Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

sycl : Optimize Q3_K mul_mat by reorder (#23725)

Browse Source
This commit is contained in:
Neo Zhang
2026-06-01 09:50:55 +03:00
committed by GitHub
parent af6528e6df
commit 44e211cecf
7 changed files with 340 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-sycl/quants.hpp
+25
View File
@@ -58,6 +58,31 @@ template <> struct block_q_t<GGML_TYPE_Q4_0> {
static constexpr int block_to_q8_1_ratio() { return traits::qk / QK8_1; }
};
template <> struct block_q_t<GGML_TYPE_Q3_K> {
struct traits {
static constexpr uint32_t qk = QK_K;
static constexpr uint32_t qi = QI3_K;
static constexpr uint32_t qr = QR3_K;
static constexpr uint32_t vdr_mmvq = 1;
};
// Reordered layout: [qs (QK_K/4 per block)] [hmask (QK_K/8 per block)] [scales] [d]
static constexpr std::pair<int, int> get_block_offset(const int block_index, const int n_blocks) {
auto qs_offset = block_index * (QK_K / 4);
auto hmask_offset = n_blocks * (QK_K / 4) + block_index * (QK_K / 8);
return { qs_offset, hmask_offset };
}
static constexpr std::pair<int, int> get_d_offset(int nrows, int ncols, const int block_index) {
auto nblocks = (nrows * (ncols / QK_K));
auto total_qs_bytes = nblocks * (QK_K / 4) + nblocks * (QK_K / 8);
return { total_qs_bytes + block_index * 12,
total_qs_bytes + nblocks * 12 + block_index * sizeof(ggml_half) };
}
static constexpr int block_to_q8_1_ratio() { return traits::qk / QK8_1; }
};
template <> struct block_q_t<GGML_TYPE_Q4_K> {
struct traits {
static constexpr uint32_t qk = QK_K;