ggml : build backends as libraries (llama/10256)

* ggml : build backends as libraries

---------

Signed-off-by: Xiaodong Ye <[email protected]>
Co-authored-by: Georgi Gerganov <[email protected]>
Co-authored-by: R0CKSTAR <[email protected]>

ggml/CMakeLists.txt

@@ -116,6 +116,7 @@ endif()
 
 # ggml core
 set(GGML_SCHED_MAX_COPIES  "4" CACHE STRING "ggml: max input copies for pipeline parallelism")
+option(GGML_CPU                             "ggml: enable CPU backend"                        ON)
 
 # 3rd party libs / backends
 option(GGML_ACCELERATE                      "ggml: enable Accelerate framework"               ON)
@@ -141,7 +142,7 @@ option(GGML_CUDA_NO_VMM                     "ggml: do not try to use CUDA VMM"
 option(GGML_CUDA_FA_ALL_QUANTS              "ggml: compile all quants for FlashAttention"     OFF)
 option(GGML_CUDA_GRAPHS                     "ggml: use CUDA graphs (llama.cpp only)"          ${GGML_CUDA_GRAPHS_DEFAULT})
 
-option(GGML_HIPBLAS                         "ggml: use hipBLAS"                               OFF)
+option(GGML_HIP                             "ggml: use HIP"                                   OFF)
 option(GGML_HIP_UMA                         "ggml: use HIP unified memory architecture"       OFF)
 option(GGML_VULKAN                          "ggml: use Vulkan"                                OFF)
 option(GGML_VULKAN_CHECK_RESULTS            "ggml: run Vulkan op checks"                      OFF)
@@ -238,12 +239,15 @@ set_target_properties(ggml PROPERTIES PUBLIC_HEADER "${GGML_PUBLIC_HEADERS}")
 install(TARGETS ggml PUBLIC_HEADER)
 
 if (BUILD_SHARED_LIBS)
-    install(TARGETS ggml LIBRARY)
+    install(TARGETS ggml      LIBRARY)
+    install(TARGETS ggml-base LIBRARY)
 endif()
 
+# FIXME: this should be done in the backend cmake files
 if (GGML_METAL)
+    # FIXME: does this need to be installed with GGML_METAL_EMBED_LIBRARY?
     install(
-        FILES src/ggml-metal.metal
+        FILES ggml/src/ggml-metal/ggml-metal.metal
         PERMISSIONS
             OWNER_READ
             OWNER_WRITE

ggml/include/ggml-amx.h

@@ -9,16 +9,16 @@ extern "C" {
 #endif
 
 // buffer_type API
-GGML_API ggml_backend_buffer_type_t ggml_backend_amx_buffer_type(void);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_amx_buffer_type(void);
 
-GGML_API bool ggml_backend_is_amx(ggml_backend_t backend);
+GGML_BACKEND_API bool ggml_backend_is_amx(ggml_backend_t backend);
 
 // backend API
-GGML_API ggml_backend_t ggml_backend_amx_init(void);
+GGML_BACKEND_API ggml_backend_t ggml_backend_amx_init(void);
 
-GGML_API void ggml_backend_amx_set_n_threads(ggml_backend_t backend_amx, int n_threads);
+GGML_BACKEND_API void ggml_backend_amx_set_n_threads(ggml_backend_t backend_amx, int n_threads);
 
-GGML_API ggml_backend_reg_t ggml_backend_amx_reg(void);
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_amx_reg(void);
 
 #ifdef  __cplusplus
 }

ggml/include/ggml-backend.h

@@ -3,6 +3,20 @@
 #include "ggml.h"
 #include "ggml-alloc.h"
 
+#ifdef GGML_BACKEND_SHARED
+#    if defined(_WIN32) && !defined(__MINGW32__)
+#        ifdef GGML_BACKEND_BUILD
+#            define GGML_BACKEND_API __declspec(dllexport) extern
+#        else
+#            define GGML_BACKEND_API __declspec(dllimport) extern
+#        endif
+#    else
+#        define GGML_BACKEND_API __attribute__ ((visibility ("default"))) extern
+#    endif
+#else
+#    define GGML_BACKEND_API extern
+#endif
+
 #ifdef  __cplusplus
 extern "C" {
 #endif

ggml/include/ggml-blas.h

@@ -9,15 +9,15 @@ extern "C" {
 #endif
 
 // backend API
-GGML_API ggml_backend_t ggml_backend_blas_init(void);
+GGML_BACKEND_API ggml_backend_t ggml_backend_blas_init(void);
 
-GGML_API bool ggml_backend_is_blas(ggml_backend_t backend);
+GGML_BACKEND_API bool ggml_backend_is_blas(ggml_backend_t backend);
 
 // number of threads used for conversion to float
 // for openblas and blis, this will also set the number of threads used for blas operations
-GGML_API void ggml_backend_blas_set_n_threads(ggml_backend_t backend_blas, int n_threads);
+GGML_BACKEND_API void ggml_backend_blas_set_n_threads(ggml_backend_t backend_blas, int n_threads);
 
-GGML_API ggml_backend_reg_t ggml_backend_blas_reg(void);
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_blas_reg(void);
 
 
 #ifdef  __cplusplus

ggml/include/ggml-cann.h

@@ -34,7 +34,7 @@ extern "C" {
  */
 #define GGML_CANN_MAX_DEVICES 16
 
-GGML_API ggml_backend_reg_t ggml_backend_cann_reg(void);
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_cann_reg(void);
 
 /**
  * @brief Initializes the CANN backend for a specified device.
@@ -46,7 +46,7 @@ GGML_API ggml_backend_reg_t ggml_backend_cann_reg(void);
  * @param device The index of the device to initialize.
  * @return A pointer to the initialized backend instance, or nullptr on failure.
  */
-GGML_API ggml_backend_t ggml_backend_cann_init(int32_t device);
+GGML_BACKEND_API ggml_backend_t ggml_backend_cann_init(int32_t device);
 
 /**
  * @brief Checks if a given backend is a CANN backend.
@@ -57,7 +57,7 @@ GGML_API ggml_backend_t ggml_backend_cann_init(int32_t device);
  * @param backend The backend instance to check.
  * @return True if the backend is a CANN backend, false otherwise.
  */
-GGML_API bool ggml_backend_is_cann(ggml_backend_t backend);
+GGML_BACKEND_API bool ggml_backend_is_cann(ggml_backend_t backend);
 
 /**
  * @brief Retrieves the CANN buffer type for a specified device.
@@ -69,7 +69,7 @@ GGML_API bool ggml_backend_is_cann(ggml_backend_t backend);
  * @return A pointer to the buffer type interface for the specified device, or
  * nullptr if the device index is out of range.
  */
-GGML_API ggml_backend_buffer_type_t
+GGML_BACKEND_API ggml_backend_buffer_type_t
 ggml_backend_cann_buffer_type(int32_t device);
 
 /**
@@ -80,14 +80,14 @@ ggml_backend_cann_buffer_type(int32_t device);
  *
  * @return The number of CANN devices available.
  */
-GGML_API int32_t ggml_backend_cann_get_device_count(void);
+GGML_BACKEND_API int32_t ggml_backend_cann_get_device_count(void);
 
 /**
  * @brief pinned host buffer for use with the CPU backend for faster copies between CPU and NPU.
  *
  * @return A pointer to the host buffer type interface.
  */
-GGML_API ggml_backend_buffer_type_t ggml_backend_cann_host_buffer_type(void);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cann_host_buffer_type(void);
 
 /**
  * @brief Retrieves the description of a specific CANN device.
@@ -99,7 +99,7 @@ GGML_API ggml_backend_buffer_type_t ggml_backend_cann_host_buffer_type(void);
  * @param description Pointer to a buffer where the description will be written.
  * @param description_size Size of the description buffer.
  */
-GGML_API void ggml_backend_cann_get_device_description(
+GGML_BACKEND_API void ggml_backend_cann_get_device_description(
     int32_t device, char* description, size_t description_size);
 
 /**
@@ -114,7 +114,7 @@ GGML_API void ggml_backend_cann_get_device_description(
  * @param total Pointer to a variable where the total memory size will be
  * stored.
  */
-GGML_API void ggml_backend_cann_get_device_memory(int32_t device,
+GGML_BACKEND_API void ggml_backend_cann_get_device_memory(int32_t device,
                                                   size_t* free,
                                                   size_t* total);
 

ggml/include/ggml-cpu.h

@@ -54,54 +54,77 @@ extern "C" {
         GGML_NUMA_STRATEGY_COUNT
     };
 
-    GGML_API void    ggml_numa_init(enum ggml_numa_strategy numa); // call once for better performance on NUMA systems
-    GGML_API bool    ggml_is_numa(void); // true if init detected that system has >1 NUMA node
+    GGML_BACKEND_API void    ggml_numa_init(enum ggml_numa_strategy numa); // call once for better performance on NUMA systems
+    GGML_BACKEND_API bool    ggml_is_numa(void); // true if init detected that system has >1 NUMA node
 
-    GGML_API struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value);
-    GGML_API struct ggml_tensor * ggml_new_f32(struct ggml_context * ctx, float value);
+    GGML_BACKEND_API struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value);
+    GGML_BACKEND_API struct ggml_tensor * ggml_new_f32(struct ggml_context * ctx, float value);
 
-    GGML_API struct ggml_tensor * ggml_set_i32 (struct ggml_tensor * tensor, int32_t value);
-    GGML_API struct ggml_tensor * ggml_set_f32 (struct ggml_tensor * tensor, float value);
+    GGML_BACKEND_API struct ggml_tensor * ggml_set_i32 (struct ggml_tensor * tensor, int32_t value);
+    GGML_BACKEND_API struct ggml_tensor * ggml_set_f32 (struct ggml_tensor * tensor, float value);
 
-    GGML_API int32_t ggml_get_i32_1d(const struct ggml_tensor * tensor, int i);
-    GGML_API void    ggml_set_i32_1d(const struct ggml_tensor * tensor, int i, int32_t value);
+    GGML_BACKEND_API int32_t ggml_get_i32_1d(const struct ggml_tensor * tensor, int i);
+    GGML_BACKEND_API void    ggml_set_i32_1d(const struct ggml_tensor * tensor, int i, int32_t value);
 
-    GGML_API int32_t ggml_get_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3);
-    GGML_API void    ggml_set_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, int32_t value);
+    GGML_BACKEND_API int32_t ggml_get_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3);
+    GGML_BACKEND_API void    ggml_set_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, int32_t value);
 
-    GGML_API float   ggml_get_f32_1d(const struct ggml_tensor * tensor, int i);
-    GGML_API void    ggml_set_f32_1d(const struct ggml_tensor * tensor, int i, float value);
+    GGML_BACKEND_API float   ggml_get_f32_1d(const struct ggml_tensor * tensor, int i);
+    GGML_BACKEND_API void    ggml_set_f32_1d(const struct ggml_tensor * tensor, int i, float value);
 
-    GGML_API float   ggml_get_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3);
-    GGML_API void    ggml_set_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, float value);
+    GGML_BACKEND_API float   ggml_get_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3);
+    GGML_BACKEND_API void    ggml_set_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, float value);
 
-    GGML_API struct ggml_threadpool_params ggml_threadpool_params_default(int n_threads);
-    GGML_API void                          ggml_threadpool_params_init   (struct ggml_threadpool_params * p, int n_threads);
-    GGML_API bool                          ggml_threadpool_params_match  (const struct ggml_threadpool_params * p0, const struct ggml_threadpool_params * p1);
-    GGML_API struct ggml_threadpool *      ggml_threadpool_new          (struct ggml_threadpool_params  * params);
-    GGML_API void                          ggml_threadpool_free         (struct ggml_threadpool * threadpool);
-    GGML_API int                           ggml_threadpool_get_n_threads(struct ggml_threadpool * threadpool);
-    GGML_API void                          ggml_threadpool_pause        (struct ggml_threadpool * threadpool);
-    GGML_API void                          ggml_threadpool_resume       (struct ggml_threadpool * threadpool);
+    GGML_BACKEND_API struct ggml_threadpool_params ggml_threadpool_params_default(int n_threads);
+    GGML_BACKEND_API void                          ggml_threadpool_params_init   (struct ggml_threadpool_params * p, int n_threads);
+    GGML_BACKEND_API bool                          ggml_threadpool_params_match  (const struct ggml_threadpool_params * p0, const struct ggml_threadpool_params * p1);
+    GGML_BACKEND_API struct ggml_threadpool *      ggml_threadpool_new          (struct ggml_threadpool_params  * params);
+    GGML_BACKEND_API void                          ggml_threadpool_free         (struct ggml_threadpool * threadpool);
+    GGML_BACKEND_API int                           ggml_threadpool_get_n_threads(struct ggml_threadpool * threadpool);
+    GGML_BACKEND_API void                          ggml_threadpool_pause        (struct ggml_threadpool * threadpool);
+    GGML_BACKEND_API void                          ggml_threadpool_resume       (struct ggml_threadpool * threadpool);
 
     // ggml_graph_plan() has to be called before ggml_graph_compute()
     // when plan.work_size > 0, caller must allocate memory for plan.work_data
-    GGML_API struct ggml_cplan ggml_graph_plan(
+    GGML_BACKEND_API struct ggml_cplan ggml_graph_plan(
                   const struct ggml_cgraph * cgraph,
                                        int   n_threads, /* = GGML_DEFAULT_N_THREADS */
                     struct ggml_threadpool * threadpool /* = NULL */ );
-    GGML_API enum ggml_status  ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan);
+    GGML_BACKEND_API enum ggml_status  ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan);
 
     // same as ggml_graph_compute() but the work data is allocated as a part of the context
     // note: the drawback of this API is that you must have ensured that the context has enough memory for the work data
-    GGML_API enum ggml_status  ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct ggml_cgraph * cgraph, int n_threads);
+    GGML_BACKEND_API enum ggml_status  ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct ggml_cgraph * cgraph, int n_threads);
 
-    // TODO: move to backend interface
-    GGML_API int ggml_cpu_has_neon       (void);
-    GGML_API int ggml_cpu_has_sve        (void);
-    GGML_API int ggml_cpu_has_matmul_int8(void);
-    // get the sve vector length in bytes
-    GGML_API int ggml_cpu_get_sve_cnt(void);
+    //
+    // system info
+    //
+
+    // x86
+    GGML_BACKEND_API int ggml_cpu_has_sse3       (void);
+    GGML_BACKEND_API int ggml_cpu_has_ssse3      (void);
+    GGML_BACKEND_API int ggml_cpu_has_avx        (void);
+    GGML_BACKEND_API int ggml_cpu_has_avx2       (void);
+    GGML_BACKEND_API int ggml_cpu_has_f16c       (void);
+    GGML_BACKEND_API int ggml_cpu_has_fma        (void);
+    GGML_BACKEND_API int ggml_cpu_has_avx_vnni   (void);
+    GGML_BACKEND_API int ggml_cpu_has_avx512     (void);
+    GGML_BACKEND_API int ggml_cpu_has_avx512_vbmi(void);
+    GGML_BACKEND_API int ggml_cpu_has_avx512_vnni(void);
+    GGML_BACKEND_API int ggml_cpu_has_avx512_bf16(void);
+    GGML_BACKEND_API int ggml_cpu_has_amx_int8   (void);
+    // ARM
+    GGML_BACKEND_API int ggml_cpu_has_neon       (void);
+    GGML_BACKEND_API int ggml_cpu_has_arm_fma    (void);
+    GGML_BACKEND_API int ggml_cpu_has_fp16_va    (void);
+    GGML_BACKEND_API int ggml_cpu_has_matmul_int8(void);
+    GGML_BACKEND_API int ggml_cpu_has_sve        (void);
+    GGML_BACKEND_API int ggml_cpu_get_sve_cnt    (void);  // sve vector length in bytes
+    // other
+    GGML_BACKEND_API int ggml_cpu_has_riscv_v    (void);
+    GGML_BACKEND_API int ggml_cpu_has_vsx        (void);
+    GGML_BACKEND_API int ggml_cpu_has_wasm_simd  (void);
+    GGML_BACKEND_API int ggml_cpu_has_llamafile  (void);
 
     // Internal types and functions exposed for tests and benchmarks
 
@@ -115,6 +138,7 @@ extern "C" {
                                        const void * GGML_RESTRICT y, int nr, int nc);
 
     struct ggml_type_traits_cpu {
+        ggml_from_float_t        from_float;
         ggml_from_float_to_mat_t from_float_to_mat;
         ggml_vec_dot_t           vec_dot;
         enum ggml_type           vec_dot_type;
@@ -124,25 +148,25 @@ extern "C" {
         ggml_gemm_t              gemm;
     };
 
-    GGML_API const struct ggml_type_traits_cpu * ggml_get_type_traits_cpu(enum ggml_type type);
+    GGML_BACKEND_API const struct ggml_type_traits_cpu * ggml_get_type_traits_cpu(enum ggml_type type);
 
-    GGML_API void ggml_cpu_init(void);
+    GGML_BACKEND_API void ggml_cpu_init(void);
 
     //
     // CPU backend
     //
 
-    GGML_API ggml_backend_t ggml_backend_cpu_init(void);
+    GGML_BACKEND_API ggml_backend_t ggml_backend_cpu_init(void);
 
-    GGML_API bool ggml_backend_is_cpu                (ggml_backend_t backend);
-    GGML_API void ggml_backend_cpu_set_n_threads     (ggml_backend_t backend_cpu, int n_threads);
-    GGML_API void ggml_backend_cpu_set_threadpool    (ggml_backend_t backend_cpu, ggml_threadpool_t threadpool);
-    GGML_API void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data);
+    GGML_BACKEND_API bool ggml_backend_is_cpu                (ggml_backend_t backend);
+    GGML_BACKEND_API void ggml_backend_cpu_set_n_threads     (ggml_backend_t backend_cpu, int n_threads);
+    GGML_BACKEND_API void ggml_backend_cpu_set_threadpool    (ggml_backend_t backend_cpu, ggml_threadpool_t threadpool);
+    GGML_BACKEND_API void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data);
 
-    GGML_API ggml_backend_reg_t ggml_backend_cpu_reg(void);
+    GGML_BACKEND_API ggml_backend_reg_t ggml_backend_cpu_reg(void);
 
 #ifdef GGML_USE_CPU_HBM
-    GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void);
+    GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void);
 #endif
 
 #ifdef __cplusplus

ggml/include/ggml-cuda.h

@@ -7,7 +7,7 @@
 extern "C" {
 #endif
 
-#ifdef GGML_USE_HIPBLAS
+#ifdef GGML_USE_HIP
 #define GGML_CUDA_NAME "ROCm"
 #define GGML_CUBLAS_NAME "hipBLAS"
 #elif defined(GGML_USE_MUSA)
@@ -20,27 +20,27 @@ extern "C" {
 #define GGML_CUDA_MAX_DEVICES       16
 
 // backend API
-GGML_API ggml_backend_t ggml_backend_cuda_init(int device);
+GGML_BACKEND_API ggml_backend_t ggml_backend_cuda_init(int device);
 
-GGML_API bool ggml_backend_is_cuda(ggml_backend_t backend);
+GGML_BACKEND_API bool ggml_backend_is_cuda(ggml_backend_t backend);
 
 // device buffer
-GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device);
 
 // split tensor buffer that splits matrices by rows across multiple devices
-GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(int main_device, const float * tensor_split);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(int main_device, const float * tensor_split);
 
 // pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
-GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);
 
-GGML_API int  ggml_backend_cuda_get_device_count(void);
-GGML_API void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size);
-GGML_API void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total);
+GGML_BACKEND_API int  ggml_backend_cuda_get_device_count(void);
+GGML_BACKEND_API void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size);
+GGML_BACKEND_API void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total);
 
-GGML_API bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size);
-GGML_API void ggml_backend_cuda_unregister_host_buffer(void * buffer);
+GGML_BACKEND_API bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size);
+GGML_BACKEND_API void ggml_backend_cuda_unregister_host_buffer(void * buffer);
 
-GGML_API ggml_backend_reg_t ggml_backend_cuda_reg(void);
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_cuda_reg(void);
 
 #ifdef  __cplusplus
 }

ggml/include/ggml-kompute.h

@@ -37,13 +37,13 @@ struct ggml_vk_device ggml_vk_current_device(void);
 // forward declaration
 typedef struct ggml_backend * ggml_backend_t;
 
-GGML_API ggml_backend_t ggml_backend_kompute_init(int device);
+GGML_BACKEND_API ggml_backend_t ggml_backend_kompute_init(int device);
 
-GGML_API bool ggml_backend_is_kompute(ggml_backend_t backend);
+GGML_BACKEND_API bool ggml_backend_is_kompute(ggml_backend_t backend);
 
-GGML_API ggml_backend_buffer_type_t ggml_backend_kompute_buffer_type(int device);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_kompute_buffer_type(int device);
 
-GGML_API ggml_backend_reg_t ggml_backend_kompute_reg(void);
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_kompute_reg(void);
 
 #ifdef __cplusplus
 }

ggml/include/ggml-metal.h

@@ -39,27 +39,27 @@ extern "C" {
 // user-code should use only these functions
 //
 
-GGML_API ggml_backend_t ggml_backend_metal_init(void);
+GGML_BACKEND_API ggml_backend_t ggml_backend_metal_init(void);
 
-GGML_API bool ggml_backend_is_metal(ggml_backend_t backend);
+GGML_BACKEND_API bool ggml_backend_is_metal(ggml_backend_t backend);
 
 GGML_DEPRECATED(
-        GGML_API ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size),
+        GGML_BACKEND_API ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size),
         "obsoleted by the new device interface - https://github.com/ggerganov/llama.cpp/pull/9713");
 
-GGML_API void ggml_backend_metal_set_abort_callback(ggml_backend_t backend, ggml_abort_callback abort_callback, void * user_data);
+GGML_BACKEND_API void ggml_backend_metal_set_abort_callback(ggml_backend_t backend, ggml_abort_callback abort_callback, void * user_data);
 
-GGML_API ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
 
 // helper to check if the device supports a specific family
 // ideally, the user code should be doing these checks
 // ref: https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf
-GGML_API bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family);
+GGML_BACKEND_API bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family);
 
 // capture all command buffers committed the next time `ggml_backend_graph_compute` is called
-GGML_API void ggml_backend_metal_capture_next_compute(ggml_backend_t backend);
+GGML_BACKEND_API void ggml_backend_metal_capture_next_compute(ggml_backend_t backend);
 
-GGML_API ggml_backend_reg_t ggml_backend_metal_reg(void);
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_metal_reg(void);
 
 #ifdef __cplusplus
 }

ggml/include/ggml-rpc.h

@@ -10,18 +10,18 @@ extern "C" {
 #define GGML_RPC_MAX_SERVERS       16
 
 // backend API
-GGML_API ggml_backend_t ggml_backend_rpc_init(const char * endpoint);
-GGML_API bool ggml_backend_is_rpc(ggml_backend_t backend);
+GGML_BACKEND_API ggml_backend_t ggml_backend_rpc_init(const char * endpoint);
+GGML_BACKEND_API bool ggml_backend_is_rpc(ggml_backend_t backend);
 
-GGML_API ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * endpoint);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * endpoint);
 
-GGML_API void ggml_backend_rpc_get_device_memory(const char * endpoint, size_t * free, size_t * total);
+GGML_BACKEND_API void ggml_backend_rpc_get_device_memory(const char * endpoint, size_t * free, size_t * total);
 
-GGML_API void ggml_backend_rpc_start_server(ggml_backend_t backend, const char * endpoint, size_t free_mem, size_t total_mem);
+GGML_BACKEND_API void ggml_backend_rpc_start_server(ggml_backend_t backend, const char * endpoint, size_t free_mem, size_t total_mem);
 
-GGML_API ggml_backend_reg_t ggml_backend_rpc_reg(void);
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_rpc_reg(void);
 
-GGML_API ggml_backend_dev_t ggml_backend_rpc_add_device(const char * endpoint);
+GGML_BACKEND_API ggml_backend_dev_t ggml_backend_rpc_add_device(const char * endpoint);
 
