CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-04 09:32:00 +00:00
Code Issues Packages Projects Releases Wiki Activity

sycl: reordered Q4_K MMVQ (#13109)

Browse Source
This commit is contained in:
Svetlozar Georgiev
2025-05-15 16:35:44 +01:00
committed by GitHub
parent 9c404ed54c
commit 64bb51cf90
7 changed files with 280 additions and 84 deletions
Download Patch File Download Diff File Expand all files Collapse all files

31
ggml/src/ggml-sycl/mmvq.cpp
View File

@@ -24,6 +24,7 @@ static void mul_mat_vec_q_reorder(const void * __restrict__ vx, const void * __r
const int blocks_per_row = ncols / block_traits::qk;
constexpr int blocks_per_subgroup = ceil_div(block_traits::vdr_mmvq * WARP_SIZE, block_traits::qi);
constexpr int block_elements_per_subgroup = block_traits::qi / block_traits::vdr_mmvq;
const int nblocks = nrows * (ncols / block_traits::qk);
static_assert(blocks_per_subgroup > 0);
static_assert(block_elements_per_subgroup > 0);
@@ -45,7 +46,7 @@ static void mul_mat_vec_q_reorder(const void * __restrict__ vx, const void * __r
// x block quant index when casting the quants to int
const int iqs = elem + block_traits::vdr_mmvq * (sg.get_local_linear_id() % block_elements_per_subgroup);
partial_sum += reorder_vec_dot_q_sycl()(vx, bx_offset, d_offset, &y[iby], iqs);
partial_sum += reorder_vec_dot_q_sycl()(vx, bx_offset, d_offset, &y[iby], iqs, nblocks);
}
}
@@ -739,6 +740,27 @@ static void mul_mat_vec_q4_K_q8_1_sycl(const void *vx, const void *vy,
}
}
static void reorder_mul_mat_vec_q4_k_q8_1_sycl(const void * vx, const void * vy, float * dst, const int ncols,
const int nrows, dpct::queue_ptr stream) {
GGML_ASSERT(ncols % QK_K == 0);
const int block_num_y = ceil_div(nrows, GGML_SYCL_MMV_Y);
constexpr size_t num_subgroups = 16;
GGML_ASSERT(block_num_y % num_subgroups == 0);
const sycl::range<3> global_size(1, GGML_SYCL_MMV_Y, block_num_y * WARP_SIZE);
const sycl::range<3> workgroup_size(1, GGML_SYCL_MMV_Y, num_subgroups * WARP_SIZE);
stream->submit([&](sycl::handler & cgh) {
cgh.parallel_for(sycl::nd_range<3>(global_size, workgroup_size),
[=](sycl::nd_item<3> nd_item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
mul_mat_vec_q_reorder<reorder_vec_dot_q_sycl<GGML_TYPE_Q4_K>>(vx, vy, dst, ncols,
nrows, nd_item);
});
});
}
static void mul_mat_vec_q5_K_q8_1_sycl(const void *vx, const void *vy,
float *dst, const int ncols,
const int nrows,
@@ -1035,7 +1057,12 @@ void ggml_sycl_op_mul_mat_vec_q(ggml_backend_sycl_context & ctx, const ggml_tens
mul_mat_vec_q3_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
break;
case GGML_TYPE_Q4_K:
mul_mat_vec_q4_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
if ((ggml_tensor_extra_gpu *) dst->src[0]->extra &&
((ggml_tensor_extra_gpu *) dst->src[0]->extra)->optimized_feature.reorder) {
reorder_mul_mat_vec_q4_k_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
} else {
mul_mat_vec_q4_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
}
break;
case GGML_TYPE_Q5_K:
mul_mat_vec_q5_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);