vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

change to one time init to prevent performance drop

Browse source
This commit is contained in:
domke 2024-07-27 02:13:24 +09:00
parent 7cc64329d7
commit ab8b748188
5 changed files with 36 additions and 18 deletions
Download patch file Download diff file Expand all files Collapse all files

10
examples/gguf-hash/deps/xxhash/xxhash.h
View file

@ -3739,6 +3739,8 @@ XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(XXH_NOESCAPE const XXH64_can
#if defined(__GNUC__) || defined(__clang__)
# if defined(__ARM_FEATURE_SVE)
# include <arm_sve.h>
# include <sys/prctl.h>
int sve_cnt_b = 0;
# endif
# if defined(__ARM_NEON__) || defined(__ARM_NEON) \
|| (defined(_M_ARM) && _M_ARM >= 7) \
@ -5495,7 +5497,9 @@ XXH3_accumulate_512_sve( void* XXH_RESTRICT acc,
const uint64_t *xinput = (const uint64_t *)(const void *)input;
const uint64_t *xsecret = (const uint64_t *)(const void *)secret;
svuint64_t kSwap = sveor_n_u64_z(svptrue_b64(), svindex_u64(0, 1), 1);
uint64_t element_count = svcntd();
if (!sve_cnt_b)
sve_cnt_b = PR_SVE_VL_LEN_MASK & prctl(PR_SVE_GET_VL);
uint64_t element_count = sve_cnt_b / 8;
if (element_count >= 8) {
svbool_t mask = svptrue_pat_b64(SV_VL8);
svuint64_t vacc = svld1_u64(mask, xacc);
@ -5537,7 +5541,9 @@ XXH3_accumulate_sve(xxh_u64* XXH_RESTRICT acc,
const uint64_t *xinput = (const uint64_t *)(const void *)input;
const uint64_t *xsecret = (const uint64_t *)(const void *)secret;
svuint64_t kSwap = sveor_n_u64_z(svptrue_b64(), svindex_u64(0, 1), 1);
uint64_t element_count = svcntd();
if (!sve_cnt_b)
sve_cnt_b = PR_SVE_VL_LEN_MASK & prctl(PR_SVE_GET_VL);
uint64_t element_count = sve_cnt_b / 8;
if (element_count >= 8) {
svbool_t mask = svptrue_pat_b64(SV_VL8);
svuint64_t vacc = svld1_u64(mask, xacc + 0);