 #ifdef  __cplusplus
 }

ggml/include/ggml-sycl.h

@@ -17,32 +17,32 @@ extern "C" {
 #endif
 
 // backend API
-GGML_API ggml_backend_t ggml_backend_sycl_init(int device);
+GGML_BACKEND_API ggml_backend_t ggml_backend_sycl_init(int device);
 
-GGML_API bool ggml_backend_is_sycl(ggml_backend_t backend);
+GGML_BACKEND_API bool ggml_backend_is_sycl(ggml_backend_t backend);
 
 // devide buffer
-GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device);
 
 // split tensor buffer that splits matrices by rows across multiple devices
-GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split);
 
 // pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
-GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type(void);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type(void);
 
-GGML_API void ggml_backend_sycl_print_sycl_devices(void);
-GGML_API void ggml_backend_sycl_get_gpu_list(int *id_list, int max_len);
-GGML_API void ggml_backend_sycl_get_device_description(int device,
+GGML_BACKEND_API void ggml_backend_sycl_print_sycl_devices(void);
+GGML_BACKEND_API void ggml_backend_sycl_get_gpu_list(int *id_list, int max_len);
+GGML_BACKEND_API void ggml_backend_sycl_get_device_description(int device,
                                                        char *description,
                                                        size_t description_size);
-GGML_API int  ggml_backend_sycl_get_device_count();
-GGML_API void ggml_backend_sycl_get_device_memory(int device, size_t *free, size_t *total);
+GGML_BACKEND_API int  ggml_backend_sycl_get_device_count();
+GGML_BACKEND_API void ggml_backend_sycl_get_device_memory(int device, size_t *free, size_t *total);
 
 // SYCL doesn't support registering host memory, keep here for reference
-// GGML_API bool ggml_backend_sycl_register_host_buffer(void * buffer, size_t size);
-// GGML_API void ggml_backend_sycl_unregister_host_buffer(void * buffer);
+// GGML_BACKEND_API bool ggml_backend_sycl_register_host_buffer(void * buffer, size_t size);
+// GGML_BACKEND_API void ggml_backend_sycl_unregister_host_buffer(void * buffer);
 
-GGML_API ggml_backend_reg_t ggml_backend_sycl_reg(void);
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_sycl_reg(void);
 
 #ifdef  __cplusplus
 }

ggml/include/ggml-vulkan.h

@@ -10,21 +10,21 @@ extern "C" {
 #define GGML_VK_NAME "Vulkan"
 #define GGML_VK_MAX_DEVICES 16
 
-GGML_API void ggml_vk_instance_init(void);
+GGML_BACKEND_API void ggml_vk_instance_init(void);
 
 // backend API
-GGML_API ggml_backend_t ggml_backend_vk_init(size_t dev_num);
+GGML_BACKEND_API ggml_backend_t ggml_backend_vk_init(size_t dev_num);
 
-GGML_API bool ggml_backend_is_vk(ggml_backend_t backend);
-GGML_API int  ggml_backend_vk_get_device_count(void);
-GGML_API void ggml_backend_vk_get_device_description(int device, char * description, size_t description_size);
-GGML_API void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total);
+GGML_BACKEND_API bool ggml_backend_is_vk(ggml_backend_t backend);
+GGML_BACKEND_API int  ggml_backend_vk_get_device_count(void);
+GGML_BACKEND_API void ggml_backend_vk_get_device_description(int device, char * description, size_t description_size);
+GGML_BACKEND_API void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total);
 
-GGML_API ggml_backend_buffer_type_t ggml_backend_vk_buffer_type(size_t dev_num);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_vk_buffer_type(size_t dev_num);
 // pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
-GGML_API ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type(void);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type(void);
 
-GGML_API ggml_backend_reg_t ggml_backend_vk_reg(void);
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_vk_reg(void);
 
 #ifdef  __cplusplus
 }

ggml/include/ggml.h

@@ -176,15 +176,15 @@
 #ifdef GGML_SHARED
 #    if defined(_WIN32) && !defined(__MINGW32__)
 #        ifdef GGML_BUILD
-#            define GGML_API __declspec(dllexport)
+#            define GGML_API __declspec(dllexport) extern
 #        else
-#            define GGML_API __declspec(dllimport)
+#            define GGML_API __declspec(dllimport) extern
 #        endif
 #    else
-#        define GGML_API __attribute__ ((visibility ("default")))
+#        define GGML_API __attribute__ ((visibility ("default"))) extern
 #    endif
 #else
-#    define GGML_API
+#    define GGML_API extern
 #endif
 
 // TODO: support for clang
@@ -1490,7 +1490,7 @@ extern "C" {
         "use ggml_rope_ext_inplace instead");
 
     // compute correction dims for YaRN RoPE scaling
-    void ggml_rope_yarn_corr_dims(
+    GGML_API void ggml_rope_yarn_corr_dims(
         int n_dims, int n_ctx_orig, float freq_base, float beta_fast, float beta_slow, float dims[2]);
 
     // rotary position embedding backward, i.e compute dx from dy
@@ -2384,38 +2384,6 @@ extern "C" {
     GGML_API size_t gguf_get_meta_size(const struct gguf_context * ctx);
     GGML_API void   gguf_get_meta_data(const struct gguf_context * ctx, void * data);
 
-    //
-    // system info
-    //
-
-    GGML_API int ggml_cpu_has_avx        (void);
-    GGML_API int ggml_cpu_has_avx_vnni   (void);
-    GGML_API int ggml_cpu_has_avx2       (void);
-    GGML_API int ggml_cpu_has_avx512     (void);
-    GGML_API int ggml_cpu_has_avx512_vbmi(void);
-    GGML_API int ggml_cpu_has_avx512_vnni(void);
-    GGML_API int ggml_cpu_has_avx512_bf16(void);
-    GGML_API int ggml_cpu_has_amx_int8   (void);
-    GGML_API int ggml_cpu_has_fma        (void);
-    GGML_API int ggml_cpu_has_arm_fma    (void);
-    GGML_API int ggml_cpu_has_metal      (void);
-    GGML_API int ggml_cpu_has_f16c       (void);
-    GGML_API int ggml_cpu_has_fp16_va    (void);
-    GGML_API int ggml_cpu_has_wasm_simd  (void);
-    GGML_API int ggml_cpu_has_blas       (void);
-    GGML_API int ggml_cpu_has_cuda       (void);
-    GGML_API int ggml_cpu_has_vulkan     (void);
-    GGML_API int ggml_cpu_has_kompute    (void);
-    GGML_API int ggml_cpu_has_gpublas    (void);
-    GGML_API int ggml_cpu_has_sse3       (void);
-    GGML_API int ggml_cpu_has_ssse3      (void);
-    GGML_API int ggml_cpu_has_riscv_v    (void);
-    GGML_API int ggml_cpu_has_sycl       (void);
-    GGML_API int ggml_cpu_has_rpc        (void);
-    GGML_API int ggml_cpu_has_vsx        (void);
-    GGML_API int ggml_cpu_has_cann       (void);
-    GGML_API int ggml_cpu_has_llamafile  (void);
-
 #ifdef  __cplusplus
 // restrict not standard in C++
 #define GGML_RESTRICT
@@ -2432,7 +2400,6 @@ extern "C" {
         size_t                   type_size;
         bool                     is_quantized;
         ggml_to_float_t          to_float;
-        ggml_from_float_t        from_float;
         ggml_from_float_t        from_float_ref;
     };
 

ggml/src/ggml-aarch64.h

@@ -1,9 +1,5 @@
-// SPDX-FileCopyrightText: Copyright 2024 Arm Ltd.
 #pragma once
 
-#define GGML_COMMON_DECL_C
-#include "ggml-common.h"
-
 #include "ggml.h"
 
 // GGML internal header
@@ -12,27 +8,11 @@
 extern "C" {
 #endif
 
-// Quantization
-void quantize_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-
-void quantize_mat_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t nrows, int64_t n_per_row, int64_t blck_size_interleave);
-
 // Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization")
 size_t quantize_q4_0_4x4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
 size_t quantize_q4_0_4x8(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
 size_t quantize_q4_0_8x8(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
 
-// GEMV
-void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-
-// GEMM
-void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-
 #ifdef __cplusplus
 }
 #endif

ggml/src/ggml-amx/CMakeLists.txt

@@ -0,0 +1,107 @@
+if (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LWR MATCHES "^(x86_64|i686|amd64|x64|win32)$" OR
+        (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND
+         CMAKE_SYSTEM_PROCESSOR MATCHES "^(x86_64|i686|AMD64)$") AND
+        CMAKE_COMPILER_IS_GNUCC AND CMAKE_CXX_COMPILER_VERSION VERSION_GREATER 11.0)
+    message(STATUS "Using AMX")
+
+    file(GLOB   GGML_HEADERS_AMX "*.h")
+    list(APPEND GGML_HEADERS_AMX "../../include/ggml-amx.h")
+
+    file(GLOB   GGML_SOURCES_AMX "*.cpp")
+
+    add_library(ggml-amx
+                ${GGML_HEADERS_AMX}
+                ${GGML_SOURCES_AMX})
+
+    target_link_libraries(ggml-amx PRIVATE ggml-base)
+    target_include_directories(ggml-amx PRIVATE . ..)
+
+    # this is duplicated from the CPU backend, since the AMX backend also depends on the architecture flags
+    # TODO: integrate AMX backend into the CPU backend
+    if (MSVC)
+        # instruction set detection for MSVC only
+        if (GGML_NATIVE)
+            # TODO: improve, should not reference files from the parent folder
+            include(../ggml-cpu/cmake/FindSIMD.cmake)
+        endif ()
+        if (GGML_AVX512)
+            list(APPEND ARCH_FLAGS /arch:AVX512)
+            # MSVC has no compile-time flags enabling specific
+            # AVX512 extensions, neither it defines the
+            # macros corresponding to the extensions.
+            # Do it manually.
+            if (GGML_AVX512_VBMI)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VBMI__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VBMI__>)
+            endif()
+            if (GGML_AVX512_VNNI)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VNNI__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VNNI__>)
+            endif()
+            if (GGML_AVX512_BF16)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512BF16__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512BF16__>)
+            endif()
+            if (GGML_AMX_TILE)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AMX_TILE__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AMX_TILE__>)
+            endif()
+            if (GGML_AMX_INT8)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AMX_INT8__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AMX_INT8__>)
+            endif()
+            if (GGML_AMX_BF16)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AMX_BF16__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AMX_BF16__>)
+            endif()
+        elseif (GGML_AVX2)
+            list(APPEND ARCH_FLAGS /arch:AVX2)
+        elseif (GGML_AVX)
+            list(APPEND ARCH_FLAGS /arch:AVX)
+        endif()
+    else()
+        if (GGML_NATIVE)
+            list(APPEND ARCH_FLAGS -march=native)
+        endif()
+        if (GGML_F16C)
+            list(APPEND ARCH_FLAGS -mf16c)
+        endif()
+        if (GGML_FMA)
+            list(APPEND ARCH_FLAGS -mfma)
+        endif()
+        if (GGML_AVX)
+            list(APPEND ARCH_FLAGS -mavx)
+        endif()
+        if (GGML_AVX2)
+            list(APPEND ARCH_FLAGS -mavx2)
+        endif()
+        if (GGML_AVX512)
+            list(APPEND ARCH_FLAGS -mavx512f)
+            list(APPEND ARCH_FLAGS -mavx512dq)
+            list(APPEND ARCH_FLAGS -mavx512bw)
+        endif()
+        if (GGML_AVX512_VBMI)
+            list(APPEND ARCH_FLAGS -mavx512vbmi)
+        endif()
+        if (GGML_AVX512_VNNI)
+            list(APPEND ARCH_FLAGS -mavx512vnni)
+        endif()
+        if (GGML_AVX512_BF16)
+            list(APPEND ARCH_FLAGS -mavx512bf16)
+        endif()
+        if (GGML_AMX_TILE)
+            list(APPEND ARCH_FLAGS -mamx-tile)
+        endif()
+        if (GGML_AMX_INT8)
+            list(APPEND ARCH_FLAGS -mamx-int8)
+        endif()
+        if (GGML_AMX_BF16)
+            list(APPEND ARCH_FLAGS -mamx-bf16)
+        endif()
+    endif()
+
+    target_compile_options(ggml-amx PRIVATE ${ARCH_FLAGS})
+else()
+    set(GGML_AMX OFF PARENT_SCOPE)
+    message(WARNING "AMX requires x86 and gcc version > 11.0. Turning off GGML_AMX.")
+endif()

ggml/src/ggml-amx/common.h

@@ -1,7 +1,8 @@
 #pragma once
 
 #include "ggml.h"
-#include "ggml-cpu-impl.h" // <immintrin.h>
+// hack until AMX is moved into the CPU backend
+#include "../ggml-cpu/ggml-cpu-impl.h" // <immintrin.h>
 
 #include <algorithm>
 #include <memory>

ggml/src/ggml-amx/ggml-amx.cpp

@@ -0,0 +1,449 @@
+#include "ggml-amx.h"
+#include "ggml-amx/common.h"
+#include "ggml-amx/mmq.h"
+#include "ggml-backend-impl.h"
+#include "ggml-impl.h"
+
+#if defined(__gnu_linux__)
+#include <sys/syscall.h>
+#include <unistd.h>
+#endif
+
+#include <cstdlib>
+#include <cstring>
+#include <memory>
+
+#if defined(__AMX_INT8__)
+
+// AMX buffer interface
+static void ggml_backend_amx_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    free(buffer->context);
+}
+
+static void * ggml_backend_amx_buffer_get_base(ggml_backend_buffer_t buffer) {
+    return (void *)(buffer->context);
+}
+
+static void ggml_backend_amx_buffer_memset_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
+    memset((char *)tensor->data + offset, value, size);
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_amx_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    if (qtype_has_amx_kernels(tensor->type)) {
+        ggml_backend_amx_convert_weight(tensor, data, offset, size);
+    } else {
+        memcpy((char *)tensor->data + offset, data, size);
+    }
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_amx_buffer_get_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    GGML_ASSERT(!qtype_has_amx_kernels(tensor->type));
+    memcpy(data, (const char *)tensor->data + offset, size);
+
+    GGML_UNUSED(buffer);
+}
+
+static bool ggml_backend_amx_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst) {
+    if (ggml_backend_buffer_is_host(src->buffer)) {
+        if (qtype_has_amx_kernels(src->type)) {
+            ggml_backend_amx_convert_weight(dst, src->data, 0, ggml_backend_amx_get_alloc_size(dst));
+        } else {
+            memcpy(dst->data, src->data, ggml_nbytes(src));
+        }
+        return true;
+    }
+    return false;
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_amx_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    memset(buffer->context, value, buffer->size);
+}
+
+static ggml_backend_buffer_i ggml_backend_amx_buffer_interface = {
+    /* .free_buffer     = */ ggml_backend_amx_buffer_free_buffer,
+    /* .get_base        = */ ggml_backend_amx_buffer_get_base,
+    /* .init_tensor     = */ NULL, // no initialization required
+    /* .memset_tensor   = */ ggml_backend_amx_buffer_memset_tensor,
+    /* .set_tensor      = */ ggml_backend_amx_buffer_set_tensor,
+    /* .get_tensor      = */ ggml_backend_amx_buffer_get_tensor,
+    /* .cpy_tensor      = */ ggml_backend_amx_buffer_cpy_tensor,
+    /* .clear           = */ ggml_backend_amx_buffer_clear,
+    /* .reset           = */ NULL,
+};
+
+static const char * ggml_backend_amx_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
+    return "AMX";
+
+    GGML_UNUSED(buft);
+}
+
+static ggml_backend_buffer_t ggml_backend_amx_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    void * data = aligned_alloc(TENSOR_ALIGNMENT, size);
+    if (data == NULL) {
+        fprintf(stderr, "%s: failed to allocate buffer of size %zu\n", __func__, size);
+        return NULL;
+    }
+
+    return ggml_backend_buffer_init(buft, ggml_backend_amx_buffer_interface, data, size);
+}
+
+static size_t ggml_backend_amx_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
+    return TENSOR_ALIGNMENT;
+
+    GGML_UNUSED(buft);
+}
+
+static size_t ggml_backend_amx_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor* tensor) {
+    return ggml_backend_amx_get_alloc_size(tensor);
+
+    GGML_UNUSED(buft);
+}
+
+static bool ggml_backend_amx_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
+    return false;
+
+    GGML_UNUSED(buft);
+}
+
+ggml_backend_buffer_type_t ggml_backend_amx_buffer_type() {
+    static struct ggml_backend_buffer_type ggml_backend_buffer_type_amx = {
+        /* .iface = */ {
+            /* .get_name         = */ ggml_backend_amx_buffer_type_get_name,
+            /* .alloc_buffer     = */ ggml_backend_amx_buffer_type_alloc_buffer,
+            /* .get_alignment    = */ ggml_backend_amx_buffer_type_get_alignment,
+            /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
+            /* .get_alloc_size   = */ ggml_backend_amx_buffer_type_get_alloc_size,
+            /* .is_host          = */ ggml_backend_amx_buffer_type_is_host,
+        },
+        /* .device  = */ ggml_backend_reg_dev_get(ggml_backend_amx_reg(), 0),
+        /* .context = */ NULL,
+    };
+
+    return &ggml_backend_buffer_type_amx;
+}
+
+// backend interface
+
+static const char * ggml_backend_amx_name(ggml_backend_t backend) {
+    return "AMX";
+
+    GGML_UNUSED(backend);
+}
+
+static void ggml_backend_amx_free(ggml_backend_t backend) {
+    ggml_backend_amx_context * ctx = (ggml_backend_amx_context *)backend->context;
+    delete ctx;
+    delete backend;
+}
+
+static enum ggml_status ggml_backend_amx_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    ggml_backend_amx_context * ctx = (ggml_backend_amx_context *)backend->context;
+
+    for (int i = 0; i < cgraph->n_nodes; i++) {
+        struct ggml_tensor * node = cgraph->nodes[i];
+
+        switch (node->op) {
+        case GGML_OP_MUL_MAT:
+            ggml_backend_amx_mul_mat(ctx, node);
+            break;
+
+        case GGML_OP_NONE:
+        case GGML_OP_RESHAPE:
+        case GGML_OP_VIEW:
+        case GGML_OP_PERMUTE:
+        case GGML_OP_TRANSPOSE:
+            break;
+
+        default:
+            fprintf(stderr, "%s: unsupported op %s\n", __func__, ggml_op_desc(node));
+            GGML_ASSERT(false);
+        }
+    }
+
+    return GGML_STATUS_SUCCESS;
+
+    GGML_UNUSED(backend);
+}
+
+static struct ggml_backend_i ggml_backend_amx_i = {
+    /* .get_name                = */ ggml_backend_amx_name,
+    /* .free                    = */ ggml_backend_amx_free,
+    /* .set_tensor_async        = */ NULL,
+    /* .get_tensor_async        = */ NULL,
+    /* .cpy_tensor_async        = */ NULL,
+    /* .synchronize             = */ NULL,
+    /* .graph_plan_create       = */ NULL,
+    /* .graph_plan_free         = */ NULL,
+    /* .graph_plan_update       = */ NULL,
+    /* .graph_plan_compute      = */ NULL,
+    /* .graph_compute           = */ ggml_backend_amx_graph_compute,
+    /* .event_record            = */ NULL,
+    /* .event_wait              = */ NULL,
+};
+
+static ggml_guid_t ggml_backend_amx_guid() {
+    static ggml_guid guid = { 0x13, 0xb8, 0xa4, 0xc4, 0xba, 0xfe, 0x51, 0x67, 0x87, 0x44, 0x55, 0x15, 0xb2, 0x35, 0x62, 0x3e };
+    return &guid;
+}
+
+#define ARCH_GET_XCOMP_PERM     0x1022
+#define ARCH_REQ_XCOMP_PERM     0x1023
+#define XFEATURE_XTILECFG       17
+#define XFEATURE_XTILEDATA      18
+
+static bool ggml_amx_init() {
+#if defined(__gnu_linux__)
+    if (syscall(SYS_arch_prctl, ARCH_REQ_XCOMP_PERM, XFEATURE_XTILEDATA)) {
+        fprintf(stderr, "AMX is not ready to be used!\n");
+        return false;
+    }
+    return true;
+#elif defined(_WIN32)
+    return true;
+#endif
+}
+
+ggml_backend_t ggml_backend_amx_init() {
+
+    // invoke a Linux system call to request access to AMX features
+    ggml_amx_init();
+
+    // backend context
+    ggml_backend_amx_context * ctx = new ggml_backend_amx_context;
+
+    // ggml amx backend
+    ggml_backend_t backend = new ggml_backend {
+        /* .guid      = */ ggml_backend_amx_guid(),
+        /* .interface = */ ggml_backend_amx_i,
+        /* .device    = */ ggml_backend_reg_dev_get(ggml_backend_amx_reg(), 0),
+        /* .context   = */ ctx,
+    };
+
+    return backend;
+}
+
+bool ggml_backend_is_amx(ggml_backend_t backend) {
+    return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_amx_guid());
+}
+
+void ggml_backend_amx_set_n_threads(ggml_backend_t backend_amx, int n_threads) {
+    GGML_ASSERT(ggml_backend_is_amx(backend_amx));
+
+    ggml_backend_amx_context * ctx = (ggml_backend_amx_context *)backend_amx->context;
+    ctx->n_threads = n_threads;
+}
+
+// device interface
+
+static const char * ggml_backend_amx_device_get_name(ggml_backend_dev_t dev) {
+    return "AMX";
+
+    GGML_UNUSED(dev);
+}
+
+static const char * ggml_backend_amx_device_get_description(ggml_backend_dev_t dev) {
+    return "Intel Advanced Matrix Extensions";
+
+    GGML_UNUSED(dev);
+}
+
+static void ggml_backend_amx_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
+    // TODO
+    *free = 0;
+    *total = 0;
+
+    GGML_UNUSED(dev);
+}
+
+static enum ggml_backend_dev_type ggml_backend_amx_device_get_type(ggml_backend_dev_t dev) {
+    return GGML_BACKEND_DEVICE_TYPE_ACCEL;
+
+    GGML_UNUSED(dev);
+}
+
+static void ggml_backend_amx_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) {
+    props->name        = ggml_backend_amx_device_get_name(dev);
+    props->description = ggml_backend_amx_device_get_description(dev);
+    props->type        = ggml_backend_amx_device_get_type(dev);
+    ggml_backend_amx_device_get_memory(dev, &props->memory_free, &props->memory_total);
+
+    // `buffer_from_host_ptr` is intended to be used in mmap, when memory layout unchanged
+    props->caps = {
+        /* .async                 = */ false,
+        /* .host_buffer           = */ false,
+        /* .buffer_from_host_ptr  = */ false,
+        /* .events                = */ false,
+    };
+}
+
+static ggml_backend_t ggml_backend_amx_device_init(ggml_backend_dev_t dev, const char * params) {
+    return ggml_backend_amx_init();
+
+    GGML_UNUSED(dev);
+    GGML_UNUSED(params);
+}
+
+static ggml_backend_buffer_type_t ggml_backend_amx_device_get_buffer_type(ggml_backend_dev_t dev) {
+    return ggml_backend_amx_buffer_type();
+
+    GGML_UNUSED(dev);
+}
+
+static bool ggml_backend_amx_device_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) {
+
+    // handle only 2d gemm for now
+    auto is_contiguous_2d = [](const struct ggml_tensor * t) {
+        return ggml_is_contiguous(t) && t->ne[3] == 1 && t->ne[2] == 1;
+    };
+
+    switch (op->op) {
+        case GGML_OP_NONE:
+        case GGML_OP_RESHAPE:
+        case GGML_OP_VIEW:
+        case GGML_OP_PERMUTE:
+        case GGML_OP_TRANSPOSE:
+            return true;
+
+        case GGML_OP_MUL_MAT: {
+            const struct ggml_tensor * src0 = op->src[0];
+            const struct ggml_tensor * src1 = op->src[1];
+
+            const enum ggml_type type = src0->type;
+            const int64_t ne0 = op->ne[0];
+
+            bool is_training = src0->grad || src1->grad;
+
+            // amx kernels enables for Q4_0, Q4_1, Q8_0, F16
+            // Q4_K, Q5_K, Q6_K, IQ4_XS enabled for QK_K = 256
+            bool has_amx_kernels = qtype_has_amx_kernels(type) || (type == GGML_TYPE_F16);
+
+            bool can_use_amx =
+                is_contiguous_2d(src0) &&       // src0 must be contiguous
+                is_contiguous_2d(src1) &&       // src1 must be contiguous
+                !is_training &&                 // inference only
+                src1->type == GGML_TYPE_F32 &&  // src1 must be float32
+                has_amx_kernels &&              // with amx kernel impls
+                ne0 % (TILE_N * 2) == 0;        // out_features is 32x
+
+            return can_use_amx;
+        }
+        default:
+            return false;
+    }
+
+    GGML_UNUSED(dev);
+}
+
+static bool ggml_backend_amx_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
+    return buft->iface.get_name == ggml_backend_amx_buffer_type_get_name;
+
+    GGML_UNUSED(dev);
+}
+
+static const struct ggml_backend_device_i ggml_backend_amx_device_i = {
+    /* .get_name             = */ ggml_backend_amx_device_get_name,
+    /* .get_description      = */ ggml_backend_amx_device_get_description,
+    /* .get_memory           = */ ggml_backend_amx_device_get_memory,
+    /* .get_type             = */ ggml_backend_amx_device_get_type,
+    /* .get_props            = */ ggml_backend_amx_device_get_props,
+    /* .init_backend         = */ ggml_backend_amx_device_init,
+    /* .get_buffer_type      = */ ggml_backend_amx_device_get_buffer_type,
+    /* .get_host_buffer_type = */ NULL,
+    /* .buffer_from_host_ptr = */ NULL,
+    /* .supports_op          = */ ggml_backend_amx_device_supports_op,
+    /* .supports_buft        = */ ggml_backend_amx_device_supports_buft,
+    /* .offload_op           = */ NULL,
+    /* .event_new            = */ NULL,
+    /* .event_free           = */ NULL,
+    /* .event_synchronize    = */ NULL,
+};
+
+// backend reg interface
+
+static const char * ggml_backend_amx_reg_get_name(ggml_backend_reg_t reg) {
+    return "AMX";
+
+    GGML_UNUSED(reg);
+}
+
+static size_t ggml_backend_amx_reg_get_device_count(ggml_backend_reg_t reg) {
+    return 1;
+
+    GGML_UNUSED(reg);
+}
+
+static ggml_backend_dev_t ggml_backend_amx_reg_get_device(ggml_backend_reg_t reg, size_t index) {
+    GGML_ASSERT(index == 0);
+
+    static ggml_backend_device ggml_backend_amx_device = {
+        /* .iface   = */ ggml_backend_amx_device_i,
+        /* .reg     = */ reg,
+        /* .context = */ nullptr,
+    };
+
+    return &ggml_backend_amx_device;
+
+    GGML_UNUSED(reg);
+    GGML_UNUSED(index);
+}
+
+static void * ggml_backend_amx_get_proc_address(ggml_backend_reg_t reg, const char * name) {
+    if (std::strcmp(name, "ggml_backend_set_n_threads") == 0) {
+        return (void *)ggml_backend_amx_set_n_threads;
+    }
+    return NULL;
+
+    GGML_UNUSED(reg);
+    GGML_UNUSED(name);
+}
+
+static const struct ggml_backend_reg_i ggml_backend_amx_reg_i = {
+    /* .get_name         = */ ggml_backend_amx_reg_get_name,
+    /* .get_device_count = */ ggml_backend_amx_reg_get_device_count,
+    /* .get_device       = */ ggml_backend_amx_reg_get_device,
+    /* .get_proc_address = */ ggml_backend_amx_get_proc_address,
+};
+
+ggml_backend_reg_t ggml_backend_amx_reg(void) {
+    static struct ggml_backend_reg ggml_backend_amx_reg = {
+        /* .iface   = */ ggml_backend_amx_reg_i,
+        /* .context = */ NULL,
+    };
+
+    return &ggml_backend_amx_reg;
+}
+
+#else // if defined(__AMX_INT8__)
+
+ggml_backend_buffer_type_t ggml_backend_amx_buffer_type(void) {
+    return nullptr;
+}
+
+bool ggml_backend_is_amx(ggml_backend_t backend) {
+    GGML_UNUSED(backend);
+    return false;
+}
+
+ggml_backend_t ggml_backend_amx_init(void) {
+    fprintf(stderr, "GGML is not compiled with AMX support!\n");
+    return nullptr;
+}
+
+void ggml_backend_amx_set_n_threads(ggml_backend_t backend_amx, int n_threads) {
+    fprintf(stderr, "GGML is not compiled with AMX support!\n");
+
+    GGML_UNUSED(backend_amx);
+    GGML_UNUSED(n_threads);
+}
+
+ggml_backend_reg_t ggml_backend_amx_reg(void) {
+    return nullptr;
+}
+
+#endif

