CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-12 10:47:01 +00:00
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

58
ggml/src/ggml-cpu/kleidiai/kernels.h
View File

@@ -4,6 +4,9 @@
#pragma once
#include <functional>
#include "ggml.h"
enum cpu_feature {
CPU_FEATURE_NONE = 0,
CPU_FEATURE_DOTPROD = 1,
@@ -26,26 +29,53 @@ struct kernel_info {
size_t (*get_nr)(void);
size_t (*get_kr)(void);
size_t (*get_sr)(void);
size_t (*get_lhs_offset)(size_t m_idx, size_t k, size_t bl);
size_t (*get_rhs_packed_offset)(size_t n_idx, size_t k, size_t bl);
std::variant<
std::function<size_t(size_t n_idx, size_t k, size_t bl)>,
std::function<size_t(size_t m_idx, size_t k)>
> get_lhs_offset;
std::variant<
std::function<size_t(size_t n_idx, size_t k, size_t bl)>,
std::function<size_t(size_t n_idx, size_t k)>
> get_rhs_packed_offset;
size_t (*get_dst_offset)(size_t m_idx, size_t n_idx, size_t stride);
size_t (*get_dst_size)(size_t m, size_t n);
void (*run_kernel)(size_t m, size_t n, size_t k, size_t bl, const void* lhs_packed, const void* rhs_packed,
float* dst, size_t dst_stride_row, size_t dst_stride_col, float scalar_min, float scalar_max);
std::variant<
std::function<void(size_t m, size_t n, size_t k, size_t bl, const void* lhs_packed, const void* rhs_packed,
float* dst, size_t dst_stride_row, size_t dst_stride_col, float scalar_min, float scalar_max)>,
std::function<void(size_t m, size_t n, size_t k, const void* lhs_packed, const void* rhs_packed, void* dst, size_t dst_stride_row,
size_t dst_stride_col, float clamp_min, float clamp_max)>
> run_kernel;
};
struct lhs_packing_info {
size_t (*get_offset)(size_t m_idx, size_t lhs_stride);
size_t (*get_packed_offset)(size_t m_idx, size_t k, size_t bl, size_t mr, size_t kr, size_t sr);
size_t (*packed_size)(size_t m, size_t k, size_t bl, size_t mr, size_t kr, size_t sr);
void (*pack_func)(size_t m, size_t k, size_t bl, size_t mr, size_t kr, size_t sr, size_t m_idx_start, const float* lhs,
size_t lhs_stride, void* lhs_packed);
std::variant<
std::function<size_t(size_t m_idx, size_t k, size_t bl, size_t mr, size_t kr, size_t sr)>,
std::function<size_t(size_t m_idx, size_t k, size_t mr, size_t kr, size_t sr)>
> get_packed_offset;
std::variant<
std::function<size_t(size_t m_idx, size_t k, size_t bl, size_t mr, size_t kr, size_t sr)>,
std::function<size_t(size_t m, size_t k, size_t mr, size_t kr, size_t sr)>
> packed_size;
std::variant<
std::function<void(size_t m, size_t k, size_t bl, size_t mr, size_t kr, size_t sr, size_t m_idx_start, const float* lhs,
size_t lhs_stride, void* lhs_packed)>,
std::function<void(size_t m, size_t k, size_t mr, size_t kr, size_t sr, size_t m_idx_start, const void* lhs, size_t lhs_stride,
void* lhs_packed)>
> pack_func;
};
struct rhs_packing_info {
size_t (*packed_size)(size_t n, size_t k, size_t nr, size_t kr, size_t bl);
void (*pack_func)(size_t num_groups, size_t n, size_t k, size_t nr, size_t kr, size_t sr, size_t bl, const uint8_t* rhs,
const float* bias, void* rhs_packed, size_t extra_bytes, const struct kai_rhs_pack_qs4cxs1s0_param* params);
std::variant<
std::function<size_t(size_t n, size_t k, size_t nr, size_t kr, size_t bl)>,
std::function<size_t(size_t n, size_t k)>
> packed_size;
std::variant<
std::function<void(size_t num_groups, size_t n, size_t k, size_t nr, size_t kr, size_t sr, size_t bl, const uint8_t* rhs,
const float* bias, void* rhs_packed, size_t extra_bytes, const struct kai_rhs_pack_qs4cxs1s0_param* params)>,
std::function<void(size_t num_groups, size_t n, size_t k, size_t nr, size_t kr, size_t sr, size_t rhs_stride, const void* rhs,
const void* bias, const void* scale, void* rhs_packed, size_t extra_bytes, const void* params)>
> pack_func;
};
struct ggml_kleidiai_kernels {
@@ -55,6 +85,10 @@ struct ggml_kleidiai_kernels {
rhs_packing_info rhs_info;
cpu_feature required_cpu;
ggml_type lhs_type;
ggml_type rhs_type;
ggml_type op_type;
};
ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature cpu_features);
ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature cpu_features, const ggml_tensor * tensor);
ggml_kleidiai_kernels * ggml_kleidiai_select_kernels_q4_0(cpu_feature features);