28
ggml/src/ggml-aarch64.c
View file

@ -384,8 +384,8 @@ void ggml_gemv_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void *
UNUSED(blocklen);
#if defined(__ARM_FEATURE_SVE)
if (svcntw() == 8) {
GGML_ASSERT(!(ggml_cpu_has_sve() && (svcntw() == 8)) &&
if (sve_cnt_b == QK8_0) {
GGML_ASSERT(!(ggml_cpu_has_sve() && (sve_cnt_b == QK8_0)) &&
"__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance");
}
#endif
@ -496,8 +496,8 @@ void ggml_gemv_q4_0_4x8_q8_0(int n, float * restrict s, size_t bs, const void *
UNUSED(blocklen);
#if defined(__ARM_FEATURE_SVE)
if (svcntw() == 8) {
GGML_ASSERT(!(ggml_cpu_has_sve() && (svcntw() == 8)) &&
if (sve_cnt_b == QK8_0) {
GGML_ASSERT(!(ggml_cpu_has_sve() && (sve_cnt_b == QK8_0)) &&
"__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance");
}
#endif
@ -614,7 +614,7 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
UNUSED(blocklen);
#if defined(__ARM_FEATURE_SVE) && ! ((defined(_MSC_VER)) && ! defined(__clang__))
if (svcntw() == 8) {
if (sve_cnt_b == QK8_0) {
const void * b_ptr = vx;
const void * a_ptr = vy;
float * res_ptr = s;
@ -680,12 +680,12 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
return;
}
else if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) {
GGML_ASSERT((ggml_cpu_has_sve() && (svcntw() == 8)) &&
GGML_ASSERT((ggml_cpu_has_sve() && (sve_cnt_b == QK8_0)) &&
"__ARM_FEATURE_SVE for vector size of 256-bits not defined, use the Q4_0_4_8 quantization format for optimal "
"performance");
}
else if (ggml_cpu_has_neon()) {
GGML_ASSERT(((ggml_cpu_has_sve() && (svcntw() == 8)) || ggml_cpu_has_matmul_int8()) &&
GGML_ASSERT(((ggml_cpu_has_sve() && (sve_cnt_b == QK8_0)) || ggml_cpu_has_matmul_int8()) &&
"__ARM_FEATURE_SVE for vector size of 256-bits and __ARM_FEATURE_MATMUL_INT8 not defined, use the Q4_0_4_4 "
"quantization format for optimal performance");
}
@ -745,8 +745,8 @@ void ggml_gemm_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void *
UNUSED(blocklen);
#if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8)
if (svcntw() == 8) {
GGML_ASSERT(!(ggml_cpu_has_sve() && (svcntw() == 8)) &&
if (sve_cnt_b == QK8_0) {
GGML_ASSERT(!(ggml_cpu_has_sve() && (sve_cnt_b == QK8_0)) &&
"__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance");
}
#endif
@ -1266,8 +1266,8 @@ void ggml_gemm_q4_0_4x8_q8_0(int n, float * restrict s, size_t bs, const void *
UNUSED(blocklen);
#if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8)
if (svcntw() == 8) {
GGML_ASSERT(!(ggml_cpu_has_sve() && (svcntw() == 8)) &&
if (sve_cnt_b == QK8_0) {
GGML_ASSERT(!(ggml_cpu_has_sve() && (sve_cnt_b == QK8_0)) &&
"__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance");
}
#endif
@ -1728,7 +1728,7 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
UNUSED(blocklen);
#if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8) && ! ((defined(_MSC_VER)) && ! defined(__clang__))
if (svcntw() == 8) {
if (sve_cnt_b == QK8_0) {
const void * b_ptr = vx;
const void * a_ptr = vy;
float * res_ptr = s;
@ -2139,12 +2139,12 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
return;
}
else if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) {
GGML_ASSERT((ggml_cpu_has_sve() && (svcntw() == 8)) &&
GGML_ASSERT((ggml_cpu_has_sve() && (sve_cnt_b == QK8_0)) &&
"__ARM_FEATURE_SVE for vector size of 256-bits not defined, use the Q4_0_4_8 quantization format for optimal "
"performance");
}
else if (ggml_cpu_has_neon()) {
GGML_ASSERT(((ggml_cpu_has_sve() && (svcntw() == 8)) || ggml_cpu_has_matmul_int8()) &&
GGML_ASSERT(((ggml_cpu_has_sve() && (sve_cnt_b == QK8_0)) || ggml_cpu_has_matmul_int8()) &&
"__ARM_FEATURE_SVE for vector size of 256-bits and __ARM_FEATURE_MATMUL_INT8 not defined, use the Q4_0_4_4 "
"quantization format for optimal performance");
}

4
ggml/src/ggml-quants.c
View file

@ -3818,7 +3818,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r
float sumf = 0;
#if defined(__ARM_FEATURE_SVE)
if ((PR_SVE_VL_LEN_MASK & prctl(PR_SVE_GET_VL)) == QK8_0) {
if (sve_cnt_b == QK8_0) {
const svbool_t ptrueh = svptrue_pat_b8(SV_VL16);
const svbool_t ptruel = svnot_b_z(svptrue_b8(), ptrueh);
@ -5291,7 +5291,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r
float sumf = 0;
#if defined(__ARM_FEATURE_SVE)
if ((PR_SVE_VL_LEN_MASK & prctl(PR_SVE_GET_VL)) == QK8_0) {
if (sve_cnt_b == QK8_0) {
svfloat32_t sumv0 = svdup_n_f32(0.0f);
svfloat32_t sumv1 = svdup_n_f32(0.0f);

4
ggml/src/ggml-quants.h
View file

@ -127,6 +127,10 @@ void iq2xs_free_impl(enum ggml_type type);
void iq3xs_init_impl(int grid_size);
void iq3xs_free_impl(int grid_size);
#if defined(__ARM_FEATURE_SVE)
extern int sve_cnt_b;
#endif
#ifdef __cplusplus
}
#endif

8
ggml/src/ggml.c
View file

@ -37,6 +37,9 @@
#include <unistd.h>
#endif
#if defined(__ARM_FEATURE_SVE)
int sve_cnt_b = 0;
#endif
#if defined(__ARM_FEATURE_SVE) || defined(__ARM_FEATURE_MATMUL_INT8)
#undef GGML_USE_LLAMAFILE
#endif
@ -3475,6 +3478,11 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {
ggml_assert_aligned(ctx->mem_buffer);
#if defined(__ARM_FEATURE_SVE)
if (!sve_cnt_b)
sve_cnt_b = PR_SVE_VL_LEN_MASK & prctl(PR_SVE_GET_VL);
#endif
GGML_PRINT_DEBUG("%s: context initialized\n", __func__);
ggml_critical_section_end();