ggml/src/ggml-amx/mmq.cpp

@@ -496,19 +496,20 @@ inline void from_float(const float * x, char * vy, int64_t k);
 
 template <>
 inline void from_float<block_q8_0>(const float * x, char * vy, int64_t k) {
-    quantize_row_q8_0(x, vy, k);
+    // FIXME: using unoptimized reference impl until moved to CPU backend
+    quantize_row_q8_0_ref(x, (block_q8_0 *)vy, k);
 }
 
 template <>
 inline void from_float<block_q8_1>(const float * x, char * vy, int64_t k) {
-    quantize_row_q8_1(x, vy, k);
+    quantize_row_q8_1_ref(x, (block_q8_1 *)vy, k);
 }
 
 template <>
 inline void from_float<block_q8_K>(const float * x, char * vy, int64_t k) {
 #if 1
     // TODO: this is reference impl!
-    quantize_row_q8_K(x, vy, k);
+    quantize_row_q8_K_ref(x, (block_q8_K *)vy, k);
 #else
     quantize_row_q8_K_vnni(x, vy, k);
 #endif

ggml/src/ggml-backend-reg.cpp

@@ -0,0 +1,195 @@
+#include "ggml-backend-impl.h"
+#include "ggml-backend.h"
+#include "ggml-cpu.h"
+#include "ggml-impl.h"
+#include <cstring>
+#include <vector>
+
+// Backend registry
+
+#ifdef GGML_USE_CUDA
+#include "ggml-cuda.h"
+#endif
+
+#ifdef GGML_USE_METAL
+#include "ggml-metal.h"
+#endif
+
+#ifdef GGML_USE_SYCL
+#include "ggml-sycl.h"
+#endif
+
+#ifdef GGML_USE_VULKAN
+#include "ggml-vulkan.h"
+#endif
+
+#ifdef GGML_USE_BLAS
+#include "ggml-blas.h"
+#endif
+
+#ifdef GGML_USE_RPC
+#include "ggml-rpc.h"
+#endif
+
+#ifdef GGML_USE_AMX
+#  include "ggml-amx.h"
+#endif
+
+#ifdef GGML_USE_CANN
+#include "ggml-cann.h"
+#endif
+
+#ifdef GGML_USE_KOMPUTE
+#include "ggml-kompute.h"
+#endif
+
+struct ggml_backend_registry {
+    std::vector<ggml_backend_reg_t> backends;
+    std::vector<ggml_backend_dev_t> devices;
+
+    ggml_backend_registry() {
+#ifdef GGML_USE_CUDA
+        register_backend(ggml_backend_cuda_reg());
+#endif
+#ifdef GGML_USE_METAL
+        register_backend(ggml_backend_metal_reg());
+#endif
+#ifdef GGML_USE_SYCL
+        register_backend(ggml_backend_sycl_reg());
+#endif
+#ifdef GGML_USE_VULKAN
+        register_backend(ggml_backend_vk_reg());
+#endif
+#ifdef GGML_USE_CANN
+        register_backend(ggml_backend_cann_reg());
+#endif
+#ifdef GGML_USE_BLAS
+        register_backend(ggml_backend_blas_reg());
+#endif
+#ifdef GGML_USE_RPC
+        register_backend(ggml_backend_rpc_reg());
+#endif
+#ifdef GGML_USE_AMX
+        register_backend(ggml_backend_amx_reg());
+#endif
+#ifdef GGML_USE_KOMPUTE
+        register_backend(ggml_backend_kompute_reg());
+#endif
+
+        register_backend(ggml_backend_cpu_reg());
+    }
+
+    void register_backend(ggml_backend_reg_t reg) {
+        if (!reg) {
+            return;
+        }
+
+#ifndef NDEBUG
+        GGML_LOG_DEBUG("%s: registered backend %s (%zu devices)\n",
+            __func__, ggml_backend_reg_name(reg), ggml_backend_reg_dev_count(reg));
+#endif
+        backends.push_back(reg);
+        for (size_t i = 0; i < ggml_backend_reg_dev_count(reg); i++) {
+            register_device(ggml_backend_reg_dev_get(reg, i));
+        }
+    }
+
+    void register_device(ggml_backend_dev_t device) {
+#ifndef NDEBUG
+        GGML_LOG_DEBUG("%s: registered device %s (%s)\n", __func__, ggml_backend_dev_name(device), ggml_backend_dev_description(device));
+#endif
+        devices.push_back(device);
+    }
+};
+
+static ggml_backend_registry & get_reg() {
+    static ggml_backend_registry reg;
+    return reg;
+}
+
+// Internal API
+void ggml_backend_register(ggml_backend_reg_t reg) {
+    get_reg().register_backend(reg);
+}
+
+void ggml_backend_device_register(ggml_backend_dev_t device) {
+    get_reg().register_device(device);
+}
+
+// Backend (reg) enumeration
+size_t ggml_backend_reg_count() {
+    return get_reg().backends.size();
+}
+
+ggml_backend_reg_t ggml_backend_reg_get(size_t index) {
+    GGML_ASSERT(index < ggml_backend_reg_count());
+    return get_reg().backends[index];
+}
+
+ggml_backend_reg_t ggml_backend_reg_by_name(const char * name) {
+    for (size_t i = 0; i < ggml_backend_reg_count(); i++) {
+        ggml_backend_reg_t reg = ggml_backend_reg_get(i);
+        if (std::strcmp(ggml_backend_reg_name(reg), name) == 0) {
+            return reg;
+        }
+    }
+    return NULL;
+}
+
+// Device enumeration
+size_t ggml_backend_dev_count() {
+    return get_reg().devices.size();
+}
+
+ggml_backend_dev_t ggml_backend_dev_get(size_t index) {
+    GGML_ASSERT(index < ggml_backend_dev_count());
+    return get_reg().devices[index];
+}
+
+ggml_backend_dev_t ggml_backend_dev_by_name(const char * name) {
+    for (size_t i = 0; i < ggml_backend_dev_count(); i++) {
+        ggml_backend_dev_t dev = ggml_backend_dev_get(i);
+        if (strcmp(ggml_backend_dev_name(dev), name) == 0) {
+            return dev;
+        }
+    }
+    return NULL;
+}
+
+ggml_backend_dev_t ggml_backend_dev_by_type(enum ggml_backend_dev_type type) {
+    for (size_t i = 0; i < ggml_backend_dev_count(); i++) {
+        ggml_backend_dev_t dev = ggml_backend_dev_get(i);
+        if (ggml_backend_dev_type(dev) == type) {
+            return dev;
+        }
+    }
+    return NULL;
+}
+
+// Convenience functions
+ggml_backend_t ggml_backend_init_by_name(const char * name, const char * params) {
+    ggml_backend_dev_t dev = ggml_backend_dev_by_name(name);
+    if (!dev) {
+        return NULL;
+    }
+    return ggml_backend_dev_init(dev, params);
+}
+
+ggml_backend_t ggml_backend_init_by_type(enum ggml_backend_dev_type type, const char * params) {
+    ggml_backend_dev_t dev = ggml_backend_dev_by_type(type);
+    if (!dev) {
+        return NULL;
+    }
+    return ggml_backend_dev_init(dev, params);
+}
+
+ggml_backend_t ggml_backend_init_best(void) {
+    ggml_backend_dev_t dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_GPU);
+    if (!dev) {
+        dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+    }
+    if (!dev) {
+        return NULL;
+    }
+    return ggml_backend_dev_init(dev, NULL);
+}

ggml/src/ggml-blas/CMakeLists.txt

@@ -0,0 +1,91 @@
+if (GGML_STATIC)
+    set(BLA_STATIC ON)
+endif()
+#if (CMAKE_VERSION VERSION_GREATER_EQUAL 3.22)
+#    set(BLA_SIZEOF_INTEGER 8)
+#endif()
+
+set(BLA_VENDOR ${GGML_BLAS_VENDOR})
+find_package(BLAS)
+
+if (BLAS_FOUND)
+    message(STATUS "BLAS found, Libraries: ${BLAS_LIBRARIES}")
+
+    add_library(ggml-blas
+                ggml-blas.cpp
+                )
+
+    target_link_libraries(ggml-blas PRIVATE ggml-base)
+    target_include_directories(ggml-blas PRIVATE . ..)
+
+    if (${GGML_BLAS_VENDOR} MATCHES "Apple")
+        add_compile_definitions(ACCELERATE_NEW_LAPACK)
+        add_compile_definitions(ACCELERATE_LAPACK_ILP64)
+        add_compile_definitions(GGML_BLAS_USE_ACCELERATE)
+    elseif ("${BLAS_INCLUDE_DIRS}" STREQUAL "")
+        # BLAS_INCLUDE_DIRS is missing in FindBLAS.cmake.
+        # see https://gitlab.kitware.com/cmake/cmake/-/issues/20268
+        find_package(PkgConfig REQUIRED)
+        if (${GGML_BLAS_VENDOR} MATCHES "Generic")
+            pkg_check_modules(DepBLAS blas)
+        elseif (${GGML_BLAS_VENDOR} MATCHES "OpenBLAS")
+            # As of openblas v0.3.22, the 64-bit is named openblas64.pc
+            pkg_check_modules(DepBLAS openblas64)
+            if (NOT DepBLAS_FOUND)
+                pkg_check_modules(DepBLAS openblas)
+            endif()
+        elseif (${GGML_BLAS_VENDOR} MATCHES "FLAME")
+            add_compile_definitions(GGML_BLAS_USE_BLIS)
+            pkg_check_modules(DepBLAS blis)
+        elseif (${GGML_BLAS_VENDOR} MATCHES "ATLAS")
+            pkg_check_modules(DepBLAS blas-atlas)
+        elseif (${GGML_BLAS_VENDOR} MATCHES "FlexiBLAS")
+            pkg_check_modules(DepBLAS flexiblas_api)
+        elseif (${GGML_BLAS_VENDOR} MATCHES "Intel")
+            add_compile_definitions(GGML_BLAS_USE_MKL)
+            # all Intel* libraries share the same include path
+            pkg_check_modules(DepBLAS mkl-sdl)
+        elseif (${GGML_BLAS_VENDOR} MATCHES "NVHPC")
+            # this doesn't provide pkg-config
+            # suggest to assign BLAS_INCLUDE_DIRS on your own
+            if ("${NVHPC_VERSION}" STREQUAL "")
+                message(WARNING "Better to set NVHPC_VERSION")
+            else()
+                set(DepBLAS_FOUND ON)
+                set(DepBLAS_INCLUDE_DIRS "/opt/nvidia/hpc_sdk/${CMAKE_SYSTEM_NAME}_${CMAKE_SYSTEM_PROCESSOR}/${NVHPC_VERSION}/math_libs/include")
+            endif()
+        endif()
+        if (DepBLAS_FOUND)
+            set(BLAS_INCLUDE_DIRS ${DepBLAS_INCLUDE_DIRS})
+        else()
+            message(WARNING "BLAS_INCLUDE_DIRS neither been provided nor been automatically"
+            " detected by pkgconfig, trying to find cblas.h from possible paths...")
+            find_path(BLAS_INCLUDE_DIRS
+                NAMES cblas.h
+                HINTS
+                    /usr/include
+                    /usr/local/include
+                    /usr/include/openblas
+                    /opt/homebrew/opt/openblas/include
+                    /usr/local/opt/openblas/include
+                    /usr/include/x86_64-linux-gnu/openblas/include
+            )
+        endif()
+    endif()
+
+    message(STATUS "BLAS found, Includes: ${BLAS_INCLUDE_DIRS}")
+
+    #add_compile_options(${BLAS_LINKER_FLAGS})
+    target_compile_options(ggml-blas PRIVATE ${BLAS_LINKER_FLAGS})
+
+    if (${BLAS_INCLUDE_DIRS} MATCHES "mkl" AND (${GGML_BLAS_VENDOR} MATCHES "Generic" OR ${GGML_BLAS_VENDOR} MATCHES "Intel"))
+        add_compile_definitions(GGML_BLAS_USE_MKL)
+    endif()
+
+    target_link_libraries     (ggml-blas PRIVATE ${BLAS_LIBRARIES})
+    target_include_directories(ggml-blas PRIVATE ${BLAS_INCLUDE_DIRS})
+else()
+    message(ERROR "BLAS not found, please refer to "
+                  "https://cmake.org/cmake/help/latest/module/FindBLAS.html#blas-lapack-vendors"
+                  " to set correct GGML_BLAS_VENDOR")
+endif()

