devtools/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2026-04-10 11:34:33 +08:00
Code Issues Packages Projects Releases Wiki Activity

Consolidate multiple implementations of divup/div_up/div_ceil.

Browse Source
This commit is contained in:
Rasmus Munk Larsen
2023-10-10 17:16:59 +00:00
parent e8515f78ac
commit a96545777b
12 changed files with 71 additions and 79 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
Eigen/src/Core/arch/AVX512/GemmKernel.h
View File

@@ -90,8 +90,6 @@ class gemm_class {
const Index a_stride, b_stride;
const Index a_off, b_off;
static EIGEN_ALWAYS_INLINE constexpr int div_up(int a, int b) { return a == 0 ? 0 : (a - 1) / b + 1; }
EIGEN_ALWAYS_INLINE void prefetch_a(const Scalar *a_addr) {
_mm_prefetch((char *)(a_prefetch_size + a_addr - a_shift), _MM_HINT_T0);
}
@@ -479,7 +477,7 @@ class gemm_class {
*
* const Scalar *cox = (idx == 0) ? co1 : co2;
*
* const int um_vecs = div_up(a_unroll, nelems_in_cache_line);
* const int um_vecs = numext::div_ceil(a_unroll, nelems_in_cache_line);
* scale_load_c<0, um_vecs, idx, a_unroll>(cox, alpha_reg);
* write_c<0, um_vecs, idx, a_unroll>(cox);
*
@@ -498,7 +496,7 @@ class gemm_class {
EIGEN_ALWAYS_INLINE void c_update_1count(Scalar *&cox) {
if (pow >= 4) cox += ldc;
const int um_vecs = div_up(a_unroll, nelems_in_cache_line);
const int um_vecs = numext::div_ceil(a_unroll, nelems_in_cache_line);
auto &alpha_reg = zmm[alpha_load_reg];
scale_load_c<0, um_vecs, idx, a_unroll>(cox, alpha_reg);
@@ -644,7 +642,7 @@ class gemm_class {
template <int uk, int max_b_unroll, int a_unroll, int b_unroll, bool ktail, bool fetch_x, bool c_fetch, bool no_a_preload = false>
EIGEN_ALWAYS_INLINE void innerkernel_1uk(const Scalar *&aa, const Scalar *const &ao, const Scalar *const &bo,
Scalar *&co2, int &fetchA_idx, int &fetchB_idx) {
const int um_vecs = div_up(a_unroll, nelems_in_cache_line);
const int um_vecs = numext::div_ceil(a_unroll, nelems_in_cache_line);
if (max_b_unroll >= 1)
innerkernel_1pow<uk, 1, 0, um_vecs, b_unroll, ktail, fetch_x, c_fetch>(aa, ao, bo, co2, fetchA_idx, fetchB_idx);
@@ -729,7 +727,7 @@ class gemm_class {
template <int a_unroll, int b_unroll, int max_b_unroll>
EIGEN_ALWAYS_INLINE void kloop(const Scalar *&aa, const Scalar *&ao, const Scalar *&bo, Scalar *&co1, Scalar *&co2) {
const int um_vecs = div_up(a_unroll, nelems_in_cache_line);
const int um_vecs = numext::div_ceil(a_unroll, nelems_in_cache_line);
if (!use_less_a_regs && k > 1)
a_loads<0, 2, 0, um_vecs, a_unroll>(ao);
else