2,3, 2,3, 2,3, 2,3
return _mm256_add_epi32(_mm256_unpacklo_epi64(p01, p23), _mm256_unpackhi_epi64(p01, p23)); // 0,1,2,3, 0,1,2,3
}
} }
}; };
// If I use this, it negatively impacts q4_1/q5_1 performance.
//template <typename Q8, typename Q8x4, typename Dot> struct Sum4 {
// Dot dot;
// inline __m256i compute(const __m256i * qx, const Q8 * y) const {
// const Q8x4 * y4 = (const Q8x4 *)y;
// const __m256i p0 = dot.compute(qx[0], _mm256_loadu_si256((const __m256i *)y4->qs+0)); // 8x block 0
// const __m256i p1 = dot.compute(qx[1], _mm256_loadu_si256((const __m256i *)y4->qs+1)); // 8x block 1
// const __m256i p2 = dot.compute(qx[2], _mm256_loadu_si256((const __m256i *)y4->qs+2)); // 8x block 2
// const __m256i p3 = dot.compute(qx[3], _mm256_loadu_si256((const __m256i *)y4->qs+3)); // 8x block 3
// auto p01 = _mm256_add_epi32(_mm256_unpacklo_epi32(p0, p1), _mm256_unpackhi_epi32(p0, p1)); // 0,1, 0,1, 0,1, 0,1
// auto p23 = _mm256_add_epi32(_mm256_unpacklo_epi32(p2, p3), _mm256_unpackhi_epi32(p2, p3)); // 2,3, 2,3, 2,3, 2,3
// return _mm256_add_epi32(_mm256_unpacklo_epi64(p01, p23), _mm256_unpackhi_epi64(p01, p23)); // 0,1,2,3, 0,1,2,3
// }
//};
struct ScaleHelperQ8_0 { struct ScaleHelperQ8_0 {
inline __m128 prepare4(const block_q8_0 * y) { inline __m128 prepare4(const block_q8_0 * y) {
@@ -1908,11 +1930,12 @@ using AccumType1 = AccumT<MinusType1<nrc_y>, nrc_y, is_multiple_of_4>;
using Sum4Type0 = Sum4<block_q8_0, block_q8_0_x4, SignedDot>; using Sum4Type0 = Sum4<block_q8_0, block_q8_0_x4, SignedDot>;
using Sum4Type1 = Sum4<block_q8_1, block_q8_1_x4, UnsignedDot>; using Sum4Type1 = Sum4<block_q8_1, block_q8_1_x4, UnsignedDot>;
using Sum4TypeQ80 = Sum4<block_q8_0, block_q8_0_x4, SignedDot, false>;
template <typename Unpacker, typename Sum4Type, typename AccumType, typename Scales, typename Q8, int nrc_y> template <typename Unpacker, typename AccumType, typename Scales, typename Q8, int nrc_y>
void mul_mat_qX_q8_Helper(int nb, const void * vx, size_t bx, const DataInfo& info, const Q8 ** y, int nrc_x) { void mul_mat_qX_q8_Helper(int nb, const void * vx, size_t bx, const DataInfo& info, const Q8 ** y, int nrc_x) {
Unpacker unp(vx, bx); Unpacker unp(vx, bx);
Sum4Type sum4; typename Unpacker::Sum4T sum4;
Scales scales; Scales scales;
for (int ix = 0; ix < nrc_x; ++ix) { for (int ix = 0; ix < nrc_x; ++ix) {
unp.set_row(ix); unp.set_row(ix);
@@ -1927,11 +1950,11 @@ void mul_mat_qX_0_q8_0_T(int n, const void * vx, size_t bx, const DataInfo& info
Q8<nrc_y, block_q8_0> q8(info); Q8<nrc_y, block_q8_0> q8(info);
int nb = n/Unpacker::block_size(); int nb = n/Unpacker::block_size();
if (nb%4 == 0) { if (nb%4 == 0) {
mul_mat_qX_q8_Helper<Unpacker, Sum4Type0, AccumType0<nrc_y, true>, ScaleHelperQ8_0, block_q8_0, nrc_y>( mul_mat_qX_q8_Helper<Unpacker, AccumType0<nrc_y, true>, ScaleHelperQ8_0, block_q8_0, nrc_y>(
nb, vx, bx, info, q8.y, nrc_x nb, vx, bx, info, q8.y, nrc_x
); );
} else { } else {
mul_mat_qX_q8_Helper<Unpacker, Sum4Type0, AccumType0<nrc_y, false>, ScaleHelperQ8_0, block_q8_0, nrc_y>( mul_mat_qX_q8_Helper<Unpacker, AccumType0<nrc_y, false>, ScaleHelperQ8_0, block_q8_0, nrc_y>(
nb, vx, bx, info, q8.y, nrc_x nb, vx, bx, info, q8.y, nrc_x