ggml/src/ggml-blas/ggml-blas.cpp

@@ -0,0 +1,514 @@
+#include "ggml-impl.h"
+#include "ggml-blas.h"
+#include "ggml-backend-impl.h"
+
+#include <future>
+#include <vector>
+#include <cstring>
+
+#if defined(GGML_BLAS_USE_ACCELERATE)
+#   include <Accelerate/Accelerate.h>
+#elif defined(GGML_BLAS_USE_MKL)
+#   include <mkl.h>
+#elif defined(GGML_BLAS_USE_BLIS)
+#   include <blis.h>
+#elif defined(GGML_BLAS_USE_NVPL)
+#   include <nvpl_blas.h>
+#else
+#   include <cblas.h>
+#endif
+
+struct ggml_backend_blas_context {
+    int n_threads = GGML_DEFAULT_N_THREADS;
+    std::unique_ptr<char[]> work_data;
+    size_t work_size = 0;
+#ifndef GGML_USE_OPENMP
+    std::vector<std::future<void>> tasks;
+#endif
+};
+
+static void ggml_backend_blas_mul_mat(ggml_backend_blas_context * ctx, struct ggml_tensor * dst) {
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    const enum ggml_type type = src0->type;
+
+    GGML_ASSERT(ne0 == ne01);
+    GGML_ASSERT(ne1 == ne11);
+    GGML_ASSERT(ne2 == ne12);
+    GGML_ASSERT(ne3 == ne13);
+
+    // we don't support permuted src0 or src1
+    GGML_ASSERT(nb00 == ggml_type_size(type));
+    GGML_ASSERT(nb10 == ggml_type_size(src1->type));
+
+    // dst cannot be transposed or permuted
+    GGML_ASSERT(nb0 == sizeof(float));
+    GGML_ASSERT(nb0 <= nb1);
+    GGML_ASSERT(nb1 <= nb2);
+    GGML_ASSERT(nb2 <= nb3);
+
+    // broadcast factors
+    const int64_t r2 = ne12/ne02;
+    const int64_t r3 = ne13/ne03;
+
+    const int64_t ne_plane      = ne01*ne00;
+    const size_t  desired_wsize = type == GGML_TYPE_F32 ? 0 : ne03*ne02*ne_plane*sizeof(float);
+
+    if (ctx->work_size < desired_wsize) {
+        ctx->work_data.reset(new char[desired_wsize]);
+        ctx->work_size = desired_wsize;
+    }
+    void * wdata = ctx->work_data.get();
+
+    // convert src0 to float
+    if (type != GGML_TYPE_F32) {
+        const auto * type_traits = ggml_get_type_traits(type);
+        ggml_to_float_t const to_float = type_traits->to_float;
+
+        for (int64_t i03 = 0; i03 < ne03; i03++) {
+            for (int64_t i02 = 0; i02 < ne02; i02++) {
+                const void  *       x      = (char *)  src0->data + i02*nb02          + i03*nb03;
+                      float * const wplane = (float *) wdata      + i02*ne_plane      + i03*ne02*ne_plane;
+
+                const int min_cols_per_thread = 4096;
+                const int min_rows_per_thread = std::max((int)(min_cols_per_thread/ne00), 1);
+                const int n_threads = std::max(std::min(ctx->n_threads, (int)(ne01/min_rows_per_thread)), 1);
+
+#ifdef GGML_USE_OPENMP
+                #pragma omp parallel for num_threads(n_threads)
+                for (int64_t i01 = 0; i01 < ne01; i01++) {
+                    to_float((const char *) x + i01*nb01, wplane + i01*ne00, ne00);
+                }
+#else
+                for (int i = 1; i < n_threads; i++) {
+                    const int64_t start =       i*ne01/n_threads;
+                    const int64_t end   = (i + 1)*ne01/n_threads;
+                    if (start < end) {
+                        ctx->tasks.push_back(std::async(std::launch::async, [=]() {
+                            for (int64_t i01 = start; i01 < end; i01++) {
+                                to_float((const char *) x + i01*nb01, wplane + i01*ne00, ne00);
+                            }
+                        }));
+                    }
+                }
+                {
+                    // reuse the current thread for the first task
+                    const int64_t start = 0;
+                    const int64_t end   = ne01/n_threads;
+                    for (int64_t i01 = start; i01 < end; i01++) {
+                        to_float((const char *) x + i01*nb01, wplane + i01*ne00, ne00);
+                    }
+                }
+#endif
+            }
+        }
+
+#ifndef GGML_USE_OPENMP
+        // wait for all tasks to finish
+        for (auto & task : ctx->tasks) {
+            task.get();
+        }
+        ctx->tasks.clear();
+#endif
+    }
+
+#if defined(OPENBLAS_VERSION)
+    openblas_set_num_threads(ctx->n_threads);
+#endif
+
+#if defined(GGML_BLAS_USE_BLIS)
+    bli_thread_set_num_threads(ctx->n_threads);
+#endif
+
+#if defined(GGML_BLAS_USE_NVPL)
+    nvpl_blas_set_num_threads(ctx->n_threads);
+#endif
+
+    for (int64_t i13 = 0; i13 < ne13; i13++) {
+        for (int64_t i12 = 0; i12 < ne12; i12++) {
+            const int64_t i03 = i13/r3;
+            const int64_t i02 = i12/r2;
+
+            const float * x = (float *) ((char *) src0->data + i02*nb02 + i03*nb03);
+            const float * y = (float *) ((char *) src1->data + i12*nb12 + i13*nb13);
+                  float * d = (float *) ((char *)  dst->data + i12*nb2  + i13*nb3);
+
+            if (type != GGML_TYPE_F32) {
+                x = (float *) wdata + i02*ne_plane + i03*ne02*ne_plane;
+            }
+
+            cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasTrans,
+                        ne1, ne01, ne10,
+                        1.0f,   y, ne10,
+                                x, ne00,
+                        0.0f,   d, ne01);
+        }
+    }
+}
+
+static void ggml_backend_blas_out_prod(ggml_backend_blas_context * ctx, struct ggml_tensor * dst) {
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    GGML_ASSERT(ne0  == ne00);
+    GGML_ASSERT(ne1  == ne10);
+    GGML_ASSERT(ne2  == ne02);
+    GGML_ASSERT(ne02 == ne12);
+    GGML_ASSERT(ne3  == ne13);
+    GGML_ASSERT(ne03 == ne13);
+
+    // we don't support permuted src0 or src1
+    GGML_ASSERT(nb00 == sizeof(float));
+
+    // dst cannot be transposed or permuted
+    GGML_ASSERT(nb0 == sizeof(float));
+    // GGML_ASSERT(nb0 <= nb1);
+    // GGML_ASSERT(nb1 <= nb2);
+    // GGML_ASSERT(nb2 <= nb3);
+
+    // Arguments to ggml_compute_forward_out_prod (expressed as major,minor)
+    // src0: (k,n)
+    // src1: (k,m)
+    // dst:  (m,n)
+    //
+    // Arguments to sgemm (see https://github.com/Reference-LAPACK/lapack/blob/master/BLAS/SRC/sgemm.f)
+    // Also expressed as (major,minor)
+    // a: (m,k): so src1 transposed
+    // b: (k,n): so src0
+    // c: (m,n)
+    //
+    // However, if ggml_is_transposed(src1) is true, then
+    // src1->data already contains a transposed version, so sgemm mustn't
+    // transpose it further.
+
+    int n = src0->ne[0];
+    int k = src0->ne[1];
+    int m = src1->ne[0];
+
+    CBLAS_TRANSPOSE transposeA;
+    int lda;
+
+    if (!ggml_is_transposed(src1)) {
+        transposeA = CblasTrans;
+        lda = m;
+    } else {
+        transposeA = CblasNoTrans;
+        lda = k;
+    }
+
+    float * a = (float *) ((char *) src1->data);
+    float * b = (float *) ((char *) src0->data);
+    float * c = (float *) ((char *) dst->data);
+
+    cblas_sgemm(CblasRowMajor, transposeA, CblasNoTrans, m, n, k, 1.0, a, lda, b, n, 0.0, c, n);
+
+    GGML_UNUSED(ctx);
+}
+
+// backend interface
+
+static const char * ggml_backend_blas_get_name(ggml_backend_t backend) {
+    return "BLAS";
+
+    GGML_UNUSED(backend);
+}
+
+static void ggml_backend_blas_free(ggml_backend_t backend) {
+    ggml_backend_blas_context * ctx = (ggml_backend_blas_context *)backend->context;
+    delete ctx;
+    delete backend;
+}
+
+static enum ggml_status ggml_backend_blas_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    ggml_backend_blas_context * ctx = (ggml_backend_blas_context *)backend->context;
+
+    for (int i = 0; i < cgraph->n_nodes; i++) {
+        struct ggml_tensor * node = cgraph->nodes[i];
+
+        switch (node->op) {
+            case GGML_OP_MUL_MAT:
+                ggml_backend_blas_mul_mat(ctx, node);
+                break;
+
+            case GGML_OP_OUT_PROD:
+                ggml_backend_blas_out_prod(ctx, node);
+                break;
+
+            case GGML_OP_NONE:
+            case GGML_OP_RESHAPE:
+            case GGML_OP_VIEW:
+            case GGML_OP_PERMUTE:
+            case GGML_OP_TRANSPOSE:
+                break;
+
+            default:
+                GGML_ABORT("%s: unsupported op %s\n", __func__, ggml_op_desc(node));
+        }
+    }
+
+    return GGML_STATUS_SUCCESS;
+
+    GGML_UNUSED(backend);
+}
+
+static struct ggml_backend_i blas_backend_i = {
+    /* .get_name                = */ ggml_backend_blas_get_name,
+    /* .free                    = */ ggml_backend_blas_free,
+    /* .set_tensor_async        = */ NULL,
+    /* .get_tensor_async        = */ NULL,
+    /* .cpy_tensor_async        = */ NULL,
+    /* .synchronize             = */ NULL,
+    /* .graph_plan_create       = */ NULL,
+    /* .graph_plan_free         = */ NULL,
+    /* .graph_plan_update       = */ NULL,
+    /* .graph_plan_compute      = */ NULL,
+    /* .graph_compute           = */ ggml_backend_blas_graph_compute,
+    /* .event_record            = */ NULL,
+    /* .event_wait              = */ NULL,
+};
+
+static ggml_guid_t ggml_backend_blas_guid(void) {
+    static ggml_guid guid = { 0x12, 0xa8, 0xae, 0xf4, 0xc0, 0x1e, 0x61, 0x97, 0x8f, 0xeb, 0x33, 0x04, 0xa1, 0x33, 0x51, 0x2d };
+    return &guid;
+}
+
+ggml_backend_t ggml_backend_blas_init(void) {
+    ggml_backend_blas_context * ctx = new ggml_backend_blas_context;
+
+    ggml_backend_t backend = new ggml_backend {
+        /* .guid      = */ ggml_backend_blas_guid(),
+        /* .interface = */ blas_backend_i,
+        /* .device    = */ ggml_backend_reg_dev_get(ggml_backend_blas_reg(), 0),
+        /* .context   = */ ctx,
+    };
+
+#if defined(OPENBLAS_VERSION) && defined(GGML_USE_OPENMP)
+    if (openblas_get_parallel() != OPENBLAS_OPENMP) {
+        GGML_LOG_DEBUG("%s: warning: ggml is using OpenMP, but OpenBLAS was compiled without OpenMP support\n", __func__);
+    }
+#endif
+
+#if defined(BLIS_ENABLE_CBLAS) && defined(GGML_USE_OPENMP) && !defined(BLIS_ENABLE_OPENMP)
+    GGML_LOG_DEBUG("%s: warning: ggml is using OpenMP, but BLIS was compiled without OpenMP support\n", __func__);
+#endif
+
+    return backend;
+}
+
+bool ggml_backend_is_blas(ggml_backend_t backend) {
+    return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_blas_guid());
+}
+
+void ggml_backend_blas_set_n_threads(ggml_backend_t backend_blas, int n_threads) {
+    GGML_ASSERT(ggml_backend_is_blas(backend_blas));
+
+    ggml_backend_blas_context * ctx = (ggml_backend_blas_context *)backend_blas->context;
+    ctx->n_threads = n_threads;
+}
+
+// device interface
+
+static const char * ggml_backend_blas_device_get_name(ggml_backend_dev_t dev) {
+    return "BLAS";
+
+    GGML_UNUSED(dev);
+}
+
+static const char * ggml_backend_blas_device_get_description(ggml_backend_dev_t dev) {
+    #if defined(GGML_BLAS_USE_ACCELERATE)
+        return "Accelerate";
+    #elif defined(GGML_BLAS_USE_MKL)
+        return "MKL";
+    #elif defined(GGML_BLAS_USE_BLIS)
+        return "BLIS";
+    #elif defined(GGML_BLAS_USE_NVPL)
+        return "NVPL";
+    #elif defined(OPENBLAS_VERSION)
+        return "OpenBLAS";
+    #else
+        return "BLAS";
+    #endif
+
+    GGML_UNUSED(dev);
+}
+
+static void ggml_backend_blas_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
+    // TODO
+    *free = 0;
+    *total = 0;
+
+    GGML_UNUSED(dev);
+}
+
+static enum ggml_backend_dev_type ggml_backend_blas_device_get_type(ggml_backend_dev_t dev) {
+    return GGML_BACKEND_DEVICE_TYPE_ACCEL;
+
+    GGML_UNUSED(dev);
+}
+
+static void ggml_backend_blas_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) {
+    props->name        = ggml_backend_blas_device_get_name(dev);
+    props->description = ggml_backend_blas_device_get_description(dev);
+    props->type        = ggml_backend_blas_device_get_type(dev);
+    ggml_backend_blas_device_get_memory(dev, &props->memory_free, &props->memory_total);
+    props->caps = {
+        /* .async                 = */ false,
+        /* .host_buffer           = */ false,
+        /* .buffer_from_host_ptr  = */ true,
+        /* .events                = */ false,
+    };
+}
+
+static ggml_backend_t ggml_backend_blas_device_init_backend(ggml_backend_dev_t dev, const char * params) {
+    return ggml_backend_blas_init();
+
+    GGML_UNUSED(dev);
+    GGML_UNUSED(params);
+}
+
+static ggml_backend_buffer_type_t ggml_backend_blas_device_get_buffer_type(ggml_backend_dev_t dev) {
+    return ggml_backend_cpu_buffer_type();
+
+    GGML_UNUSED(dev);
+}
+
+static ggml_backend_buffer_t ggml_backend_blas_device_buffer_from_host_ptr(ggml_backend_dev_t dev, void * ptr, size_t size, size_t max_tensor_size) {
+    return ggml_backend_cpu_buffer_from_ptr(ptr, size);
+
+    GGML_UNUSED(dev);
+    GGML_UNUSED(max_tensor_size);
+}
+
+static bool ggml_backend_blas_device_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) {
+    const struct ggml_tensor * src0 = op->src[0];
+    const struct ggml_tensor * src1 = op->src[1];
+
+    switch (op->op) {
+        case GGML_OP_NONE:
+        case GGML_OP_RESHAPE:
+        case GGML_OP_VIEW:
+        case GGML_OP_PERMUTE:
+        case GGML_OP_TRANSPOSE:
+            return true;
+
+        case GGML_OP_MUL_MAT:
+        {
+            // BLAS usually is only faster for large matrices
+            const struct ggml_tensor * src0 = op->src[0];
+            const struct ggml_tensor * src1 = op->src[1];
+
+            const int64_t ne10 = src1->ne[0];
+
+            const int64_t ne0 = op->ne[0];
+            const int64_t ne1 = op->ne[1];
+
+            // TODO: find the optimal value
+            const int64_t min_batch = 32;
+
+            return ggml_is_contiguous(src0) &&
+                   ggml_is_contiguous(src1) &&
+                   src1->type == GGML_TYPE_F32 &&
+                   (ne0 >= min_batch && ne1 >= min_batch && ne10 >= min_batch) &&
+                   (src0->type == GGML_TYPE_F32 || ggml_get_type_traits(src0->type)->to_float != NULL);
+        }
+
+        case GGML_OP_OUT_PROD:
+            return op->src[0]->type == GGML_TYPE_F32 &&
+                   op->src[1]->type == GGML_TYPE_F32 &&
+                   ggml_is_matrix(src0) &&
+                   ggml_is_matrix(src1) &&
+                   ggml_is_contiguous(src0) &&
+                   (ggml_is_contiguous(src1) || ggml_is_transposed(src1)) &&
+                   (src0->type == GGML_TYPE_F32 || ggml_get_type_traits(src0->type)->to_float != NULL);
+
+        default:
+            return false;
+
+    }
+
+    GGML_UNUSED(dev);
+}
+
+static bool ggml_backend_blas_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
+    return ggml_backend_buft_is_host(buft);
+
+    GGML_UNUSED(dev);
+}
+
+static const struct ggml_backend_device_i ggml_backend_blas_device_i = {
+    /* .get_name             = */ ggml_backend_blas_device_get_name,
+    /* .get_description      = */ ggml_backend_blas_device_get_description,
+    /* .get_memory           = */ ggml_backend_blas_device_get_memory,
+    /* .get_type             = */ ggml_backend_blas_device_get_type,
+    /* .get_props            = */ ggml_backend_blas_device_get_props,
+    /* .init_backend         = */ ggml_backend_blas_device_init_backend,
+    /* .get_buffer_type      = */ ggml_backend_blas_device_get_buffer_type,
+    /* .get_host_buffer_type = */ NULL,
+    /* .buffer_from_host_ptr = */ ggml_backend_blas_device_buffer_from_host_ptr,
+    /* .supports_op          = */ ggml_backend_blas_device_supports_op,
+    /* .supports_buft        = */ ggml_backend_blas_device_supports_buft,
+    /* .offload_op           = */ NULL,
+    /* .event_new            = */ NULL,
+    /* .event_free           = */ NULL,
+    /* .event_synchronize    = */ NULL,
+};
+
+// backend reg interface
+
+static const char * ggml_backend_blas_reg_get_name(ggml_backend_reg_t reg) {
+    return "BLAS";
+
+    GGML_UNUSED(reg);
+}
+
+static size_t ggml_backend_blas_reg_get_device_count(ggml_backend_reg_t reg) {
+    return 1;
+
+    GGML_UNUSED(reg);
+}
+
+static ggml_backend_dev_t ggml_backend_blas_reg_get_device(ggml_backend_reg_t reg, size_t index) {
+    GGML_ASSERT(index == 0);
+
+    static ggml_backend_device ggml_backend_blas_device = {
+        /* .iface   = */ ggml_backend_blas_device_i,
+        /* .reg     = */ reg,
+        /* .context = */ nullptr,
+    };
+
+    return &ggml_backend_blas_device;
+
+    GGML_UNUSED(reg);
+    GGML_UNUSED(index);
+}
+
+static void * ggml_backend_blas_get_proc_address(ggml_backend_reg_t reg, const char * name) {
+    if (std::strcmp(name, "ggml_backend_set_n_threads") == 0) {
+        return (void *)ggml_backend_blas_set_n_threads;
+    }
+    return NULL;
+
+    GGML_UNUSED(reg);
+    GGML_UNUSED(name);
+}
+
+static const struct ggml_backend_reg_i ggml_backend_blas_reg_i = {
+    /* .get_name         = */ ggml_backend_blas_reg_get_name,
+    /* .get_device_count = */ ggml_backend_blas_reg_get_device_count,
+    /* .get_device       = */ ggml_backend_blas_reg_get_device,
+    /* .get_proc_address = */ ggml_backend_blas_get_proc_address,
+};
+
+ggml_backend_reg_t ggml_backend_blas_reg(void) {
+    static struct ggml_backend_reg ggml_backend_blas_reg = {
+        /* .iface   = */ ggml_backend_blas_reg_i,
+        /* .context = */ NULL,
+    };
+
+    return &ggml_backend_blas_reg;
+}

ggml/src/ggml-cann/CMakeLists.txt

@@ -0,0 +1,46 @@
+if ("cann${CANN_INSTALL_DIR}" STREQUAL "cann" AND DEFINED ENV{ASCEND_TOOLKIT_HOME})
+    set(CANN_INSTALL_DIR $ENV{ASCEND_TOOLKIT_HOME})
+    message(STATUS "CANN: updated CANN_INSTALL_DIR from ASCEND_TOOLKIT_HOME=$ENV{ASCEND_TOOLKIT_HOME}")
+endif()
+
+if (CANN_INSTALL_DIR)
+    # Only Support Linux.
+    if (NOT UNIX)
+        message(FATAL_ERROR "CANN: CANN toolkit supports unix but not ${CMAKE_SYSTEM_NAME}")
+    endif()
+
+    # Supported platforms: x86-64, arm64
+    if (CMAKE_SYSTEM_PROCESSOR STREQUAL "aarch64")
+    elseif (CMAKE_SYSTEM_PROCESSOR STREQUAL "x86_64" OR CMAKE_SYSTEM_PROCESSOR STREQUAL "amd64")
+    else()
+        message(FATAL_ERROR "CANN: CANN toolkit supports x86-64 and arm64 but not ${CMAKE_SYSTEM_PROCESSOR}")
+    endif()
+
+    # Set header and libs
+    set(CANN_INCLUDE_DIRS
+        ${CANN_INSTALL_DIR}/include
+        ${CANN_INSTALL_DIR}/include/aclnn
+        ${CANN_INSTALL_DIR}/acllib/include
+    )
+
+    add_subdirectory(kernels)
+    list(APPEND CANN_LIBRARIES
+        ascendcl
+        nnopbase
+        opapi
+        acl_op_compiler
+        ascendc_kernels
+    )
+
+    file(GLOB GGML_SOURCES_CANN "*.cpp")
+
+    add_library(ggml-cann ${GGML_SOURCES_CANN})
+    target_link_libraries(ggml-cann PRIVATE ggml-base ${CANN_LIBRARIES})
+    target_include_directories(ggml-cann PRIVATE . .. ${CANN_INCLUDE_DIRS})
+    target_link_directories(ggml-cann PRIVATE ${CANN_INSTALL_DIR}/lib64)
+
+    message(STATUS "CANN: CANN_INCLUDE_DIRS =  ${CANN_INCLUDE_DIRS}")
+    message(STATUS "CANN: CANN_LIBRARIES =  ${CANN_LIBRARIES}")
+else()
+    message(FATAL_ERROR "CANN: Can't find CANN_INSTALL_DIR, did you forget to source set_var.sh?")
+endif()

