extern/whisper.cpp
1
0
Fork 1
mirror of https://github.com/ggerganov/whisper.cpp.git synced 2025-06-23 19:21:39 +02:00
Code Issues Packages Projects Releases Wiki Activity

ggml-cpu : remove unnecesary arm feature detection (llama/14281)

Browse Source 
Support for Arm runtime feature detection has now been added to GGML_CPU_ALL_VARIANTS. This removes the old and not very functional code.
This commit is contained in:
Diego Devesa 2025-06-19 12:24:14 -07:00 committed by Georgi Gerganov
parent cef59c1e26
commit 71adde9203
2 changed files with 1004 additions and 1093 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2002
ggml/src/ggml-cpu/arch/arm/repack.cpp
View File

File diff suppressed because it is too large Load Diff

95
ggml/src/ggml-cpu/ggml-cpu.c
View File

@ -74,13 +74,8 @@
#if defined(__ARM_ARCH) #if defined(__ARM_ARCH)
struct ggml_arm_arch_features_type { struct ggml_arm_arch_features_type {
int has_neon;
int has_dotprod;
int has_i8mm;
int has_sve;
int sve_cnt; int sve_cnt;
int has_sme; } ggml_arm_arch_features = { 0 };
} ggml_arm_arch_features = {-1, -1, -1, -1, 0, -1};
#endif #endif
@ -678,87 +673,15 @@ bool ggml_is_numa(void) {
#if defined(__linux__) && defined(__aarch64__) #if defined(__linux__) && defined(__aarch64__)
#include <sys/auxv.h> #include <sys/auxv.h>
#elif defined(__APPLE__)
#include <sys/sysctl.h>
#endif
#if !defined(HWCAP2_I8MM)
#define HWCAP2_I8MM (1 << 13)
#endif
#if !defined(HWCAP2_SME)
#define HWCAP2_SME (1 << 23)
#endif #endif
static void ggml_init_arm_arch_features(void) { static void ggml_init_arm_arch_features(void) {
#if defined(__linux__) && defined(__aarch64__) #if defined(__linux__) && defined(__aarch64__) && defined(__ARM_FEATURE_SVE)
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_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_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); ggml_arm_arch_features.sve_cnt = PR_SVE_VL_LEN_MASK & prctl(PR_SVE_GET_VL);
#endif #endif
#elif defined(__APPLE__)
int oldp = 0;
size_t size = sizeof(oldp);
if (sysctlbyname("hw.optional.AdvSIMD", &oldp, &size, NULL, 0) != 0) {
oldp = 0;
}
ggml_arm_arch_features.has_neon = oldp;
if (sysctlbyname("hw.optional.arm.FEAT_DotProd", &oldp, &size, NULL, 0) != 0) {
oldp = 0;
}
ggml_arm_arch_features.has_dotprod = oldp;
if (sysctlbyname("hw.optional.arm.FEAT_I8MM", &oldp, &size, NULL, 0) != 0) {
oldp = 0;
}
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
// Run-time CPU feature detection not implemented for this platform, fallback to compile time
#if defined(__ARM_NEON)
ggml_arm_arch_features.has_neon = 1;
#else
ggml_arm_arch_features.has_neon = 0;
#endif
#if defined(__ARM_FEATURE_MATMUL_INT8)
ggml_arm_arch_features.has_i8mm = 1;
#else
ggml_arm_arch_features.has_i8mm = 0;
#endif
#if defined(__ARM_FEATURE_SVE)
ggml_arm_arch_features.has_sve = 1;
ggml_arm_arch_features.sve_cnt = 16;
#else
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
#endif // __ARM_ARCH
struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value) { struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value) {
GGML_ASSERT(!ggml_get_no_alloc(ctx)); GGML_ASSERT(!ggml_get_no_alloc(ctx));
@ -3448,7 +3371,7 @@ int ggml_cpu_has_vxe(void) {
int ggml_cpu_has_neon(void) { int ggml_cpu_has_neon(void) {
#if defined(__ARM_ARCH) && defined(__ARM_NEON) #if defined(__ARM_ARCH) && defined(__ARM_NEON)
return ggml_arm_arch_features.has_neon; return 1;
#else #else
return 0; return 0;
#endif #endif
@ -3456,7 +3379,7 @@ int ggml_cpu_has_neon(void) {
int ggml_cpu_has_dotprod(void) { int ggml_cpu_has_dotprod(void) {
#if defined(__ARM_ARCH) && defined(__ARM_FEATURE_DOTPROD) #if defined(__ARM_ARCH) && defined(__ARM_FEATURE_DOTPROD)
return ggml_arm_arch_features.has_dotprod; return 1;
#else #else
return 0; return 0;
#endif #endif
@ -3464,7 +3387,7 @@ int ggml_cpu_has_dotprod(void) {
int ggml_cpu_has_sve(void) { int ggml_cpu_has_sve(void) {
#if defined(__ARM_ARCH) && defined(__ARM_FEATURE_SVE) #if defined(__ARM_ARCH) && defined(__ARM_FEATURE_SVE)
return ggml_arm_arch_features.has_sve; return 1;
#else #else
return 0; return 0;
#endif #endif
@ -3472,7 +3395,7 @@ int ggml_cpu_has_sve(void) {
int ggml_cpu_has_matmul_int8(void) { int ggml_cpu_has_matmul_int8(void) {
#if defined(__ARM_ARCH) && defined(__ARM_FEATURE_MATMUL_INT8) #if defined(__ARM_ARCH) && defined(__ARM_FEATURE_MATMUL_INT8)
return ggml_arm_arch_features.has_i8mm; return 1;
#else #else
return 0; return 0;
#endif #endif
@ -3488,7 +3411,7 @@ int ggml_cpu_get_sve_cnt(void) {
int ggml_cpu_has_sme(void) { int ggml_cpu_has_sme(void) {
#if defined(__ARM_ARCH) && defined(__ARM_FEATURE_SME) #if defined(__ARM_ARCH) && defined(__ARM_FEATURE_SME)
return ggml_arm_arch_features.has_sme; return 1;
#else #else
return 0; return 0;
#endif #endif