); );
} }
@@ -1943,11 +1966,11 @@ void mul_mat_qX_1_q8_1_T(int n, const void * vx, size_t bx, const DataInfo& info
Q8<nrc_y, block_q8_1> q8(info); Q8<nrc_y, block_q8_1> q8(info);
int nb = n/Unpacker::block_size(); int nb = n/Unpacker::block_size();
if (nb%4 == 0) { if (nb%4 == 0) {
mul_mat_qX_q8_Helper<Unpacker, Sum4Type1, AccumType1<nrc_y, true>, ScaleHelperQ8_1, block_q8_1, nrc_y>( mul_mat_qX_q8_Helper<Unpacker, AccumType1<nrc_y, true>, ScaleHelperQ8_1, block_q8_1, nrc_y>(
nb, vx, bx, info, q8.y, nrc_x nb, vx, bx, info, q8.y, nrc_x
); );
} else { } else {
mul_mat_qX_q8_Helper<Unpacker, Sum4Type1, AccumType1<nrc_y, false>, ScaleHelperQ8_1, block_q8_1, nrc_y>( mul_mat_qX_q8_Helper<Unpacker, AccumType1<nrc_y, false>, ScaleHelperQ8_1, block_q8_1, nrc_y>(
nb, vx, bx, info, q8.y, nrc_x nb, vx, bx, info, q8.y, nrc_x
); );
} }
@@ -2050,22 +2073,27 @@ struct Q_Unpacker {
struct Q8_0_Unpacker final : public Q_Unpacker<block_q8_0, ScaleHelperQ_0, Q8_0_Dequantizer> { struct Q8_0_Unpacker final : public Q_Unpacker<block_q8_0, ScaleHelperQ_0, Q8_0_Dequantizer> {
Q8_0_Unpacker(const void * vx, size_t bx) : Q_Unpacker(vx, bx) {} Q8_0_Unpacker(const void * vx, size_t bx) : Q_Unpacker(vx, bx) {}
using Sum4T = Sum4TypeQ80;
inline static int block_size() { return QK8_0; } inline static int block_size() { return QK8_0; }
}; };
struct Q4_0_Unpacker final : public Q_Unpacker<block_q4_0, ScaleHelperQ_0, Q4_0_Dequantizer> { struct Q4_0_Unpacker final : public Q_Unpacker<block_q4_0, ScaleHelperQ_0, Q4_0_Dequantizer> {
Q4_0_Unpacker(const void * vx, size_t bx) : Q_Unpacker(vx, bx) {} Q4_0_Unpacker(const void * vx, size_t bx) : Q_Unpacker(vx, bx) {}
using Sum4T = Sum4TypeQ80;
inline static int block_size() { return QK4_0; } inline static int block_size() { return QK4_0; }
}; };
struct Q5_0_Unpacker final : public Q_Unpacker<block_q5_0, ScaleHelperQ_0, Q5_0_Dequantizer> { struct Q5_0_Unpacker final : public Q_Unpacker<block_q5_0, ScaleHelperQ_0, Q5_0_Dequantizer> {
Q5_0_Unpacker(const void * vx, size_t bx) : Q_Unpacker(vx, bx) {} Q5_0_Unpacker(const void * vx, size_t bx) : Q_Unpacker(vx, bx) {}
using Sum4T = Sum4TypeQ80;
inline static int block_size() { return QK5_0; } inline static int block_size() { return QK5_0; }
}; };
struct Q4_1_Unpacker final : public Q_Unpacker<block_q4_1, ScaleHelperQ_1, Q4_1_Dequantizer> { struct Q4_1_Unpacker final : public Q_Unpacker<block_q4_1, ScaleHelperQ_1, Q4_1_Dequantizer> {
Q4_1_Unpacker(const void * vx, size_t bx) : Q_Unpacker(vx, bx) {} Q4_1_Unpacker(const void * vx, size_t bx) : Q_Unpacker(vx, bx) {}
using Sum4T = Sum4Type1;
inline static int block_size() { return QK4_1; } inline static int block_size() { return QK4_1; }
}; };
struct Q5_1_Unpacker final : public Q_Unpacker<block_q5_1, ScaleHelperQ_1, Q5_1_Dequantizer> { struct Q5_1_Unpacker final : public Q_Unpacker<block_q5_1, ScaleHelperQ_1, Q5_1_Dequantizer> {
Q5_1_Unpacker(const void * vx, size_t bx) : Q_Unpacker(vx, bx) {} Q5_1_Unpacker(const void * vx, size_t bx) : Q_Unpacker(vx, bx) {}
using Sum4T = Sum4Type1;
inline static int block_size() { return QK4_1; } inline static int block_size() { return QK4_1; }
}; };