ggml/src/ggml-cann/ggml-cann.cpp

@@ -0,0 +1,2128 @@
+/*
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include "ggml-cann.h"
+
+#include <acl/acl.h>
+#include <stdarg.h>
+
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <mutex>
+
+#include "ggml-impl.h"
+#include "ggml-backend-impl.h"
+#include "ggml-cann/aclnn_ops.h"
+#include "ggml-cann/common.h"
+
+#define GGML_COMMON_DECL_C
+
+#include "ggml-common.h"
+
+#define GGML_CANN_NAME "CANN"
+
+/**
+ * @brief Handles CANN errors by printing an error message and aborting.
+ *
+ * @param stmt The statement that caused the error.
+ * @param func The function in which the error occurred.
+ * @param file The file in which the error occurred.
+ * @param line The line number where the error occurred.
+ * @param msg The error message.
+ */
+[[noreturn]] void ggml_cann_error(const char* stmt, const char* func,
+                                  const char* file, int line, const char* msg) {
+    int32_t id = -1;
+    aclrtGetDevice(&id);
+
+    GGML_LOG_ERROR("CANN error: %s\n", msg);
+    GGML_LOG_ERROR("  current device: %d, in function %s at %s:%d\n", id, func,
+            file, line);
+    GGML_LOG_ERROR("  %s\n", stmt);
+    // abort with GGML_ASSERT to get a stack trace
+    GGML_ABORT("CANN error");
+}
+
+/**
+ * @brief Sets the device to be used by CANN.
+ *
+ * @param device The device ID to set.
+ */
+void ggml_cann_set_device(const int32_t device) {
+    // TODO: uncomment these lines after empty context has fixed.
+    // int current_device;
+    // ACL_CHECK(aclrtGetDevice(&current_device));
+
+    // if (device == current_device) {
+    //   return;
+    // }
+    ACL_CHECK(aclrtSetDevice(device));
+}
+
+/**
+ * @brief Retrieves the current device ID.
+ *
+ * @return The current device ID.
+ */
+int32_t ggml_cann_get_device() {
+    int32_t id;
+    ACL_CHECK(aclrtGetDevice(&id));
+    return id;
+}
+
+/**
+ * @brief Initialize the CANN device information.
+ *
+ * This function initializes the CANN device information by obtaining the
+ * device count and setting the memory allocation granularity for each device.
+ *
+ * @return A structure containing the device information.
+ */
+static ggml_cann_device_info ggml_cann_init() {
+    ggml_cann_device_info info = {};
+
+    aclError err = aclrtGetDeviceCount((uint32_t*)&info.device_count);
+
+    if (err != ACL_SUCCESS) {
+        GGML_LOG_ERROR("%s: failed to initialize CANN: %s\n",
+                __func__, aclGetRecentErrMsg());
+        return info;
+    }
+
+    GGML_ASSERT(info.device_count <= GGML_CANN_MAX_DEVICES);
+
+    for (int id = 0; id < info.device_count; ++id) {
+        aclrtPhysicalMemProp prop = {};
+        prop.handleType = ACL_MEM_HANDLE_TYPE_NONE;
+        prop.allocationType = ACL_MEM_ALLOCATION_TYPE_PINNED;
+        prop.memAttr = ACL_HBM_MEM_HUGE;
+        prop.location.type = ACL_MEM_LOCATION_TYPE_DEVICE;
+        prop.location.id = id;
+        prop.reserve = 0;
+        ACL_CHECK(aclrtMemGetAllocationGranularity(
+            &prop, ACL_RT_MEM_ALLOC_GRANULARITY_RECOMMENDED,
+            &info.devices[id].vmm_granularity));
+    }
+
+    // TODO: add more device info later.
+    return info;
+}
+
+/**
+ * @brief Retrieve the CANN device information.
+ *
+ * This function returns a reference to a structure containing the CANN device
+ * information. The device information is initialized once and reused on
+ * subsequent calls.
+ *
+ * @return A reference to the structure containing the device information.
+ */
+const ggml_cann_device_info& ggml_cann_info() {
+    static ggml_cann_device_info info = ggml_cann_init();
+    return info;
+}
+
+//#define DEBUG_CANN_MALLOC
+/**
+ * @brief A pool of CANN buffers(legacy).
+ *
+ * This class manages a pool of CANN buffers for a specific device.
+ */
+struct ggml_cann_pool_leg : public ggml_cann_pool {
+    /**
+     * @brief The maximum number of buffers in the pool.
+     */
+    static const int MAX_BUFFERS = 256;
+
+    /**
+     * @brief The device ID associated with this buffer pool.
+     */
+    int device;
+
+    /**
+     * @brief Structure representing a CANN buffer.
+     */
+    struct ggml_cann_buffer {
+        void* ptr = nullptr;  ///< Pointer to the buffer memory.
+        size_t size = 0;      ///< Size of the buffer.
+    };
+
+    /**
+     * @brief Array of CANN buffers in the pool.
+     */
+    ggml_cann_buffer buffer_pool[MAX_BUFFERS] = {};
+
+    /**
+     * @brief Total size of all buffers in the pool.
+     */
+    size_t pool_size = 0;
+
+    /**
+     * @brief Constructor to initialize the buffer pool for a specific device.
+     *
+     * @param device The device ID to associate with this buffer pool.
+     */
+    explicit ggml_cann_pool_leg(int device) : device(device) {}
+
+    /**
+     * @brief Destructor to free all buffers in the pool.
+     */
+    ~ggml_cann_pool_leg() {
+        ggml_cann_set_device(device);
+        for (int i = 0; i < MAX_BUFFERS; ++i) {
+            ggml_cann_buffer& b = buffer_pool[i];
+            if (b.ptr != nullptr) {
+                ACL_CHECK(aclrtFree(b.ptr));
+                pool_size -= b.size;
+            }
+        }
+        GGML_ASSERT(pool_size == 0);
+    }
+
+    /**
+     * @brief Allocate a buffer of the given size.
+     *
+     * @param size The size of the buffer to allocate.
+     * @param actual_size A pointer to a variable to receive the actual size of
+     * the allocated buffer.
+     * @return A pointer to the allocated buffer.
+     */
+    void* alloc(size_t size, size_t* actual_size) override {
+#ifdef DEBUG_CANN_MALLOC
+        int nnz = 0;
+        size_t max_size = 0;
+#endif
+        size_t best_diff = 1ull << 36;
+        int ibest = -1;
+        for (int i = 0; i < MAX_BUFFERS; ++i) {
+            ggml_cann_buffer& b = buffer_pool[i];
+            if (b.ptr != nullptr) {
+#ifdef DEBUG_CANN_MALLOC
+                ++nnz;
+                if (b.size > max_size) max_size = b.size;
+#endif
+                if (b.size >= size) {
+                    size_t diff = b.size - size;
+                    if (diff < best_diff) {
+                        best_diff = diff;
+                        ibest = i;
+                        if (!best_diff) {
+                            void* ptr = b.ptr;
+                            *actual_size = b.size;
+                            b.ptr = nullptr;
+                            b.size = 0;
+                            return ptr;
+                        }
+                    }
+                }
+            }
+        }
+        if (ibest >= 0) {
+            ggml_cann_buffer& b = buffer_pool[ibest];
+            void* ptr = b.ptr;
+            *actual_size = b.size;
+            b.ptr = nullptr;
+            b.size = 0;
+            return ptr;
+        }
+        void* ptr;
+        size_t look_ahead_size = (size_t)(1.05 * size);
+        look_ahead_size = 256 * ((look_ahead_size + 255) / 256);
+        ggml_cann_set_device(device);
+        ACL_CHECK(
+            aclrtMalloc(&ptr, look_ahead_size, ACL_MEM_MALLOC_HUGE_FIRST));
+        *actual_size = look_ahead_size;
+        pool_size += look_ahead_size;
+#ifdef DEBUG_CANN_MALLOC
+        GGML_LOG_INFO(
+            "%s[%d]: %d buffers, max_size = %u MB, pool_size = %u MB, "
+            "requested %u MB\n",
+            __func__, device, nnz, (uint32_t)(max_size / 1024 / 1024),
+            (uint32_t)(pool_size / 1024 / 1024),
+            (uint32_t)(size / 1024 / 1024));
+#endif
+        return ptr;
+    }
+
+    /**
+     * @brief Free a buffer and return it to the pool.
+     *
+     * @param ptr Pointer to the buffer to free.
+     * @param size Size of the buffer to free.
+     */
+    void free(void* ptr, size_t size) override {
+        for (int i = 0; i < MAX_BUFFERS; ++i) {
+            ggml_cann_buffer& b = buffer_pool[i];
+            if (b.ptr == nullptr) {
+                b.ptr = ptr;
+                b.size = size;
+                return;
+            }
+        }
+        // memory should always buffered. these memory may still needed by
+        // tasks in stream.
+        // TODO, fix me.
+        GGML_ABORT("Cann buffer pool full, increase MAX_CANN_BUFFERS\n");
+    }
+};
+
+/**
+ * @brief A pool of CANN buffers with virtual memory.
+ *
+ * This class manages a pool of CANN buffers with virtual memory for a specific
+ * device.
+ */
+struct ggml_cann_pool_vmm : public ggml_cann_pool {
+    /**
+     * @brief The maximum size of the virtual memory pool (32 GB).
+     */
+    static const size_t CANN_POOL_VMM_MAX_SIZE = 1ull << 35;  // 32 GB
+
+    /**
+     * @brief The device ID associated with this buffer pool.
+     */
+    int device;
+
+    /**
+     * @brief Pointer to the start of the virtual memory pool.
+     */
+    void* pool_addr = 0;
+
+    /**
+     * @brief Amount of virtual memory used in the pool.
+     */
+    size_t pool_used = 0;
+
+    /**
+     * @brief Total size of the virtual memory pool.
+     */
+    size_t pool_size = 0;
+
+    /**
+     * @brief Allocation granularity for the virtual memory pool.
+     */
+    size_t granularity;
+
+    /**
+     * @brief Handles for the physical memory allocated.
+     */
+    std::vector<aclrtDrvMemHandle> handles;
+
+    /**
+     * @brief Offsets for the mapped memory regions.
+     */
+    std::vector<void*> map_offsets;
+
+    /**
+     * @brief Constructor to initialize the buffer pool with virtual memory for
+     * a specific device.
+     *
+     * @param device The device ID to associate with this buffer pool.
+     */
+    explicit ggml_cann_pool_vmm(int device)
+        : device(device),
+          granularity(ggml_cann_info().devices[device].vmm_granularity) {}
+
+    /**
+     * @brief Destructor to free all buffers in the virtual memory pool.
+     */
+    ~ggml_cann_pool_vmm() {
+        if (pool_addr != 0) {
+            for (auto& offset : map_offsets) {
+                ACL_CHECK(aclrtUnmapMem(offset));
+            }
+            for (auto& handle : handles) {
+                ACL_CHECK(aclrtFreePhysical(handle));
+            }
+            ACL_CHECK(aclrtReleaseMemAddress(pool_addr));
+        }
+    }
+
+    /**
+     * @brief Allocate a buffer of the given size in the virtual memory pool.
+     *
+     * @param size The size of the buffer to allocate.
+     * @param actual_size A pointer to a variable to receive the actual size of
+     * the allocated buffer.
+     * @return A pointer to the allocated buffer.
+     */
+    void* alloc(size_t size, size_t* actual_size) override {
+        // round up the allocation size to the alignment to ensure that all
+        // allocations are aligned for all data types
+        const size_t alignment = 128;
+        size = alignment * ((size + alignment - 1) / alignment);
+
+        size_t avail = pool_size - pool_used;
+
+        if (size > avail) {
+            // round up to the next multiple of the granularity
+            size_t reserve_size = size - avail;
+            reserve_size =
+                granularity * ((reserve_size + granularity - 1) / granularity);
+
+            GGML_ASSERT(pool_size + reserve_size <= CANN_POOL_VMM_MAX_SIZE);
+
+            // allocate more physical memory
+            aclrtPhysicalMemProp prop = {};
+            prop.handleType = ACL_MEM_HANDLE_TYPE_NONE;
+            prop.allocationType = ACL_MEM_ALLOCATION_TYPE_PINNED;
+            prop.memAttr = ACL_HBM_MEM_HUGE;
+            prop.location.type = ACL_MEM_LOCATION_TYPE_DEVICE;
+            prop.location.id = device;
+            prop.reserve = 0;
+            aclrtDrvMemHandle handle;
+            ACL_CHECK(aclrtMallocPhysical(&handle, reserve_size, &prop, 0));
+
+            // reserve virtual address space (if not already reserved)
+            if (pool_addr == 0) {
+                ACL_CHECK(aclrtReserveMemAddress(
+                    &pool_addr, CANN_POOL_VMM_MAX_SIZE, 0, NULL, 1));
+            }
+
+            // map at the end of the pool
+            ACL_CHECK(aclrtMapMem((char*)pool_addr + pool_size, reserve_size, 0,
+                                  handle, 0));
+
+            handles.push_back(handle);
+            map_offsets.push_back((char*)pool_addr + pool_size);
+
+            // add to the pool
+            pool_size += reserve_size;
+
+            // GGML_LOG_INFO("cann pool[%d]: size increased to %llu MB (
+            // reserved %llu MB)\n",
+            //       device, (unsigned long long) (pool_size/1024/1024),
+            //       (unsigned long long) (reserve_size/1024/1024));
+        }
+
+        GGML_ASSERT(pool_addr != 0);
+
+        void* ptr = (void*)((char*)pool_addr + pool_used);
+        *actual_size = size;
+        pool_used += size;
+
+#ifdef DEBUG_CANN_MALLOC
+        GGML_LOG_INFO("cann pool[%d]: allocated %llu bytes at %llx\n", device,
+               (unsigned long long)size, (unsigned long long)ptr);
+#endif
+        return ptr;
+    }
+
+    /**
+     * @brief Free a buffer and return it to the virtual memory pool.
+     *
+     * @param ptr Pointer to the buffer to free.
+     * @param size Size of the buffer to free.
+     */
+    void free(void* ptr, size_t size) override {
+#ifdef DEBUG_CANN_MALLOC
+        GGML_LOG_INFO("cann pool[%d]: freed %llu bytes at %llx\n", device,
+               (unsigned long long)size, (unsigned long long)ptr);
+#endif
+
+        pool_used -= size;
+
+        // all deallocations must be in reverse order of the allocations
+        GGML_ASSERT(ptr == (void*)((char*)pool_addr + pool_used));
+    }
+};
+
+/**
+ * @brief Create a new CANN pool for a specific device.
+ *
+ * Factory method to create a new CANN pool object based on the device type.
+ *
+ * @param device The device ID for which to create the pool.
+ * @return A unique pointer to the created CANN pool.
+ */
+std::unique_ptr<ggml_cann_pool> ggml_backend_cann_context::new_pool_for_device(
+    int device) {
+    // return std::unique_ptr<ggml_cann_pool>(new ggml_cann_pool_leg(device));
+    return std::unique_ptr<ggml_cann_pool>(new ggml_cann_pool_vmm(device));
+}
+
+// cann buffer
+/**
+ * @brief Context for managing a CANN buffer associated with a specific device.
+ *
+ * This structure holds information about a CANN buffer, including the device
+ * ID, device pointer, and a name derived from GGML_CANN_NAME and the device ID.
+ */
+struct ggml_backend_cann_buffer_context {
+    int32_t device;  ///< The device ID associated with this buffer context.
+    void* dev_ptr =
+        nullptr;  ///< Pointer to the device memory allocated for the buffer.
+
+    /**
+     * @brief Constructor to initialize the CANN buffer context.
+     *
+     * @param device The device ID associated with this buffer context.
+     * @param dev_ptr Pointer to the device memory allocated for the buffer.
+     */
+    ggml_backend_cann_buffer_context(int32_t device, void* dev_ptr)
+        : device(device),
+          dev_ptr(dev_ptr) {}
+
+    /**
+     * @brief Destructor to free the device memory allocated for the buffer.
+     */
+    ~ggml_backend_cann_buffer_context() { ACL_CHECK(aclrtFree(dev_ptr)); }
+};
+
+/**
+ * @brief Check if a buffer is a CANN buffer.
+ *
+ * This function checks if a given buffer is a CANN buffer by comparing its
+ * `get_name` function pointer to `ggml_backend_cann_buffer_get_name`.
+ *
+ * @param buffer The buffer to check.
+ * @return true if the buffer is a CANN buffer, false otherwise.
+ */
+static bool ggml_backend_buft_is_cann(ggml_backend_buffer_type_t buft);
+static bool ggml_backend_buffer_is_cann(
+    ggml_backend_buffer_t buffer) {
+    return ggml_backend_buft_is_cann(buffer->buft);
+}
+
+/**
+ * @brief Free resources associated with a CANN buffer.
+ *
+ * This function frees the resources associated with a CANN buffer, including
+ * its context.
+ *
+ * @param buffer The CANN buffer to free.
+ */
+static void ggml_backend_cann_buffer_free_buffer(
+    ggml_backend_buffer_t buffer) {
+    ggml_backend_cann_buffer_context* ctx =
+        (ggml_backend_cann_buffer_context*)buffer->context;
+    delete ctx;
+}
+
+/**
+ * @brief Retrieve the base pointer of a CANN buffer.
+ *
+ * This function returns the base pointer of a CANN buffer, which points to the
+ * device memory allocated for the buffer.
+ *
+ * @param buffer The CANN buffer whose base pointer is to be retrieved.
+ * @return A pointer to the base of the device memory allocated for the buffer.
+ */
+static void* ggml_backend_cann_buffer_get_base(
+    ggml_backend_buffer_t buffer) {
+    ggml_backend_cann_buffer_context* ctx =
+        (ggml_backend_cann_buffer_context*)buffer->context;
+    return ctx->dev_ptr;
+}
+
+/**
+ * @brief Transform quantized Q4.0 tensor data into a format suitable for CANN
+ * processing.
+ *
+ * This function transforms quantized Q4.0 tensor data into a format suitable
+ * for CANN processing. It extracts quantization values and scales from the
+ * source data and prepares them in a format expected by CANN operations.
+ *
+ * @param tensor Pointer to the tensor information.
+ * @param src Pointer to the source data in Q4.0 format.
+ * @param dst Pointer to the destination buffer where transformed data will be
+ * stored.
+ */
+static void ggml_backend_cann_transform_q4_0(ggml_tensor* tensor,
+                                             const void* src,
+                                             void* dst) {
+
+    int64_t n_elems = ggml_nelements(tensor);
+    int64_t groups = n_elems / QK4_0;
+    size_t quant_bytes = n_elems * sizeof(uint8_t) / 2;
+
+    uint8_t* quant_offset = (uint8_t*)dst;
+    uint16_t* scale_offset = (uint16_t*)((char*)dst + quant_bytes);
+
+    for (int i = 0; i < groups; i++) {
+        const block_q4_0* group =
+            (const block_q4_0*)((const char*)src + i * sizeof(block_q4_0));
+        *scale_offset = group->d;
+        scale_offset++;
+
+        // 0-15
+        for (int j = 0; j < QK4_0 / 2; j += 2) {
+            (*quant_offset) = (group->qs[j] & 0x0F);
+            (*quant_offset) |= ((group->qs[j + 1] << 4));
+            quant_offset++;
+        }
+
+        // 16-31
+        for (int j = 0; j < QK4_0 / 2; j += 2) {
+            (*quant_offset) = (group->qs[j] >> 4);
+            (*quant_offset) |= (group->qs[j + 1] & 0xF0);
+            quant_offset++;
+        }
+    }
+
+    // put (uint4b_t -8) into int4b_t
+    for (quant_offset = (uint8_t*)dst;
+         quant_offset < (uint8_t*)dst + quant_bytes; quant_offset++) {
+        (*quant_offset) ^= 0x88;
+    }
+}
+
+/**
+ * @brief Transform CANN processed data back into quantized Q4.0 format.
+ *
+ * This function transforms CANN processed data back into quantized Q4.0 format.
+ * It reverses the transformation performed by
+ * ggml_backend_cann_transform_q4_0(), converting the data back into its
+ * original quantized form.
+ *
+ * @param tensor Pointer to the tensor information.
+ * @param src Pointer to the source buffer containing transformed data.
+ * @param dst Pointer to the destination buffer where the Q4.0 formatted data
+ * will be stored.
+ */
+static void ggml_backend_cann_transform_back_q4_0(
+    const ggml_tensor* tensor, void* src, void* dst) {
+
+    int64_t n_elems = ggml_nelements(tensor);
+    int64_t groups = n_elems / QK4_0;
+    size_t quant_bytes = n_elems * sizeof(uint8_t) / 2;
+
+    uint8_t* quant_offset = (uint8_t*)src;
+    uint16_t* scale_offset = (uint16_t*)((char*)src + quant_bytes);
+
+    for (; quant_offset < (uint8_t*)src + quant_bytes; quant_offset++) {
+        (*quant_offset) ^= 0x88;
+    }
+    quant_offset = (uint8_t*)src;
+
+    for (int i = 0; i < groups; i++) {
+        block_q4_0* group = (block_q4_0*)((char*)dst + i * sizeof(block_q4_0));
+        group->d = *scale_offset;
+        scale_offset++;
+
+        // 0-15
+        for (int j = 0; j < QK4_0 / 2; j += 2) {
+            group->qs[j] = ((*quant_offset) & 0x0F);
+            group->qs[j + 1] = ((*quant_offset) >> 4);
+            quant_offset++;
+        }
+
+        // 16-31
+        for (int j = 0; j < QK4_0 / 2; j += 2) {
+            group->qs[j] |= ((*quant_offset) << 4);
+            group->qs[j + 1] |= ((*quant_offset) & 0xF0);
+            quant_offset++;
+        }
+    }
+}
+
+/**
+ * @brief Transform quantized Q8.0 tensor data into a format suitable for CANN
+ * processing.
+ *
+ * This function transforms quantized Q8.0 tensor data into a format suitable
+ * for CANN processing. It extracts quantization values and scales from the
+ * source data and prepares them in a format expected by CANN operations.
+ *
+ * @param tensor Pointer to the tensor information.
+ * @param src Pointer to the source data in Q8.0 format.
+ * @param dst Pointer to the destination buffer where transformed data will be
+ * stored.
+ */
+static void ggml_backend_cann_transform_q8_0(ggml_tensor* tensor,
+                                             const void* src,
+                                             void* dst) {
+    int64_t n_elems = ggml_nelements(tensor);
+    int64_t groups = n_elems / QK8_0;
+    size_t quant_bytes = n_elems * sizeof(uint8_t);
+
+    uint8_t* quant_offset = (uint8_t*)dst;
+    uint16_t* scale_offset = (uint16_t*)((char*)dst + quant_bytes);
+
+    for (int i = 0; i < groups; i++) {
+        const block_q8_0* group =
+            (const block_q8_0*)((const char*)src + i * sizeof(block_q8_0));
+        *scale_offset = group->d;
+        scale_offset++;
+        size_t group_quant_size = QK8_0 * sizeof(uint8_t);
+        memcpy(quant_offset, group->qs, group_quant_size);
+        quant_offset += group_quant_size;
+    }
+}
+
+/**
+ * @brief Transform CANN processed data back into quantized Q8.0 format.
+ *
+ * This function transforms CANN processed data back into quantized Q8.0 format.
+ * It reverses the transformation performed by
+ * ggml_backend_cann_transform_q8_0(), converting the data back into its
+ * original quantized form.
+ *
+ * @param tensor Pointer to the tensor information.
+ * @param src Pointer to the source buffer containing transformed data.
+ * @param dst Pointer to the destination buffer where the Q8.0 formatted data
+ * will be stored.
+ */
+static void ggml_backend_cann_transform_back_q8_0(
+    const ggml_tensor* tensor, const void* src, void* dst) {
+    int64_t n_elems = ggml_nelements(tensor);
+    int64_t groups = n_elems / QK8_0;
+    size_t quant_bytes = n_elems * sizeof(uint8_t);
+
+    const uint8_t* quant_offset = (const uint8_t*)src;
+    const uint16_t* scale_offset =
+        (const uint16_t*)((const char*)src + quant_bytes);
+
+    for (int i = 0; i < groups; i++) {
+        block_q8_0* group = (block_q8_0*)((char*)dst + i * sizeof(block_q8_0));
+        group->d = *scale_offset;
+        scale_offset++;
+        size_t group_quant_size = QK8_0 * sizeof(uint8_t);
+        memcpy(group->qs, quant_offset, group_quant_size);
+        quant_offset += group_quant_size;
+    }
+}
+
+/**
+ * @brief Transform tensor data based on its type for CANN processing.
+ *
+ * This function transforms tensor data based on its quantization type for CANN
+ * processing. It dispatches the transformation based on the tensor's type to
+ * specialized functions handling Q4.0 and Q8.0 formats.
+ *
+ * @param tensor Pointer to the tensor information.
+ * @param src Pointer to the source data to be transformed.
+ * @param dst Pointer to the destination buffer where transformed data will be
+ * stored.
+ */
+static void ggml_backend_cann_transform(ggml_tensor* tensor,
+                                        const void* src, void* dst) {
+    switch (tensor->type) {
+        case GGML_TYPE_Q4_0:
+            ggml_backend_cann_transform_q4_0(tensor, src, dst);
+            break;
+        case GGML_TYPE_Q8_0:
+            ggml_backend_cann_transform_q8_0(tensor, src, dst);
+            break;
+        default:
+            break;
+    }
+}
+
+/**
+ * @brief Transform CANN processed data back into tensor data based on its type.
+ *
+ * This function transforms CANN processed data back into tensor data based on
+ * its quantization type for Q4.0 and Q8.0 formats. It dispatches the
+ * transformation based on the tensor's type to specialized functions.
+ *
+ * @param tensor Pointer to the tensor information.
+ * @param src Pointer to the source data containing CANN processed data.
+ * @param dst Pointer to the destination buffer where transformed tensor data
+ * will be stored.
+ */
+static void ggml_backend_cann_transform_back(
+    const ggml_tensor* tensor, void* src, void* dst) {
+    switch (tensor->type) {
+        case GGML_TYPE_Q4_0:
+            ggml_backend_cann_transform_back_q4_0(tensor, src, dst);
+            break;
+        case GGML_TYPE_Q8_0:
+            ggml_backend_cann_transform_back_q8_0(tensor, src, dst);
+            break;
+        default:
+            break;
+    }
+}
+
+/**
+ * @brief Check if transformation is needed for a given tensor type.
+ *
+ * This function checks if transformation is needed for a given tensor type
+ * to prepare data for CANN processing.
+ *
+ * @param type The tensor type to check.
+ * @return true if transformation is needed, false otherwise.
+ */
+static bool need_transform(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q8_0:
+            return true;
+        default:
+            return false;
+    }
+}
+
+/**
+ * @brief Initialize a tensor using data from a CANN buffer.
+ *
+ * This function initializes a tensor using data from a CANN buffer.
+ * It handles special cases such as views and quantization.
+ *
+ * @param buffer The CANN buffer from which to initialize the tensor.
+ * @param tensor Pointer to the tensor to be initialized.
