tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-07-28 21:23:55 -04:00
Code Issues Packages Projects Releases Wiki Activity

ggml-cpu: Integrate fp32=bf16xbf16 SME KleidiAI kernel (#13053)

Browse Source 
* ggml-cpu: Integrate fp32=bf16xbf16 SME KleidiAI kernel

Signed-off-by: Dan Johansson <dan.johansson@arm.com>

* * code review fixes

Signed-off-by: Dan Johansson <dan.johansson@arm.com>

* * adds a comment that clarifies barrier usage

Signed-off-by: Dan Johansson <dan.johansson@arm.com>

---------

Signed-off-by: Dan Johansson <dan.johansson@arm.com>
Co-authored-by: Charles Xu <charles.xu@arm.com>
This commit is contained in:
Dan Johansson
2025-05-12 13:06:19 +02:00
committed by GitHub
parent 95e18884fc
commit a71a4075cd
4 changed files with 416 additions and 99 deletions
Download Patch File Download Diff File Expand all files Collapse all files

93
ggml/src/ggml-cpu/kleidiai/kernels.cpp
View File

@@ -4,16 +4,22 @@
// KleidiAI micro-kernels
#include "kai_matmul_clamp_f32_qsi8d32p_qsi4c32p_interface.h"
#include "kai_lhs_quant_pack_qsi8d32p_f32.h"
#include "kai_lhs_quant_pack_qsi8d32p_f32_neon.h"
#include "kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.h"
#include "kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.h"
#include "kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.h"
#include "kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.h"
#include "kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.h"
#include "kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm.h"
#include "kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.h"
#include "kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.h"
#include "kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.h"
#include "kai_lhs_pack_bf16p2vlx2_f32_sme.h"
#include "kai_lhs_quant_pack_qsi8d32p_f32.h"
#include "kai_lhs_quant_pack_qsi8d32p_f32_neon.h"
#include "kai_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme.h"
#include "kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.h"
#include "kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.h"
#include "kai_common.h"
#include "kernels.h"
@@ -61,6 +67,53 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
/* .pack_func = */ kai_run_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon,
},
/* .required_cpu = */ CPU_FEATURE_SME,
/* .lhs_type = */ GGML_TYPE_F32,
/* .rhs_type = */ GGML_TYPE_Q4_0,
/* .op_type = */ GGML_TYPE_F32,
},
{
/* SME GEMM */
/* .kern_info = */ {
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_lhs_offset = */ kai_get_lhs_packed_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .run_kernel = */ kai_run_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
},
/* SME GEMV */
/* .kern_info = */ {
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_lhs_offset = */ kai_get_lhs_packed_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .run_kernel = */ kai_run_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
},
/* .lhs_info = */ {
/* .get_offset = */ kai_get_lhs_offset_lhs_pack_bf16p2vlx2_f32_sme,
/* .get_packed_offset = */ kai_get_lhs_packed_offset_lhs_pack_bf16p2vlx2_f32_sme,
/* .packed_size = */ kai_get_lhs_packed_size_lhs_pack_bf16p2vlx2_f32_sme,
/* .pack_func = */ kai_run_lhs_pack_bf16p2vlx2_f32_sme,
},
/* .rhs_info = */ {
/* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme,
/* .pack_func = */ kai_run_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme,
},
/* .required_cpu = */ CPU_FEATURE_SME,
/* .lhs_type = */ GGML_TYPE_F32,
/* .rhs_type = */ GGML_TYPE_F16,
/* .op_type = */ GGML_TYPE_F32,
},
#endif
#if defined(__APPLE__)
@@ -105,6 +158,9 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
/* .pack_func = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
},
/* .required_cpu = */ CPU_FEATURE_DOTPROD,
/* .lhs_type = */ GGML_TYPE_F32,
/* .rhs_type = */ GGML_TYPE_Q4_0,
/* .op_type = */ GGML_TYPE_F32,
},
#endif
#if defined(__ARM_FEATURE_MATMUL_INT8)
@@ -148,6 +204,9 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
/* .pack_func = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
},
/* .required_cpu = */ CPU_FEATURE_DOTPROD | CPU_FEATURE_I8MM,
/* .lhs_type = */ GGML_TYPE_F32,
/* .rhs_type = */ GGML_TYPE_Q4_0,
/* .op_type = */ GGML_TYPE_F32,
},
#endif
#else
@@ -192,6 +251,9 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
/* .pack_func = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
},
/* .required_cpu = */ CPU_FEATURE_DOTPROD | CPU_FEATURE_I8MM,
/* .lhs_type = */ GGML_TYPE_F32,
/* .rhs_type = */ GGML_TYPE_Q4_0,
/* .op_type = */ GGML_TYPE_F32,
},
#endif
#if defined(__ARM_FEATURE_DOTPROD)
@@ -235,12 +297,33 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
/* .pack_func = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
},
/* .required_cpu = */ CPU_FEATURE_DOTPROD,
/* .lhs_type = */ GGML_TYPE_F32,
/* .rhs_type = */ GGML_TYPE_Q4_0,
/* .op_type = */ GGML_TYPE_F32,
},
#endif
#endif
};
ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature features) {
ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature cpu_features, const ggml_tensor * tensor) {
ggml_kleidiai_kernels * kernel = nullptr;
if (tensor->op == GGML_OP_MUL_MAT && tensor->src[0] != nullptr && tensor->src[1] != nullptr) {
for (size_t i = 0; i < NELEMS(gemm_gemv_kernels); ++i) {
if ((cpu_features & gemm_gemv_kernels[i].required_cpu) == gemm_gemv_kernels[i].required_cpu &&
gemm_gemv_kernels[i].lhs_type == tensor->src[1]->type &&
gemm_gemv_kernels[i].rhs_type == tensor->src[0]->type &&
gemm_gemv_kernels[i].op_type == tensor->type) {
kernel = &gemm_gemv_kernels[i];
break;
}
}
}
return kernel;
}
ggml_kleidiai_kernels * ggml_kleidiai_select_kernels_q4_0(cpu_feature features) {
ggml_kleidiai_kernels * kernels = nullptr;
for (size_t i = 0; i < NELEMS(gemm_gemv_kernels); ++i) {