tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-06-28 20:25:20 +00:00
Code Issues Packages Projects Releases Wiki Activity

ggml-cpu: Add CPU backend support for KleidiAI library (#11390)

Browse Source 
* ggml-cpu: Add CPU backend support for KleidiAI library

* Add environmental variable GGML_KLEIDIAI_SME

* Add support for multithread LHS conversion

* Switch kernel selection order to dotprod and i8mm

* updates for review comments

* More updates for review comments

* Reorganize and rename KleidiAI files

* Move ggml-cpu-traits.h to source file

* Update cmake for SME build and add alignment for SME

* Remove append GGML_USE_CPU_KLEIDIAI to the GGML_CDEF_PUBLIC list
This commit is contained in:
Charles Xu
2025-02-20 14:06:51 +01:00
committed by GitHub
parent 4806498bf1
commit c5d91a7400
9 changed files with 767 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files

33
ggml/src/ggml-cpu/ggml-cpu.c
View File

@ -112,7 +112,8 @@ struct ggml_arm_arch_features_type {
int has_i8mm;
int has_sve;
int sve_cnt;
} ggml_arm_arch_features = {-1, -1, -1, -1, 0};
int has_sme;
} ggml_arm_arch_features = {-1, -1, -1, -1, 0, -1};
#endif
@ -2381,15 +2382,20 @@ bool ggml_is_numa(void) {
#define HWCAP2_I8MM (1 << 13)
#endif
#if !defined(HWCAP2_SME)
#define HWCAP2_SME (1 << 23)
#endif
static void ggml_init_arm_arch_features(void) {
#if defined(__linux__) && defined(__aarch64__)
uint32_t hwcap = getauxval(AT_HWCAP);
uint32_t hwcap2 = getauxval(AT_HWCAP2);
ggml_arm_arch_features.has_neon = !!(hwcap & HWCAP_ASIMD);
ggml_arm_arch_features.has_neon = !!(hwcap & HWCAP_ASIMD);
ggml_arm_arch_features.has_dotprod = !!(hwcap & HWCAP_ASIMDDP);
ggml_arm_arch_features.has_i8mm = !!(hwcap2 & HWCAP2_I8MM);
ggml_arm_arch_features.has_sve = !!(hwcap & HWCAP_SVE);
ggml_arm_arch_features.has_i8mm = !!(hwcap2 & HWCAP2_I8MM);
ggml_arm_arch_features.has_sve = !!(hwcap & HWCAP_SVE);
ggml_arm_arch_features.has_sme = !!(hwcap2 & HWCAP2_SME);
#if defined(__ARM_FEATURE_SVE)
ggml_arm_arch_features.sve_cnt = PR_SVE_VL_LEN_MASK & prctl(PR_SVE_GET_VL);
@ -2412,6 +2418,11 @@ static void ggml_init_arm_arch_features(void) {
}
ggml_arm_arch_features.has_i8mm = oldp;
if (sysctlbyname("hw.optional.arm.FEAT_SME", &oldp, &size, NULL, 0) != 0) {
oldp = 0;
}
ggml_arm_arch_features.has_sme = oldp;
ggml_arm_arch_features.has_sve = 0;
ggml_arm_arch_features.sve_cnt = 0;
#else
@ -2435,6 +2446,12 @@ static void ggml_init_arm_arch_features(void) {
ggml_arm_arch_features.has_sve = 0;
ggml_arm_arch_features.sve_cnt = 0;
#endif
#if defined(__ARM_FEATURE_SME) || defined(__ARM_FEATURE_SME2)
ggml_arm_arch_features.has_sme = 1;
#else
ggml_arm_arch_features.has_sme = 0;
#endif
#endif
}
#endif
@ -14442,6 +14459,14 @@ int ggml_cpu_get_sve_cnt(void) {
#endif
}
int ggml_cpu_has_sme(void) {
#if defined(__ARM_ARCH) && defined(__ARM_FEATURE_SME)
return ggml_arm_arch_features.has_sme;
#else
return 0;
#endif
}
void ggml_cpu_init(void) {
// needed to initialize f16 tables
{