+ */
+static void ggml_backend_cann_buffer_init_tensor(
+    ggml_backend_buffer_t buffer, ggml_tensor* tensor) {
+    if (tensor->view_src != NULL && tensor->view_offs == 0) {
+        GGML_ASSERT(tensor->view_src->buffer->buft == buffer->buft);
+        return;
+    }
+
+    // TODO: can backend doesn't support quantized yet. Just leave the code
+    // here.
+    if (ggml_is_quantized(tensor->type)) {
+        // Initialize padding to 0 to avoid possible NaN values
+        size_t original_size = ggml_nbytes(tensor);
+        size_t padded_size =
+            ggml_backend_buft_get_alloc_size(buffer->buft, tensor);
+
+        if (padded_size > original_size && tensor->view_src == nullptr) {
+            size_t memset_size = padded_size - original_size;
+            ACL_CHECK(aclrtMemset((char*)tensor->data + original_size,
+                                  memset_size, 0, memset_size));
+        }
+    }
+}
+
+// TODO: need handle tensor which has paddings.
+/**
+ * @brief Set tensor data in a CANN buffer.
+ *
+ * This function sets tensor data in a CANN buffer, handling transformations
+ * if needed based on the tensor's type.
+ *
+ * @param buffer The CANN buffer where the tensor data will be set.
+ * @param tensor Pointer to the tensor whose data will be set.
+ * @param data Pointer to the source data to be copied into the tensor.
+ * @param offset Offset in the source data from where to start copying.
+ * @param size Size of the data to be copied, in bytes.
+ */
+static void ggml_backend_cann_buffer_set_tensor(
+    ggml_backend_buffer_t buffer, ggml_tensor *tensor, const void *data,
+    size_t offset, size_t size) {
+    ggml_backend_cann_buffer_context *ctx =
+        (ggml_backend_cann_buffer_context *)buffer->context;
+
+    ggml_cann_set_device(ctx->device);
+    // TODO: refer to cann(#6017), it use thread's default stream.
+    // For acl, synchronous functions use this default stream.
+    // Why aclrtSynchronizeDevice?
+
+    if (!need_transform(tensor->type)) {
+        ACL_CHECK(aclrtMemcpy((char *)tensor->data + offset, size, data, size,
+                              ACL_MEMCPY_HOST_TO_DEVICE));
+    } else {
+        void *transform_buffer = malloc(size);
+        ggml_backend_cann_transform(tensor, data, transform_buffer);
+
+        ACL_CHECK(aclrtMemcpy((char *)tensor->data + offset, size,
+                              transform_buffer, size,
+                              ACL_MEMCPY_HOST_TO_DEVICE));
+        free(transform_buffer);
+    }
+}
+
+/**
+ * @brief Get tensor data from a CANN buffer.
+ *
+ * This function retrieves tensor data from a CANN buffer, handling
+ * transformations if needed based on the tensor's type.
+ *
+ * @param buffer The CANN buffer from which to retrieve tensor data.
+ * @param tensor Pointer to the tensor whose data will be retrieved.
+ * @param data Pointer to the destination buffer where the tensor data will be
+ * copied.
+ * @param offset Offset in the destination buffer where to start copying.
+ * @param size Size of the data to be copied, in bytes.
+ */
+static void ggml_backend_cann_buffer_get_tensor(
+    ggml_backend_buffer_t buffer, const ggml_tensor* tensor, void* data,
+    size_t offset, size_t size) {
+    ggml_backend_cann_buffer_context* ctx =
+        (ggml_backend_cann_buffer_context*)buffer->context;
+
+    ggml_cann_set_device(ctx->device);
+
+    if (!need_transform(tensor->type)) {
+        ACL_CHECK(aclrtMemcpy(data, size, (char*)tensor->data + offset, size,
+                              ACL_MEMCPY_DEVICE_TO_HOST));
+    } else {
+        void* transform_buffer = malloc(size);
+        ACL_CHECK(aclrtMemcpy(transform_buffer, size,
+                              (char*)tensor->data + offset, size,
+                              ACL_MEMCPY_DEVICE_TO_HOST));
+        ggml_backend_cann_transform_back(tensor, transform_buffer, data);
+        free(transform_buffer);
+    }
+}
+
+/**
+ * @brief Copy tensor data between CANN buffers if possible.
+ *
+ * This function copies tensor data between CANN buffers if the source and
+ * destination buffers are CANN buffers and they meet the necessary conditions
+ * (same device or devices can access each other).
+ *
+ * @param buffer The destination CANN buffer where the tensor data will be
+ * copied.
+ * @param src Pointer to the source tensor whose data will be copied.
+ * @param dst Pointer to the destination tensor where the data will be copied.
+ * @return true if the copy operation succeeded, false otherwise.
+ */
+static bool ggml_backend_cann_buffer_cpy_tensor(
+    ggml_backend_buffer_t buffer, const ggml_tensor* src, ggml_tensor* dst) {
+    if (ggml_backend_buffer_is_cann(src->buffer)) {
+        ggml_backend_cann_buffer_context* src_ctx =
+            (ggml_backend_cann_buffer_context*)src->buffer->context;
+        ggml_backend_cann_buffer_context* dst_ctx =
+            (ggml_backend_cann_buffer_context*)buffer->context;
+
+        size_t memcpy_size = ggml_nbytes(src);
+        // Same device.
+        if (src_ctx->device == dst_ctx->device) {
+            ACL_CHECK(aclrtMemcpy((char*)dst->data, memcpy_size,
+                                  (const char*)src->data, memcpy_size,
+                                  ACL_MEMCPY_DEVICE_TO_DEVICE));
+            return true;
+        } else {
+            // Different device but can access by peer.
+            int32_t canAccessPeer = 0;
+            ACL_CHECK(aclrtDeviceCanAccessPeer(&canAccessPeer, src_ctx->device,
+                                               dst_ctx->device));
+            if (canAccessPeer) {
+                ggml_cann_set_device(src_ctx->device);
+                ACL_CHECK(aclrtDeviceEnablePeerAccess(dst_ctx->device, 0));
+                ACL_CHECK(aclrtMemcpy((char*)dst->data, memcpy_size,
+                                      (const char*)src->data, memcpy_size,
+                                      ACL_MEMCPY_DEVICE_TO_DEVICE));
+                return true;
+            }
+        }
+    }
+    return false;
+}
+
+/**
+ * @brief Clear a CANN buffer by setting all its memory to a specified value.
+ *
+ * This function clears a CANN buffer by setting all its memory to a specified
+ * value.
+ *
+ * @param buffer The CANN buffer to be cleared.
+ * @param value The value to which each byte in the buffer will be set.
+ */
+static void ggml_backend_cann_buffer_clear(
+    ggml_backend_buffer_t buffer, uint8_t value) {
+    ggml_backend_cann_buffer_context* ctx =
+        (ggml_backend_cann_buffer_context*)buffer->context;
+
+    ggml_cann_set_device(ctx->device);
+    ACL_CHECK(aclrtMemset(ctx->dev_ptr, buffer->size, value, buffer->size));
+}
+
+/**
+ * @brief Interface for a CANN buffer in the backend.
+ *
+ * This structure defines function pointers to operations that can be performed
+ * on a CANN buffer within the backend.
+ */
+static const ggml_backend_buffer_i ggml_backend_cann_buffer_interface = {
+    /* .free_buffer     = */ ggml_backend_cann_buffer_free_buffer,
+    /* .get_base        = */ ggml_backend_cann_buffer_get_base,
+    /* .init_tensor     = */ ggml_backend_cann_buffer_init_tensor,
+    /* .memset_tensor   = */ NULL,
+    /* .set_tensor      = */ ggml_backend_cann_buffer_set_tensor,
+    /* .get_tensor      = */ ggml_backend_cann_buffer_get_tensor,
+    /* .cpy_tensor      = */ ggml_backend_cann_buffer_cpy_tensor,
+    /* .clear           = */ ggml_backend_cann_buffer_clear,
+    /* .reset           = */ NULL,
+};
+
+// cann buffer type
+/**
+ * @brief Structure representing context information for a specific backend
+ * buffer type.
+ */
+struct ggml_backend_cann_buffer_type_context {
+    int32_t
+        device; /**< Device identifier associated with the buffer context. */
+    std::string name; /**< Name associated with the buffer context. */
+};
+
+/**
+ * @brief Retrieves the name associated with a CANN buffer type.
+ *
+ * This function returns the descriptive name associated with the specified
+ * CANN buffer type context.
+ *
+ * @param buft Pointer to the buffer type context.
+ * @return Const pointer to the C-style string containing the name.
+ */
+static const char* ggml_backend_cann_buffer_type_name(
+    ggml_backend_buffer_type_t buft) {
+    ggml_backend_cann_buffer_type_context* buft_ctx =
+        (ggml_backend_cann_buffer_type_context*)buft->context;
+
+    return buft_ctx->name.c_str();
+}
+
+/**
+ * @brief Allocates a new CANN buffer of the specified type and size.
+ *
+ * This function allocates a new CANN buffer on the specified device with the
+ * given size.
+ *
+ * @param buft Pointer to the buffer type context.
+ * @param size Size in bytes of the buffer to allocate.
+ * @return Pointer to the allocated buffer, or nullptr if allocation fails.
+ */
+static ggml_backend_buffer_t
+ggml_backend_cann_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft,
+                                           size_t size) {
+    ggml_backend_cann_buffer_type_context* buft_ctx =
+        (ggml_backend_cann_buffer_type_context*)buft->context;
+
+    ggml_cann_set_device(buft_ctx->device);
+
+    size = std::max(size, (size_t)1);
+
+    void* dev_ptr;
+    aclError err = aclrtMalloc(&dev_ptr, size, ACL_MEM_MALLOC_HUGE_FIRST);
+    if (err != ACL_SUCCESS) {
+        GGML_LOG_ERROR(
+            "%s: allocating %.2f MiB on device %d: aclrtMalloc failed: %s\n",
+            __func__, size / 1024.0 / 1024.0, buft_ctx->device,
+            aclGetRecentErrMsg());
+        return nullptr;
+    }
+
+    ggml_backend_cann_buffer_context* ctx =
+        new ggml_backend_cann_buffer_context(buft_ctx->device, dev_ptr);
+
+    return ggml_backend_buffer_init(buft, ggml_backend_cann_buffer_interface,
+                                    ctx, size);
+}
+
+/**
+ * @brief Retrieves the memory alignment requirement for CANN buffers of this
+ * type.
+ *
+ * This function returns the alignment requirement in bytes for memory allocated
+ * by the CANN buffer type.
+ *
+ * @param buft Pointer to the buffer type context (unused in this
+ * implementation).
+ * @return The alignment requirement in bytes (fixed at 128 bytes for CANN
+ * buffers).
+ */
+static size_t ggml_backend_cann_buffer_type_get_alignment(
+    ggml_backend_buffer_type_t buft) {
+    return 128;
+
+    GGML_UNUSED(buft);
+}
+
+/**
+ * @brief Calculates the allocation size required for a tensor in a CANN buffer.
+ *
+ * Computes the total allocation size needed for storing the tensor's data in a
+ * CANN buffer, considering any necessary padding or adjustments for quantized
+ * types.
+ *
+ * @param buft Pointer to the buffer type context (unused in this
+ * implementation).
+ * @param tensor Pointer to the tensor for which the allocation size is
+ * calculated.
+ * @return The total allocation size in bytes required for the tensor in the
+ * CANN buffer.
+ */
+static size_t ggml_backend_cann_buffer_type_get_alloc_size(
+    ggml_backend_buffer_type_t buft, const ggml_tensor* tensor) {
+    size_t size = ggml_nbytes(tensor);
+    int64_t ne0 = tensor->ne[0];
+
+    // last line must bigger than 32, because every single op deal at
+    // least 32 bytes.
+    // TODO: quantized type?
+    // int64_t line_size = ne0 * ggml_element_size(tensor);
+    // int64_t line_size_align_32 = (line_size + 31) & ~31;
+    // size += (line_size_align_32 - line_size);
+
+    // TODO: not support quantized yet.
+    // TODO: consider un-continue tensor.
+    if (ggml_is_quantized(tensor->type)) {
+        if (ne0 % MATRIX_ROW_PADDING != 0) {
+            size += ggml_row_size(
+                tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING);
+        }
+    }
+
+    return size;
+
+    GGML_UNUSED(buft);
+}
+
+static bool ggml_backend_cann_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
+    return false;
+
+    GGML_UNUSED(buft);
+}
+
+/**
+ * @brief Interface for managing CANN buffer types in the GGML backend.
+ *
+ * Provides function pointers for allocating, querying properties, and managing
+ * memory for CANN buffer types in the GGML backend.
+ */
+static const ggml_backend_buffer_type_i ggml_backend_cann_buffer_type_interface = {
+    /* .get_name         = */ ggml_backend_cann_buffer_type_name,
+    /* .alloc_buffer     = */ ggml_backend_cann_buffer_type_alloc_buffer,
+    /* .get_alignment    = */ ggml_backend_cann_buffer_type_get_alignment,
+    /* .get_max_size     = */ NULL,  // defaults to SIZE_MAX
+    /* .get_alloc_size   = */ ggml_backend_cann_buffer_type_get_alloc_size,
+    /* .is_host          = */ ggml_backend_cann_buffer_type_is_host,
+};
+
+/**
+ * @brief Retrieves the CANN buffer type for a specified device.
+ *
+ * This function initializes and returns the buffer type interface associated
+ * with the given device. It ensures thread-safe access using a mutex.
+ *
+ * @param device The device index for which to retrieve the buffer type.
+ * @return A pointer to the buffer type interface for the specified device, or
+ * nullptr if the device index is out of range.
+ */
+ggml_backend_buffer_type_t
+ggml_backend_cann_buffer_type(int32_t device) {
+    static std::mutex mutex;
+    std::lock_guard<std::mutex> lock(mutex);
+
+    if (device >= ggml_backend_cann_get_device_count()) {
+        return nullptr;
+    }
+
+    static ggml_backend_buffer_type
+        ggml_backend_cann_buffer_types[GGML_CANN_MAX_DEVICES];
+
+    static bool ggml_backend_cann_buffer_type_initialized = false;
+
+    if (!ggml_backend_cann_buffer_type_initialized) {
+        for (int32_t i = 0; i < GGML_CANN_MAX_DEVICES; i++) {
+            ggml_backend_cann_buffer_types[i] = {
+                /* .iface    = */ ggml_backend_cann_buffer_type_interface,
+                /* .device    = */ ggml_backend_reg_dev_get(ggml_backend_cann_reg(), device),
+                /* .context  = */
+                 new ggml_backend_cann_buffer_type_context{
+                    i, "CANN" + std::to_string(i)},
+            };
+        }
+        ggml_backend_cann_buffer_type_initialized = true;
+    }
+
+    return &ggml_backend_cann_buffer_types[device];
+}
+
+/**
+ * @brief Retrieves the name associated with a CANN host buffer type.
+ *
+ * This function returns the descriptive name associated with the specified
+ * CANN host buffer type context.
+ *
+ * @param buft Pointer to the host buffer type context.
+ * @return Const pointer to the C-style string containing the name.
+ */
+static const char * ggml_backend_cann_host_buffer_type_name(ggml_backend_buffer_type_t buft) {
+    return "CANN_Host";
+
+    GGML_UNUSED(buft);
+}
+
+/**
+ * @brief Retrieves the name associated with a CANN host buffer.
+ *
+ * This function returns the descriptive name associated with the specified
+ * CANN host buffer context.
+ *
+ * @param buft Pointer to the host buffer context.
+ * @return Const pointer to the C-style string containing the name.
+ */
+static const char * ggml_backend_cann_host_buffer_name(ggml_backend_buffer_t buffer) {
+    return "CANN_Host";
+
+    GGML_UNUSED(buffer);
+}
+
+/**
+ * @brief Free resources associated with a CANN host buffer.
+ *
+ * This function frees the resources associated with a CANN host buffer, including
+ * its context.
+ *
+ * @param buffer The CANN host buffer to free.
+ */
+static void ggml_backend_cann_host_buffer_free(ggml_backend_buffer_t buffer) {
+    ACL_CHECK(aclrtFreeHost(buffer->context));
+}
+
+/**
+ * @brief Allocates a new CANN host buffer of the specified size.
+ *
+ * This function allocates a new CANN host buffer with the given size.
+ * @param size Size in bytes of the host buffer to allocate.
+ * @return Pointer to the allocated host buffer, or nullptr if allocation fails.
+ */
+static void * ggml_cann_host_malloc(size_t size) {
+    if (getenv("GGML_CANN_NO_PINNED") != nullptr) {
+        return nullptr;
+    }
+
+    void * hostPtr = nullptr;
+    aclError err = aclrtMallocHost((void **) &hostPtr, size);
+    if (err != ACL_SUCCESS) {
+
+        GGML_LOG_WARN("%s: failed to allocate %.2f MiB of pinned memory: %s\n", __func__,
+                           size / 1024.0 / 1024.0, aclGetRecentErrMsg());
+        return nullptr;
+    }
+    return hostPtr;
+}
+
+/**
+ * @brief Allocates a new CANN host buffer of the specified type and size.
+ *
+ * @param buft Pointer to the host buffer type context.
+ * @param size Size in bytes of the host buffer to allocate.
+ * @return Pointer to the allocated host buffer, or CPU buffer pointer if allocation fails.
+ */
+static ggml_backend_buffer_t ggml_backend_cann_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    void * hostPtr = ggml_cann_host_malloc(size);
+
+    if (hostPtr == nullptr) {
+        // fallback to cpu buffer
+        return ggml_backend_buft_alloc_buffer(ggml_backend_cpu_buffer_type(), size);
+    }
+
+    ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(hostPtr, size);
+    buffer->buft = buft;
+    buffer->iface.free_buffer = ggml_backend_cann_host_buffer_free;
+
+    return buffer;
+}
+
+/**
+ * @brief Interface for managing CANN host buffer types in the GGML backend.
+ *
+ * Provides function pointers for allocating, querying properties, and managing
+ * memory for CANN buffer types in the GGML backend.
+ */
+ggml_backend_buffer_type_t ggml_backend_cann_host_buffer_type() {
+    static struct ggml_backend_buffer_type ggml_backend_cann_buffer_type_host = {
+        /* .iface    = */ {
+            /* .get_name         = */ ggml_backend_cann_host_buffer_type_name,
+            /* .alloc_buffer     = */ ggml_backend_cann_host_buffer_type_alloc_buffer,
+            /* .get_alignment    = */ ggml_backend_cpu_buffer_type()->iface.get_alignment,
+            /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
+            /* .get_alloc_size   = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size,
+            /* .is_host          = */ ggml_backend_cpu_buffer_type()->iface.is_host,
+        },
+        /* .device   = */ ggml_backend_reg_dev_get(ggml_backend_cann_reg(), 0),
+        /* .context  = */ nullptr,
+    };
+
+    return &ggml_backend_cann_buffer_type_host;
+}
+
+/**
+ * @brief Computes the forward operation for a given tensor using CANN
+ * operations.
+ *
+ * This function selects the appropriate CANN operation based on the type of
+ * operation specified in the tensor and performs the computation.
+ *
+ * @param ctx The CANN context containing necessary resources and
+ * configurations.
+ * @param dst The destination tensor where the result of the computation will be
+ * stored.
+ * @return true if the computation was successful; false otherwise.
+ */
+static bool ggml_cann_compute_forward(ggml_backend_cann_context& ctx,
+                                      struct ggml_tensor* dst) {
+    switch (dst->op) {
+        case GGML_OP_REPEAT:
+            ggml_cann_repeat(ctx, dst);
+            break;
+        case GGML_OP_GET_ROWS:
+            ggml_cann_get_rows(ctx, dst);
+            break;
+        case GGML_OP_DUP:
+            ggml_cann_dup(ctx, dst);
+            break;
+        case GGML_OP_ADD:
+            ggml_cann_add(ctx, dst);
+            break;
+        case GGML_OP_ACC:
+            ggml_cann_acc(ctx, dst);
+            break;
+        case GGML_OP_MUL:
+            ggml_cann_mul_div<aclnnMulGetWorkspaceSize, aclnnMul>(ctx, dst);
+            break;
+        case GGML_OP_DIV:
+            ggml_cann_mul_div<aclnnDivGetWorkspaceSize, aclnnDiv>(ctx, dst);
+            break;
+        case GGML_OP_UNARY:
+            switch (ggml_get_unary_op(dst)) {
+                case GGML_UNARY_OP_GELU:
+                    ggml_cann_activation<aclnnGeluGetWorkspaceSize, aclnnGelu>(
+                        ctx, dst);
+                    break;
+                case GGML_UNARY_OP_SILU:
+                    ggml_cann_activation<aclnnSiluGetWorkspaceSize, aclnnSilu>(
+                        ctx, dst);
+                    break;
+                // TODO: Use faster gelu??
+                case GGML_UNARY_OP_GELU_QUICK:
+                    ggml_cann_activation<aclnnGeluGetWorkspaceSize, aclnnGelu>(
+                        ctx, dst);
+                    break;
+                case GGML_UNARY_OP_TANH:
+                    ggml_cann_activation<aclnnTanhGetWorkspaceSize, aclnnTanh>(
+                        ctx, dst);
+                    break;
+                case GGML_UNARY_OP_RELU:
+                    ggml_cann_activation<aclnnReluGetWorkspaceSize, aclnnRelu>(
+                        ctx, dst);
+                    break;
+                case GGML_UNARY_OP_HARDSIGMOID:
+                    ggml_cann_activation<aclnnHardsigmoidGetWorkspaceSize,
+                                         aclnnHardsigmoid>(ctx, dst);
+                    break;
+                case GGML_UNARY_OP_HARDSWISH:
+                    ggml_cann_activation<aclnnHardswishGetWorkspaceSize,
+                                         aclnnHardswish>(ctx, dst);
+                    break;
+                default:
+                    return false;
+            }
+            break;
+        case GGML_OP_NORM:
+            ggml_cann_norm(ctx, dst);
+            break;
+        case GGML_OP_GROUP_NORM:
+            ggml_cann_group_norm(ctx, dst);
+            break;
+        case GGML_OP_CONCAT:
+            ggml_cann_concat(ctx, dst);
+            break;
+        case GGML_OP_UPSCALE:
+            ggml_cann_upsample_nearest2d(ctx, dst);
+            break;
+        case GGML_OP_PAD:
+            ggml_cann_pad(ctx, dst);
+            break;
+        case GGML_OP_ARANGE:
+            ggml_cann_arange(ctx, dst);
+            break;
+        case GGML_OP_TIMESTEP_EMBEDDING:
+            ggml_cann_timestep_embedding(ctx, dst);
+            break;
+        case GGML_OP_LEAKY_RELU:
+            ggml_cann_leaky_relu(ctx, dst);
+            break;
+        case GGML_OP_RMS_NORM:
+            ggml_cann_rms_norm(ctx, dst);
+            break;
+        case GGML_OP_MUL_MAT:
+            ggml_cann_mul_mat(ctx, dst);
+            break;
+        case GGML_OP_MUL_MAT_ID:
+            return false;
+        case GGML_OP_SCALE:
+            ggml_cann_scale(ctx, dst);
+            break;
+        case GGML_OP_SQR:
+            ggml_cann_sqr(ctx, dst);
+            break;
+        case GGML_OP_CLAMP:
+            ggml_cann_clamp(ctx, dst);
+            break;
+        case GGML_OP_CPY:
+            ggml_cann_cpy(ctx, dst);
+            break;
+        case GGML_OP_CONT:
+            ggml_cann_dup(ctx, dst);
+            break;
+        case GGML_OP_NONE:
+        case GGML_OP_RESHAPE:
+        case GGML_OP_VIEW:
+        case GGML_OP_PERMUTE:
+        case GGML_OP_TRANSPOSE:
+            break;
+        case GGML_OP_DIAG_MASK_INF:
+            ggml_cann_diag_mask(ctx, dst, -INFINITY);
+            break;
+        case GGML_OP_SOFT_MAX:
+            ggml_cann_softmax(ctx, dst);
+            break;
+        case GGML_OP_ROPE:
+            ggml_cann_rope(ctx, dst);
+            break;
+        case GGML_OP_IM2COL:
+            ggml_cann_im2col(ctx, dst);
+            break;
+        case GGML_OP_POOL_2D:
+            ggml_cann_pool2d(ctx, dst);
+            break;
+        case GGML_OP_SUM_ROWS:
+            ggml_cann_sum_rows(ctx, dst);
+            break;
+        case GGML_OP_ARGSORT:
+            ggml_cann_argsort(ctx, dst);
+            break;
+        default:
+            return false;
+    }
+
+    return true;
+}
+
+// backend
+/**
+ * @brief Retrieves the name associated with the CANN backend.
+ *
+ * This function returns the name assigned to the CANN backend, which is stored
+ * in the context of the provided backend structure.
+ *
+ * @param backend Pointer to the CANN backend structure.
+ * @return A pointer to a constant string representing the backend name.
+ */
+static const char* ggml_backend_cann_name(ggml_backend_t backend) {
+    ggml_backend_cann_context* cann_ctx =
+        (ggml_backend_cann_context*)backend->context;
+
+    return cann_ctx->name.c_str();
+}
+
+/**
+ * @brief Frees resources associated with the CANN backend.
+ *
+ * This function releases resources associated with the CANN backend context
+ * and resets the device associated with the backend to its initial state.
+ *
+ * @param backend Pointer to the CANN backend structure to be freed.
+ */
+static void ggml_backend_cann_free(ggml_backend_t backend) {
+    ggml_backend_cann_context* cann_ctx =
+        (ggml_backend_cann_context*)backend->context;
+    ACL_CHECK(aclrtSynchronizeDevice());
+    ACL_CHECK(aclrtResetDevice(cann_ctx->device));
+
+    // finalize when last backend freed.
+    if (cann_ctx->device == ggml_backend_cann_get_device_count() - 1) {
+        ACL_CHECK(aclFinalize());
+    }
+
+    delete cann_ctx;
+    delete backend;
+}
+
+/**
+ * @brief Sets tensor data asynchronously in the CANN backend.
+ *
+ * This function asynchronously sets tensor data in the CANN backend. Depending
+ * on the tensor type, it may perform data transformations before copying data
+ * to the device.
+ *
+ * @param backend Pointer to the CANN backend structure.
+ * @param tensor Pointer to the tensor structure to set data for.
+ * @param data Pointer to the host data to copy to the tensor.
+ * @param offset Offset in bytes within the host data.
+ * @param size Size of the data to copy in bytes.
+ */
+static void ggml_backend_cann_set_tensor_async(ggml_backend_t backend,
+                                               ggml_tensor *tensor,
+                                               const void *data,
+                                               size_t offset,
+                                               size_t size) {
+    ggml_backend_cann_context *cann_ctx =
+        (ggml_backend_cann_context *)backend->context;
+
+    if (!need_transform(tensor->type)) {
+        ACL_CHECK(aclrtMemcpyAsync((char *)tensor->data + offset, size, data,
+                                   size, ACL_MEMCPY_HOST_TO_DEVICE,
+                                   cann_ctx->stream()));
+    } else {
+        void *transform_buffer = malloc(size);
+        ggml_backend_cann_transform(tensor, data, transform_buffer);
+
+        ACL_CHECK(aclrtMemcpyAsync(
+            (char *)tensor->data + offset, size, transform_buffer, size,
+            ACL_MEMCPY_HOST_TO_DEVICE, cann_ctx->stream()));
+        ACL_CHECK(aclrtSynchronizeStream(cann_ctx->stream()));
+        free(transform_buffer);
+    }
+}
+
+static void ggml_backend_cann_get_tensor_async(
+    ggml_backend_t backend, const ggml_tensor *tensor, void *data,
+    size_t offset, size_t size) {
+    ggml_backend_cann_context *cann_ctx =
+        (ggml_backend_cann_context *)backend->context;
+    ggml_backend_buffer_t buf =
+        tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
+
+    GGML_ASSERT(buf->buft == ggml_backend_cann_buffer_type(cann_ctx->device) &&
+                "unsupported buffer type");
+
+    if (!need_transform(tensor->type)) {
+        ACL_CHECK(aclrtMemcpyAsync(data, size, (char *)tensor->data + offset,
+                                   size, ACL_MEMCPY_DEVICE_TO_HOST,
+                                   cann_ctx->stream()));
+    } else {
+        void *transform_buffer = malloc(size);
+        ACL_CHECK(aclrtMemcpyAsync(
+            transform_buffer, size, (char *)tensor->data + offset, size,
+            ACL_MEMCPY_DEVICE_TO_HOST, cann_ctx->stream()));
+        ACL_CHECK(aclrtSynchronizeStream(cann_ctx->stream()));
+        ggml_backend_cann_transform_back(tensor, transform_buffer, data);
+        free(transform_buffer);
+    }
+}
+
+/**
+ * @brief Asynchronously copies tensor data between CANN backends.
+ *
+ * This function copies tensor data asynchronously between two CANN backends. It
+ * checks if both tensors reside in CANN buffers and whether the devices support
+ * peer-to-peer access for direct copying. If not, it returns false.
+ *
+ * @param backend_src Pointer to the source CANN backend structure.
+ * @param backend_dst Pointer to the destination CANN backend structure.
+ * @param src Pointer to the source tensor to copy data from.
+ * @param dst Pointer to the destination tensor to copy data to.
+ * @return true if the copy operation succeeds, false otherwise.
+ */
+static bool ggml_backend_cann_cpy_tensor_async(
+    ggml_backend_t backend_src, ggml_backend_t backend_dst,
+    const ggml_tensor* src, ggml_tensor* dst) {
+    GGML_ASSERT(ggml_backend_is_cann(backend_src) ||
+                ggml_backend_is_cann(backend_dst));
+
+    if (!ggml_backend_buffer_is_cann(src->buffer) ||
+        !ggml_backend_buffer_is_cann(dst->buffer)) {
+        return false;
+    }
+
+    ggml_backend_buffer_t buf_src =
+        src->view_src ? src->view_src->buffer : src->buffer;
+    ggml_backend_buffer_t buf_dst =
+        dst->view_src ? dst->view_src->buffer : dst->buffer;
+
+    ggml_backend_cann_context* cann_ctx_src =
+        (ggml_backend_cann_context*)backend_src->context;
+    ggml_backend_cann_context* cann_ctx_dst =
+        (ggml_backend_cann_context*)backend_dst->context;
+
+    size_t copy_size = ggml_nbytes(dst);
+    if (backend_src != backend_dst) {
+        ggml_backend_cann_buffer_context* buf_ctx_src =
+            (ggml_backend_cann_buffer_context*)buf_src->context;
+        ggml_backend_cann_buffer_context* buf_ctx_dst =
+            (ggml_backend_cann_buffer_context*)buf_dst->context;
+
+        GGML_ASSERT(cann_ctx_src->device == buf_ctx_src->device);
+        GGML_ASSERT(cann_ctx_dst->device == buf_ctx_dst->device);
+
+        int32_t canAccessPeer = 0;
+        ACL_CHECK(aclrtDeviceCanAccessPeer(&canAccessPeer, cann_ctx_src->device,
+                                           cann_ctx_dst->device));
+        if (!canAccessPeer) {
+            return false;
+        }
+
+        // need open both directions for memcpyasync between devices.
+        ggml_cann_set_device(cann_ctx_dst->device);
+        ACL_CHECK(aclrtDeviceEnablePeerAccess(cann_ctx_src->device, 0));
+        ggml_cann_set_device(cann_ctx_src->device);
+        ACL_CHECK(aclrtDeviceEnablePeerAccess(cann_ctx_dst->device, 0));
+
+        ACL_CHECK(aclrtMemcpyAsync(dst->data, copy_size, src->data, copy_size,
+                                   ACL_MEMCPY_DEVICE_TO_DEVICE,
+                                   cann_ctx_src->stream()));
+
+        //TODO: workaround for Event didn`t work here.
+        aclrtSynchronizeStream(cann_ctx_src->stream());
+    } else {
+        // src and dst are on the same backend
+        ACL_CHECK(aclrtMemcpyAsync(dst->data, copy_size, src->data, copy_size,
+                                   ACL_MEMCPY_DEVICE_TO_DEVICE,
+                                   cann_ctx_dst->stream()));
+    }
+
+    return true;
+}
+
+/**
+ * @brief Synchronizes a CANN backend.
+ *
+ * This function synchronizes the specified CANN backend by waiting for all
+ * operations in its associated stream to complete.
+ *
+ * @param backend Pointer to the CANN backend structure to synchronize.
+ */
+static void ggml_backend_cann_synchronize(ggml_backend_t backend) {
+    ggml_backend_cann_context* cann_ctx =
+        (ggml_backend_cann_context*)backend->context;
+
+    ggml_cann_set_device(cann_ctx->device);
+
+    ACL_CHECK(aclrtSynchronizeStream(cann_ctx->stream()));
+}
+
+/**
+ * @brief Computes a computational graph using a CANN backend.
+ *
+ * This function computes the operations defined in the computational graph
+ * using the specified CANN backend.
+ *
+ * @param backend Pointer to the CANN backend structure to use for computation.
+ * @param cgraph Pointer to the computational graph structure containing nodes
+ *               representing operations to be computed.
+ * @return enum ggml_status Returns GGML_STATUS_SUCCESS if computation
+ *         completes successfully, otherwise an appropriate error status.
+ */
+static enum ggml_status ggml_backend_cann_graph_compute(
+    ggml_backend_t backend, ggml_cgraph* cgraph) {
+    ggml_backend_cann_context* cann_ctx =
+        (ggml_backend_cann_context*)backend->context;
+
+    ggml_cann_set_device(cann_ctx->device);
+
+    for (int i = 0; i < cgraph->n_nodes; i++) {
+        ggml_tensor* node = cgraph->nodes[i];
+
+        if (ggml_is_empty(node) || node->op == GGML_OP_NONE) {
+            continue;
+        }
+
+        bool ok = ggml_cann_compute_forward(*cann_ctx, node);
+
+        if (!ok) {
+            GGML_LOG_ERROR("%s: error: op not supported %s (%s)\n", __func__,
+                    node->name, ggml_op_name(node->op));
+        }
+        GGML_ASSERT(ok);
+    }
+
+    return GGML_STATUS_SUCCESS;
+}
+
+/**
+ * @brief Checks if the CANN backend supports a specific operation.
+ *
+ * This function checks whether the specified operation is supported by the
+ * CANN backend.
+ *
+ * @param backend Pointer to the CANN backend structure to check support for
+ *                the operation.
+ * @param op Pointer to the tensor representing the operation to check.
+ * @return bool Returns true if the operation is supported by the backend,
+ *              otherwise false.
+ */
+static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
+                                                    const ggml_tensor* op) {
+    switch (op->op) {
+        case GGML_OP_UNARY:
+            switch (ggml_get_unary_op(op)) {
+                case GGML_UNARY_OP_GELU:
+                case GGML_UNARY_OP_SILU:
+                case GGML_UNARY_OP_RELU:
+                case GGML_UNARY_OP_HARDSIGMOID:
+                case GGML_UNARY_OP_HARDSWISH:
+                case GGML_UNARY_OP_GELU_QUICK:
+                case GGML_UNARY_OP_TANH:
+                    return true;
+                default:
+                    return false;
+            }
+        case GGML_OP_MUL_MAT: {
+            switch (op->src[0]->type) {
+                case GGML_TYPE_F16:
+                case GGML_TYPE_F32:
+                case GGML_TYPE_Q8_0:
+                    // TODO: fix me
+                    // Current groupsize should not be greater than k-1 in
+                    // aclnnWeightQuantBatchMatmulV2GetWorkspaceSize().
+                case GGML_TYPE_Q4_0:
+                    return true;
+                default:
+                    return false;
+            }
+        }
+        case GGML_OP_MUL_MAT_ID:
+            return false;
+        // embedding
+        case GGML_OP_GET_ROWS: {
+            switch (op->src[0]->type) {
+                case GGML_TYPE_F32:
+                case GGML_TYPE_F16:
+                case GGML_TYPE_Q4_0:
+                case GGML_TYPE_Q8_0:
+                    return true;
+                default:
+                    return false;
+            }
+        } break;
+        case GGML_OP_CPY: {
+            switch (op->type) {
+                case GGML_TYPE_F32:
+                case GGML_TYPE_F16:
+                case GGML_TYPE_Q8_0:
+                case GGML_TYPE_Q4_0:
+                    return true;
+                default:
+                    return false;
+            }
+        }
+        case GGML_OP_DUP:
+        case GGML_OP_REPEAT:
+        case GGML_OP_CONCAT:
+        case GGML_OP_NONE:
+        case GGML_OP_RESHAPE:
+        case GGML_OP_VIEW:
+        case GGML_OP_PERMUTE:
+        case GGML_OP_TRANSPOSE:
+        case GGML_OP_NORM:
+        case GGML_OP_ADD:
+        case GGML_OP_MUL:
+        case GGML_OP_DIV:
+        case GGML_OP_RMS_NORM:
+        case GGML_OP_SCALE:
+        case GGML_OP_SQR:
+        case GGML_OP_CLAMP:
+        case GGML_OP_CONT:
+        case GGML_OP_DIAG_MASK_INF:
+        case GGML_OP_SOFT_MAX:
+        case GGML_OP_ROPE:
+        case GGML_OP_IM2COL:
+        case GGML_OP_POOL_2D:
+        case GGML_OP_SUM_ROWS:
+        case GGML_OP_ARGSORT:
+        case GGML_OP_ACC:
+        case GGML_OP_GROUP_NORM:
+        case GGML_OP_UPSCALE:
+        case GGML_OP_PAD:
+        case GGML_OP_ARANGE:
+        case GGML_OP_TIMESTEP_EMBEDDING:
+        case GGML_OP_LEAKY_RELU:
+            return true;
+        default:
+            return false;
+    }
+
+    GGML_UNUSED(dev);
+}
+
+/**
+ * @brief Checks if the backend buffer type is associated with the CANN backend.
+ *
+ * This function checks whether the provided backend buffer type is associated
+ * with the CANN backend based on the comparison of its name retrieval function
+ * pointer.
+ *
+ * @param buft Pointer to the backend buffer type to check.
+ * @return bool Returns true if the buffer type is associated with the CANN
+ * backend, otherwise false.
+ */
+static bool ggml_backend_buft_is_cann(ggml_backend_buffer_type_t buft) {
+    return buft->iface.get_name == ggml_backend_cann_buffer_type_name;
+}
+
+/**
+ * @brief Determines if a tensor operation should be offloaded to the CANN
+ * backend.
+ *
+ * This function checks if a given tensor operation should be offloaded to the
+ * CANN backend based on the operation type and the size of the tensor. It
+ * returns true if the second dimension (ne[1]) of the tensor is greater than or
+ * equal to the minimum batch size and the operation is not GGML_OP_GET_ROWS.
+ *
+ * @param backend Pointer to the CANN backend.
+ * @param op Pointer to the tensor operation to check.
+ * @return bool Returns true if the operation should be offloaded, otherwise
+ * false.
+ */
+static bool ggml_backend_cann_offload_op(ggml_backend_dev_t dev,
+                                                   const ggml_tensor* op) {
+    const int min_batch_size = 32;
+    GGML_UNUSED(dev);
+
+    return op->ne[1] >= min_batch_size && op->op != GGML_OP_GET_ROWS;
+}
+
+/**
+ * @brief Records an event on the CANN backend stream.
+ *
+ * This function records the given event on the ACL runtime stream associated
+ * with the backend context.
+ *
+ * @param event Pointer to the event structure to be recorded.
+ */
+static void ggml_backend_cann_event_record(ggml_backend_t backend, ggml_backend_event_t event) {
+    ggml_backend_cann_context* cann_ctx =
+        (ggml_backend_cann_context*)backend->context;
+    ACL_CHECK(aclrtRecordEvent((aclrtEvent)event->context, cann_ctx->stream()));
+}
+
+/**
+ * @brief Waits for a recorded event to complete on the CANN backend stream.
+ *
+ * This function makes the given backend wait for the event to complete on its
+ * ACL runtime stream.
+ *
+ * @param backend Pointer to the backend structure.
+ * @param event Pointer to the event structure that the backend needs to wait
+ * for.
+ */
+static void ggml_backend_cann_event_wait(ggml_backend_t backend,
+                                         ggml_backend_event_t event) {
+    ggml_backend_cann_context* cann_ctx =
+        (ggml_backend_cann_context*)backend->context;
+    if (ggml_backend_is_cann(backend)) {
+        ACL_CHECK(aclrtStreamWaitEvent(cann_ctx->stream(),
+                                       (aclrtEvent)event->context));
+    } else {
+        GGML_ABORT("fatal error");
+    }
+}
+
+/**
+ * @brief Structure defining the interface for the CANN backend.
+ *
+ * This structure contains function pointers for various operations
+ * supported by the CANN backend, including name retrieval, memory
+ * management, tensor operations, synchronization, and event handling.
+ */
+static const ggml_backend_i ggml_backend_cann_interface = {
+    /* .get_name                = */ ggml_backend_cann_name,
+    /* .free                    = */ ggml_backend_cann_free,
+    /* .set_tensor_async        = */ ggml_backend_cann_set_tensor_async,
+    /* .get_tensor_async        = */ ggml_backend_cann_get_tensor_async,
+    /* .cpy_tensor_async        = */ ggml_backend_cann_cpy_tensor_async,
+    /* .synchronize             = */ ggml_backend_cann_synchronize,
+    /* .graph_plan_create       = */ NULL,
+    /* .graph_plan_free         = */ NULL,
+    /* .graph_plan_update       = */ NULL,
+    /* .graph_plan_compute      = */ NULL,
+    /* .graph_compute           = */ ggml_backend_cann_graph_compute,
+    /* .event_record            = */ ggml_backend_cann_event_record,
+    /* .event_wait              = */ ggml_backend_cann_event_wait,
+};
+
+/**
+ * @brief Return the hardcoded GUID for the CANN backend.
+ *
+ * This function returns a static GUID which uniquely identifies the CANN
+ * backend.
+ *
+ * @return A pointer to the static GUID.
+ */
+static ggml_guid_t ggml_backend_cann_guid() {
+    static ggml_guid guid = {0xa1, 0x94, 0xaf, 0xac, 0xbd, 0x4f, 0x47, 0x34,
+                             0xbe, 0x1a, 0x9e, 0x71, 0x1f, 0x9e, 0xed, 0x64};
+    return &guid;
+}
+
+// backend device
+struct ggml_backend_cann_device_context {
+    int device;
+    std::string name;
+    std::string description;
+};
+
+static const char * ggml_backend_cann_device_get_name(ggml_backend_dev_t dev) {
+    ggml_backend_cann_device_context * ctx = (ggml_backend_cann_device_context *)dev->context;
+    return ctx->name.c_str();
+}
+
+static const char* ggml_backend_cann_device_get_description(ggml_backend_dev_t dev) {
+    ggml_backend_cann_device_context * ctx = (ggml_backend_cann_device_context *)dev->context;
+    return ctx->description.c_str();
+}
+
+static void ggml_backend_cann_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
+    ggml_backend_cann_device_context * ctx = (ggml_backend_cann_device_context *)dev->context;
+    ggml_backend_cann_get_device_memory(ctx->device, free, total);
+}
+
+static enum ggml_backend_dev_type ggml_backend_cann_device_get_type(ggml_backend_dev_t dev) {
+    GGML_UNUSED(dev);
+    return GGML_BACKEND_DEVICE_TYPE_GPU;
+}
+
+static void ggml_backend_cann_device_get_props(ggml_backend_dev_t dev, ggml_backend_dev_props * props) {
+    props->name        = ggml_backend_cann_device_get_name(dev);
+    props->description = ggml_backend_cann_device_get_description(dev);
+    props->type        = ggml_backend_cann_device_get_type(dev);
+    ggml_backend_cann_device_get_memory(dev, &props->memory_free, &props->memory_total);
+
+    bool host_buffer = getenv("GGML_CANN_NO_PINNED") == nullptr;
+
+    props->caps = {
+        /* .async                 = */ false,
+        /* .host_buffer           = */ host_buffer,
+        /* .buffer_from_host_ptr  = */ false,
+        /* .events                = */ true,
+    };
+}
+
+static ggml_backend_t ggml_backend_cann_device_init(ggml_backend_dev_t dev, const char * params) {
+    GGML_UNUSED(params);
+    ggml_backend_cann_device_context * ctx = (ggml_backend_cann_device_context *)dev->context;
+    return ggml_backend_cann_init(ctx->device);
+}
+
+/**
+ * @brief Checks if the CANN backend supports a specific backend buffer type.
+ *
+ * This function determines whether the CANN backend supports the given backend
+ * buffer type by comparing the device context of the backend and buffer type.
+ * It returns true if the devices are same between the backend context and
+ * buffer type context.
+ *
+ * @param backend Pointer to the CANN backend.
+ * @param buft Pointer to the backend buffer type to check.
+ * @return bool Returns true if the CANN backend supports the buffer type,
+ *              otherwise false.
+ */
+static bool ggml_backend_cann_supports_buft(
+    ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
+    if (ggml_backend_buft_is_cann(buft)) {
+        ggml_backend_cann_device_context * dev_ctx = (ggml_backend_cann_device_context *)dev->context;
+        ggml_backend_cann_buffer_type_context * buft_ctx =
+                        (ggml_backend_cann_buffer_type_context *)buft->context;
+        return buft_ctx->device == dev_ctx->device;
+    }
+    return false;
+}
+
+static ggml_backend_buffer_type_t ggml_backend_cann_device_get_buffer_type(ggml_backend_dev_t dev) {
+    ggml_backend_cann_device_context * ctx = (ggml_backend_cann_device_context *)dev->context;
+    return ggml_backend_cann_buffer_type(ctx->device);
+}
+
+static ggml_backend_buffer_type_t ggml_backend_cann_device_get_host_buffer_type(ggml_backend_dev_t dev) {
+    GGML_UNUSED(dev);
+    return ggml_backend_cann_host_buffer_type();
+}
+
+/**
+ * @brief Creates a new event for the CANN backend device.
+ *
+ * This function initializes a new event for the CANN backend by setting the
+ * device and creating an ACL runtime event. The created event is then wrapped
+ * in a ggml_backend_event structure and returned.
+ *
+ * @param backend Pointer to the CANN backend.
+ * @return ggml_backend_event_t Returns a pointer to the new event structure.
+ */
+static ggml_backend_event_t ggml_backend_cann_device_event_new(
+    ggml_backend_dev_t dev) {
+    ggml_backend_cann_device_context * dev_ctx = (ggml_backend_cann_device_context *)dev->context;
+
+    ggml_cann_set_device(dev_ctx->device);
+
+    aclrtEvent event;
+    ACL_CHECK(aclrtCreateEvent(&event));
+
+    return new ggml_backend_event{
+        /* .device = */ ggml_backend_reg_dev_get(ggml_backend_cann_reg(), dev_ctx->device),
+        /* .context = */ event,
+    };
+}
+
+/**
+ * @brief Frees a CANN backend event.
+ *
+ * This function destroys the ACL runtime event associated with the given CANN
+ * backend event and then deletes the event structure itself.
+ *
+ * @param event Pointer to the event structure to be freed.
+ */
+static void ggml_backend_cann_device_event_free(ggml_backend_dev_t dev, ggml_backend_event_t event) {
+    ACL_CHECK(aclrtDestroyEvent((aclrtEvent)event->context));
+
+    delete event;
+    GGML_UNUSED(dev);
+}
+
+/**
+ * @brief Synchronizes the given event on the CANN backend.
+ *
+ * This function waits for the specified event to complete on the ACL runtime.
+ *
+ * @param event Pointer to the event structure to be synchronized.
+ */
+static void ggml_backend_cann_device_event_synchronize(ggml_backend_dev_t dev, ggml_backend_event_t event) {
+    ACL_CHECK(aclrtSynchronizeEvent((aclrtEvent)event->context));
+
+    GGML_UNUSED(dev);
+}
+
+static const ggml_backend_device_i ggml_backend_cann_device_interface = {
+    /* .get_name                = */ ggml_backend_cann_device_get_name,
+    /* .get_description         = */ ggml_backend_cann_device_get_description,
+    /* .get_memory              = */ ggml_backend_cann_device_get_memory,
+    /* .get_type                = */ ggml_backend_cann_device_get_type,
+    /* .get_props               = */ ggml_backend_cann_device_get_props,
+    /* .init_backend            = */ ggml_backend_cann_device_init,    // called for every card
+    /* .get_buffer_type         = */ ggml_backend_cann_device_get_buffer_type,
+    /* .get_host_buffer_type    = */ ggml_backend_cann_device_get_host_buffer_type,
+    /* .buffer_from_host_ptr    = */ NULL, // not supported for CANN
+    /* .supports_op             = */ ggml_backend_cann_supports_op,
+    /* .supports_buft           = */ ggml_backend_cann_supports_buft,
+    /* .offload_op              = */ ggml_backend_cann_offload_op,
+    /* .event_new               = */ ggml_backend_cann_device_event_new,
+    /* .event_free              = */ ggml_backend_cann_device_event_free,
+    /* .event_synchronize       = */ ggml_backend_cann_device_event_synchronize,
+};
+
+
+// backend reg
+struct ggml_backend_cann_reg_context {
+    std::vector<ggml_backend_dev_t> devices;
+};
+
+static const char * ggml_backend_cann_reg_get_name(ggml_backend_reg_t reg) {
+    GGML_UNUSED(reg);
+    return GGML_CANN_NAME;
+}
+
+static size_t ggml_backend_cann_reg_get_device_count(ggml_backend_reg_t reg) {
+    ggml_backend_cann_reg_context * ctx = (ggml_backend_cann_reg_context *)reg->context;
+    return ctx->devices.size();
+}
+
+static ggml_backend_dev_t ggml_backend_cann_reg_get_device(ggml_backend_reg_t reg, size_t index) {
+    ggml_backend_cann_reg_context * ctx = (ggml_backend_cann_reg_context *)reg->context;
+    GGML_ASSERT(index < ctx->devices.size());
+    return ctx->devices[index];
+}
+
+static void * ggml_backend_cann_reg_get_proc_address(ggml_backend_reg_t reg, const char * name) {
+    GGML_UNUSED(reg);
+    GGML_UNUSED(name);
+    // reserved for future use
+    return nullptr;
+}
+
+static const ggml_backend_reg_i ggml_backend_cann_reg_interface = {
+    /* .get_name          = */ ggml_backend_cann_reg_get_name,
+    /* .get_device_count  = */ ggml_backend_cann_reg_get_device_count,
+    /* .get_device_get    = */ ggml_backend_cann_reg_get_device,
+    /* .get_proc_address  = */ ggml_backend_cann_reg_get_proc_address,
+};
+
+// backend registry, called only once for cann backend
+ggml_backend_reg_t ggml_backend_cann_reg() {
+    static ggml_backend_reg reg;
+    static bool initialized = false;
+
+    {
+        static std::mutex mutex;
+        std::lock_guard<std::mutex> lock(mutex);
+        if (!initialized) {
+            aclInit(nullptr);
+            ggml_backend_cann_reg_context * ctx = new ggml_backend_cann_reg_context;
+
+            for (int i = 0; i < ggml_cann_info().device_count; i++) {
+                ggml_backend_cann_device_context* dev_ctx = new ggml_backend_cann_device_context();
+                dev_ctx->description = aclrtGetSocName();
+                dev_ctx->device = i;
+                dev_ctx->name = GGML_CANN_NAME + std::to_string(i);
+                ggml_cann_set_device(i);
+                ggml_backend_dev_t dev = new ggml_backend_device {
+                    /* .interface = */ ggml_backend_cann_device_interface,
+                    /* .reg       = */ &reg,
+                    /* .context   = */ dev_ctx
+                };
+                ctx->devices.push_back(dev);
+            }
+
+            reg = ggml_backend_reg {
+                /* .interface = */ ggml_backend_cann_reg_interface,
+                /* .context   = */ ctx
+            };
+        }
+
+        initialized = true;
+    }
+
+    return &reg;
+}
+
+ggml_backend_t ggml_backend_cann_init(int32_t device) {
+    aclInit(nullptr);
+    if (device < 0 || device >= ggml_backend_cann_get_device_count()) {
+        GGML_LOG_ERROR("%s: error: invalid device %d\n", __func__, device);
+        return nullptr;
+    }
+
+    ggml_backend_cann_context* ctx = new ggml_backend_cann_context(device);
+    if (ctx == nullptr) {
+        GGML_LOG_ERROR("%s: error: failed to allocate context\n", __func__);
+        return nullptr;
+    }
+    ggml_cann_set_device(ctx->device);
+    ggml_backend_t cann_backend =
+        new ggml_backend{/* .guid      = */ ggml_backend_cann_guid(),
+                         /* .interface = */ ggml_backend_cann_interface,
+                         /* .device    = */ ggml_backend_reg_dev_get(ggml_backend_cann_reg(), device),
+                         /* .context   = */ ctx};
+
+    return cann_backend;
+}
+
+bool ggml_backend_is_cann(ggml_backend_t backend) {
+    return backend != NULL &&
+           ggml_guid_matches(backend->guid, ggml_backend_cann_guid());
+}
+
+int32_t ggml_backend_cann_get_device_count() {
+    return ggml_cann_info().device_count;
+}
+
+void ggml_backend_cann_get_device_description(
+    int32_t device, char* description, size_t description_size) {
+    ggml_cann_set_device(device);
+    const char* soc_name = aclrtGetSocName();
+    snprintf(description, description_size, "%s", soc_name);
+}
+
+void ggml_backend_cann_get_device_memory(int32_t device, size_t* free,
+                                         size_t* total) {
+    ggml_cann_set_device(device);
+    ACL_CHECK(aclrtGetMemInfo(ACL_HBM_MEM, free, total));
+}

ggml/src/ggml-cpu/CMakeLists.txt

@@ -0,0 +1,244 @@
+add_library(ggml-cpu
+            ggml-cpu.c
+            ggml-cpu.cpp
+            ggml-cpu-aarch64.c
+            ggml-cpu-aarch64.h
+            ggml-cpu-quants.c
+            ggml-cpu-quants.h
+            )
+
+target_link_libraries(ggml-cpu PRIVATE ggml-base)
+target_include_directories(ggml-cpu PRIVATE . ..)
+
+if (APPLE AND GGML_ACCELERATE)
+    find_library(ACCELERATE_FRAMEWORK Accelerate)
+    if (ACCELERATE_FRAMEWORK)
+        message(STATUS "Accelerate framework found")
+
+        add_compile_definitions(GGML_USE_ACCELERATE)
+        add_compile_definitions(ACCELERATE_NEW_LAPACK)
+        add_compile_definitions(ACCELERATE_LAPACK_ILP64)
+
+        target_link_libraries(ggml-cpu PRIVATE ${ACCELERATE_FRAMEWORK})
+    else()
+        message(WARNING "Accelerate framework not found")
+    endif()
+endif()
+
+if (GGML_OPENMP)
+    find_package(OpenMP)
+    if (OpenMP_FOUND)
+        message(STATUS "OpenMP found")
+
+        add_compile_definitions(GGML_USE_OPENMP)
+
+        target_link_libraries(ggml-cpu PRIVATE OpenMP::OpenMP_C OpenMP::OpenMP_CXX)
+
+        # FIXME: should be replaced with a compiler id check
+        #if (GGML_MUSA)
+        #    list(APPEND GGML_CPU_EXTRA_INCLUDES     "/usr/lib/llvm-14/lib/clang/14.0.0/include")
+        #    list(APPEND GGML_CPU_EXTRA_LIBS_PRIVATE "/usr/lib/llvm-14/lib/libomp.so")
+        #endif()
+    else()
+        message(WARNING "OpenMP not found")
+    endif()
+endif()
+
+if (GGML_LLAMAFILE)
+    message(STATUS "Using llamafile")
+
+    add_compile_definitions(GGML_USE_LLAMAFILE)
+
+    target_sources(ggml-cpu PRIVATE
+                    llamafile/sgemm.cpp
+                    llamafile/sgemm.h)
+endif()
+
+if (GGML_CPU_HBM)
+    find_library(memkind memkind REQUIRED)
+
+    message(STATUS "Using memkind for CPU HBM")
+
+    add_compile_definitions(GGML_USE_CPU_HBM)
+
+    target_link_libraries(ggml-cpu PUBLIC memkind)
+endif()
+
+if (CMAKE_OSX_ARCHITECTURES      STREQUAL "arm64" OR
+    CMAKE_GENERATOR_PLATFORM_LWR STREQUAL "arm64" OR
+    (NOT CMAKE_OSX_ARCHITECTURES      AND
+     NOT CMAKE_GENERATOR_PLATFORM_LWR AND
+         CMAKE_SYSTEM_PROCESSOR MATCHES "^(aarch64|arm.*|ARM64)$"))
+
+    message(STATUS "ARM detected")
+
+    if (MSVC)
+        add_compile_definitions(__aarch64__) # MSVC defines _M_ARM64 instead
+        add_compile_definitions(__ARM_NEON)
+        add_compile_definitions(__ARM_FEATURE_FMA)
+
+        set(CMAKE_REQUIRED_FLAGS_PREV ${CMAKE_REQUIRED_FLAGS})
+        string(JOIN " " CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS} "/arch:armv8.2")
+
+        check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int8x16_t _a, _b; int32x4_t _s = vdotq_s32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_DOTPROD)
+        if (GGML_COMPILER_SUPPORT_DOTPROD)
+            add_compile_definitions(__ARM_FEATURE_DOTPROD)
+        endif ()
+
+        check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int8x16_t _a, _b; int32x4_t _s = vmlaq_f32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_MATMUL_INT8)
+
+        if (GGML_COMPILER_SUPPORT_MATMUL_INT8)
+            add_compile_definitions(__ARM_FEATURE_MATMUL_INT8)
+        endif ()
+
+        check_cxx_source_compiles("#include <arm_neon.h>\nint main() { float16_t _a; float16x8_t _s = vdupq_n_f16(_a); return 0; }" GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC)
+        if (GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC)
+            add_compile_definitions(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
+        endif ()
+
+        set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_PREV})
+    else()
+        check_cxx_compiler_flag(-mfp16-format=ieee COMPILER_SUPPORTS_FP16_FORMAT_I3E)
+        if (NOT "${COMPILER_SUPPORTS_FP16_FORMAT_I3E}" STREQUAL "")
+            list(APPEND ARCH_FLAGS -mfp16-format=ieee)
+        endif()
+        if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv6")
+            # Raspberry Pi 1, Zero
+            list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access)
+        endif()
+        if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv7")
+            if ("${CMAKE_SYSTEM_NAME}" STREQUAL "Android")
+                # Android armeabi-v7a
+                list(APPEND ARCH_FLAGS -mfpu=neon-vfpv4 -mno-unaligned-access -funsafe-math-optimizations)
+            else()
+                # Raspberry Pi 2
+                list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access -funsafe-math-optimizations)
+            endif()
+        endif()
+        if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv8")
+            # Android arm64-v8a
+            # Raspberry Pi 3, 4, Zero 2 (32-bit)
+            list(APPEND ARCH_FLAGS -mno-unaligned-access)
+        endif()
+        if (GGML_SVE)
+            list(APPEND ARCH_FLAGS -march=armv8.6-a+sve)
+        endif()
+    endif()
+elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LWR MATCHES "^(x86_64|i686|amd64|x64|win32)$" OR
+        (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND
+         CMAKE_SYSTEM_PROCESSOR MATCHES "^(x86_64|i686|AMD64)$"))
+    message(STATUS "x86 detected")
+    if (MSVC)
+        # instruction set detection for MSVC only
+        if (GGML_NATIVE)
+            # TODO: improve, should not reference files from the parent folder
+            include(cmake/FindSIMD.cmake)
+        endif ()
+        if (GGML_AVX512)
+            list(APPEND ARCH_FLAGS /arch:AVX512)
+            # MSVC has no compile-time flags enabling specific
+            # AVX512 extensions, neither it defines the
+            # macros corresponding to the extensions.
+            # Do it manually.
+            if (GGML_AVX512_VBMI)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VBMI__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VBMI__>)
+            endif()
+            if (GGML_AVX512_VNNI)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VNNI__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VNNI__>)
+            endif()
+            if (GGML_AVX512_BF16)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512BF16__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512BF16__>)
+            endif()
+            if (GGML_AMX_TILE)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AMX_TILE__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AMX_TILE__>)
+            endif()
+            if (GGML_AMX_INT8)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AMX_INT8__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AMX_INT8__>)
+            endif()
+            if (GGML_AMX_BF16)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AMX_BF16__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AMX_BF16__>)
+            endif()
+        elseif (GGML_AVX2)
+            list(APPEND ARCH_FLAGS /arch:AVX2)
+        elseif (GGML_AVX)
+            list(APPEND ARCH_FLAGS /arch:AVX)
+        endif()
+    else()
+        if (GGML_NATIVE)
+            list(APPEND ARCH_FLAGS -march=native)
+        endif()
+        if (GGML_F16C)
+            list(APPEND ARCH_FLAGS -mf16c)
+        endif()
+        if (GGML_FMA)
+            list(APPEND ARCH_FLAGS -mfma)
+        endif()
+        if (GGML_AVX)
+            list(APPEND ARCH_FLAGS -mavx)
+        endif()
+        if (GGML_AVX2)
+            list(APPEND ARCH_FLAGS -mavx2)
+        endif()
+        if (GGML_AVX512)
+            list(APPEND ARCH_FLAGS -mavx512f)
+            list(APPEND ARCH_FLAGS -mavx512dq)
+            list(APPEND ARCH_FLAGS -mavx512bw)
+        endif()
+        if (GGML_AVX512_VBMI)
+            list(APPEND ARCH_FLAGS -mavx512vbmi)
+        endif()
+        if (GGML_AVX512_VNNI)
+            list(APPEND ARCH_FLAGS -mavx512vnni)
+        endif()
+        if (GGML_AVX512_BF16)
+            list(APPEND ARCH_FLAGS -mavx512bf16)
+        endif()
+        if (GGML_AMX_TILE)
+            list(APPEND ARCH_FLAGS -mamx-tile)
+        endif()
+        if (GGML_AMX_INT8)
+            list(APPEND ARCH_FLAGS -mamx-int8)
+        endif()
+        if (GGML_AMX_BF16)
+            list(APPEND ARCH_FLAGS -mamx-bf16)
+        endif()
+    endif()
+elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64")
+    message(STATUS "PowerPC detected")
+    execute_process(COMMAND bash -c "grep POWER10 /proc/cpuinfo | head -n 1"
+                   OUTPUT_VARIABLE POWER10_M)
+    string(FIND ${POWER10_M} "POWER10" substring_index)
+    if(${substring_index} GREATER_EQUAL 0)
+       list(APPEND ARCH_FLAGS -mcpu=power10)
+    elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64le")
+       list(APPEND ARCH_FLAGS -mcpu=powerpc64le)
+    else()
+        list(APPEND ARCH_FLAGS -mcpu=native -mtune=native)
+        #TODO: Add  targets for Power8/Power9 (Altivec/VSX) and Power10(MMA) and query for big endian systems (ppc64/le/be)
+    endif()
+elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64")
+    message(STATUS "loongarch64 detected")
+
+    list(APPEND ARCH_FLAGS -march=loongarch64)
+    if (GGML_LASX)
+        list(APPEND ARCH_FLAGS -mlasx)
+    endif()
+    if (GGML_LSX)
+        list(APPEND ARCH_FLAGS -mlsx)
+    endif()
+else()
+    message(STATUS "Unknown architecture")
+endif()
+
+target_compile_options(ggml-cpu PRIVATE "$<$<COMPILE_LANGUAGE:CXX>:${ARCH_FLAGS}>")
+target_compile_options(ggml-cpu PRIVATE "$<$<COMPILE_LANGUAGE:C>:${ARCH_FLAGS}>")
+
+if (EMSCRIPTEN)
+    set_target_properties(ggml-cpu PROPERTIES COMPILE_FLAGS "-msimd128")
+endif()

ggml/src/ggml-cpu/cmake/FindSIMD.cmake

@@ -0,0 +1,100 @@
+include(CheckCSourceRuns)
+
+set(AVX_CODE "
+    #include <immintrin.h>
+    int main()
+    {
+        __m256 a;
+        a = _mm256_set1_ps(0);
+        return 0;
+    }
+")
+
+set(AVX512_CODE "
+    #include <immintrin.h>
+    int main()
+    {
+        __m512i a = _mm512_set_epi8(0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0);
+        __m512i b = a;
+        __mmask64 equality_mask = _mm512_cmp_epi8_mask(a, b, _MM_CMPINT_EQ);
+        return 0;
+    }
+")
+
+set(AVX2_CODE "
+    #include <immintrin.h>
+    int main()
+    {
+        __m256i a = {0};
+        a = _mm256_abs_epi16(a);
+        __m256i x;
+        _mm256_extract_epi64(x, 0); // we rely on this in our AVX2 code
+        return 0;
+    }
+")
+
+set(FMA_CODE "
+    #include <immintrin.h>
+    int main()
+    {
+        __m256 acc = _mm256_setzero_ps();
+        const __m256 d = _mm256_setzero_ps();
+        const __m256 p = _mm256_setzero_ps();
+        acc = _mm256_fmadd_ps( d, p, acc );
+        return 0;
+    }
+")
+
+macro(check_sse type flags)
+    set(__FLAG_I 1)
+    set(CMAKE_REQUIRED_FLAGS_SAVE ${CMAKE_REQUIRED_FLAGS})
+    foreach (__FLAG ${flags})
+        if (NOT ${type}_FOUND)
+            set(CMAKE_REQUIRED_FLAGS ${__FLAG})
+            check_c_source_runs("${${type}_CODE}" HAS_${type}_${__FLAG_I})
+            if (HAS_${type}_${__FLAG_I})
+                set(${type}_FOUND TRUE CACHE BOOL "${type} support")
+                set(${type}_FLAGS "${__FLAG}" CACHE STRING "${type} flags")
+            endif()
+            math(EXPR __FLAG_I "${__FLAG_I}+1")
+        endif()
+    endforeach()
+    set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_SAVE})
+
+    if (NOT ${type}_FOUND)
+        set(${type}_FOUND FALSE CACHE BOOL "${type} support")
+        set(${type}_FLAGS "" CACHE STRING "${type} flags")
+    endif()
+
+    mark_as_advanced(${type}_FOUND ${type}_FLAGS)
+endmacro()
+
+# flags are for MSVC only!
+check_sse("AVX" " ;/arch:AVX")
+if (NOT ${AVX_FOUND})
+    set(GGML_AVX OFF)
+else()
+    set(GGML_AVX ON)
+endif()
+
+check_sse("AVX2" " ;/arch:AVX2")
+check_sse("FMA" " ;/arch:AVX2")
+if ((NOT ${AVX2_FOUND}) OR (NOT ${FMA_FOUND}))
+    set(GGML_AVX2 OFF)
+else()
+    set(GGML_AVX2 ON)
+endif()
+
+check_sse("AVX512" " ;/arch:AVX512")
+if (NOT ${AVX512_FOUND})
+    set(GGML_AVX512 OFF)
+else()
+    set(GGML_AVX512 ON)
+endif()

ggml/src/ggml-cpu/ggml-cpu-aarch64.h

@@ -0,0 +1,27 @@
+#pragma once
+
+#include "ggml.h"
+
+// GGML internal header
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Quantization
+void quantize_mat_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t nrows, int64_t n_per_row, int64_t blck_size_interleave);
+
+// GEMV
+void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+
+// GEMM
+void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+
+#ifdef __cplusplus
+}
+#endif
+

ggml/src/ggml-cpu/ggml-cpu-impl.h

@@ -0,0 +1,371 @@
+#pragma once
+
+// GGML CPU internal header
+
+#include "ggml.h"
+#include "ggml-impl.h"
+#include <stdlib.h> // load `stdlib.h` before other headers to work around MinGW bug: https://sourceforge.net/p/mingw-w64/bugs/192/
+//#include <stddef.h>
+#include <stdbool.h>
+#include <string.h> // memcpy
+#include <math.h>   // fabsf
+
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#if defined(_MSC_VER)
+
+#define m512bh(p) p
+#define m512i(p) p
+
+#else
+
+#define m512bh(p) (__m512bh)(p)
+#define m512i(p) (__m512i)(p)
+
+#endif
+
+// __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512
+#if defined(_MSC_VER) && (defined(__AVX2__) || defined(__AVX512F__))
+#ifndef __FMA__
+#define __FMA__
+#endif
+#ifndef __F16C__
+#define __F16C__
+#endif
+#endif
+
+// __SSE3__ and __SSSE3__ are not defined in MSVC, but SSE3/SSSE3 are present when AVX/AVX2/AVX512 are available
+#if defined(_MSC_VER) && (defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__))
+#ifndef __SSE3__
+#define __SSE3__
+#endif
+#ifndef __SSSE3__
+#define __SSSE3__
+#endif
+#endif
+
+#if defined(__ARM_FEATURE_SVE)
+#include <arm_sve.h>
+#include <sys/prctl.h>
+#endif
+
+// 16-bit float
+// on Arm, we use __fp16
+// on x86, we use uint16_t
+#if defined(__ARM_NEON)
+
+// if YCM cannot find <arm_neon.h>, make a symbolic link to it, for example:
+//
+//   $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/
+//
+#include <arm_neon.h>
+
+#ifdef _MSC_VER
+
+typedef uint16_t ggml_fp16_internal_t;
+
+#define ggml_vld1q_u32(w,x,y,z) { ((w) + ((uint64_t)(x) << 32)), ((y) + ((uint64_t)(z) << 32)) }
+
+#else
+
+typedef __fp16 ggml_fp16_internal_t;
+
+#define ggml_vld1q_u32(w,x,y,z) { (w), (x), (y), (z) }
+
+#endif // _MSC_VER
+
+#if !defined(__aarch64__)
+
+// 32-bit ARM compatibility
+
+// vaddlvq_s16
+// vpaddq_s16
+// vpaddq_s32
+// vaddvq_s32
+// vaddvq_f32
+// vmaxvq_f32
+// vcvtnq_s32_f32
+// vzip1_u8
+// vzip2_u8
+
+inline static int32_t vaddlvq_s16(int16x8_t v) {
+    int32x4_t v0 = vreinterpretq_s32_s64(vpaddlq_s32(vpaddlq_s16(v)));
+    return vgetq_lane_s32(v0, 0) + vgetq_lane_s32(v0, 2);
+}
+
+inline static int16x8_t vpaddq_s16(int16x8_t a, int16x8_t b) {
+    int16x4_t a0 = vpadd_s16(vget_low_s16(a), vget_high_s16(a));
+    int16x4_t b0 = vpadd_s16(vget_low_s16(b), vget_high_s16(b));
+    return vcombine_s16(a0, b0);
+}
+
+inline static int32x4_t vpaddq_s32(int32x4_t a, int32x4_t b) {
+    int32x2_t a0 = vpadd_s32(vget_low_s32(a), vget_high_s32(a));
+    int32x2_t b0 = vpadd_s32(vget_low_s32(b), vget_high_s32(b));
+    return vcombine_s32(a0, b0);
+}
+
+inline static int32_t vaddvq_s32(int32x4_t v) {
+    return vgetq_lane_s32(v, 0) + vgetq_lane_s32(v, 1) + vgetq_lane_s32(v, 2) + vgetq_lane_s32(v, 3);
+}
+
+inline static float vaddvq_f32(float32x4_t v) {
+    return vgetq_lane_f32(v, 0) + vgetq_lane_f32(v, 1) + vgetq_lane_f32(v, 2) + vgetq_lane_f32(v, 3);
+}
+
+inline static float vmaxvq_f32(float32x4_t v) {
+    return
+        MAX(MAX(vgetq_lane_f32(v, 0), vgetq_lane_f32(v, 1)),
+            MAX(vgetq_lane_f32(v, 2), vgetq_lane_f32(v, 3)));
+}
+
+inline static int32x4_t vcvtnq_s32_f32(float32x4_t v) {
+    int32x4_t res;
+
+    res[0] = roundf(vgetq_lane_f32(v, 0));
+    res[1] = roundf(vgetq_lane_f32(v, 1));
+    res[2] = roundf(vgetq_lane_f32(v, 2));
+    res[3] = roundf(vgetq_lane_f32(v, 3));
+
+    return res;
+}
+
+inline static uint8x8_t vzip1_u8(uint8x8_t a, uint8x8_t b) {
+    uint8x8_t res;
+
+    res[0] = a[0]; res[1] = b[0];
+    res[2] = a[1]; res[3] = b[1];
+    res[4] = a[2]; res[5] = b[2];
+    res[6] = a[3]; res[7] = b[3];
+
+    return res;
+}
+
+inline static uint8x8_t vzip2_u8(uint8x8_t a, uint8x8_t b) {
+    uint8x8_t res;
+
+    res[0] = a[4]; res[1] = b[4];
+    res[2] = a[5]; res[3] = b[5];
+    res[4] = a[6]; res[5] = b[6];
+    res[6] = a[7]; res[7] = b[7];
+
+    return res;
+}
+
+// vld1q_s16_x2
+// vld1q_u8_x2
+// vld1q_u8_x4
+// vld1q_s8_x2
+// vld1q_s8_x4
+// TODO: double-check these work correctly
+
+typedef struct ggml_int16x8x2_t {
+    int16x8_t val[2];
+} ggml_int16x8x2_t;
+
+inline static ggml_int16x8x2_t ggml_vld1q_s16_x2(const int16_t * ptr) {
+    ggml_int16x8x2_t res;
+
+    res.val[0] = vld1q_s16(ptr + 0);
+    res.val[1] = vld1q_s16(ptr + 8);
+
+    return res;
+}
+
+typedef struct ggml_uint8x16x2_t {
+    uint8x16_t val[2];
+} ggml_uint8x16x2_t;
+
+inline static ggml_uint8x16x2_t ggml_vld1q_u8_x2(const uint8_t * ptr) {
+    ggml_uint8x16x2_t res;
+
+    res.val[0] = vld1q_u8(ptr + 0);
+    res.val[1] = vld1q_u8(ptr + 16);
+
+    return res;
+}
+
+typedef struct ggml_uint8x16x4_t {
+    uint8x16_t val[4];
+} ggml_uint8x16x4_t;
+
+inline static ggml_uint8x16x4_t ggml_vld1q_u8_x4(const uint8_t * ptr) {
+    ggml_uint8x16x4_t res;
+
+    res.val[0] = vld1q_u8(ptr + 0);
+    res.val[1] = vld1q_u8(ptr + 16);
+    res.val[2] = vld1q_u8(ptr + 32);
+    res.val[3] = vld1q_u8(ptr + 48);
+
+    return res;
+}
+
+typedef struct ggml_int8x16x2_t {
+    int8x16_t val[2];
+} ggml_int8x16x2_t;
+
+inline static ggml_int8x16x2_t ggml_vld1q_s8_x2(const int8_t * ptr) {
+    ggml_int8x16x2_t res;
+
+    res.val[0] = vld1q_s8(ptr + 0);
+    res.val[1] = vld1q_s8(ptr + 16);
+
+    return res;
+}
+
+typedef struct ggml_int8x16x4_t {
+    int8x16_t val[4];
+} ggml_int8x16x4_t;
+
+inline static ggml_int8x16x4_t ggml_vld1q_s8_x4(const int8_t * ptr) {
+    ggml_int8x16x4_t res;
+
+    res.val[0] = vld1q_s8(ptr + 0);
+    res.val[1] = vld1q_s8(ptr + 16);
+    res.val[2] = vld1q_s8(ptr + 32);
+    res.val[3] = vld1q_s8(ptr + 48);
+
+    return res;
+}
+
+// NOTE: not tested
+inline static int8x16_t ggml_vqtbl1q_s8(int8x16_t a, uint8x16_t b) {
+    int8x16_t res;
+
+    res[ 0] = a[b[ 0]];
+    res[ 1] = a[b[ 1]];
+    res[ 2] = a[b[ 2]];
+    res[ 3] = a[b[ 3]];
+    res[ 4] = a[b[ 4]];
+    res[ 5] = a[b[ 5]];
+    res[ 6] = a[b[ 6]];
+    res[ 7] = a[b[ 7]];
+    res[ 8] = a[b[ 8]];
+    res[ 9] = a[b[ 9]];
+    res[10] = a[b[10]];
+    res[11] = a[b[11]];
+    res[12] = a[b[12]];
+    res[13] = a[b[13]];
+    res[14] = a[b[14]];
+    res[15] = a[b[15]];
+
+    return res;
+}
+
+// NOTE: not tested
+inline static uint8x16_t ggml_vqtbl1q_u8(uint8x16_t a, uint8x16_t b) {
+    uint8x16_t res;
+
+    res[ 0] = a[b[ 0]];
+    res[ 1] = a[b[ 1]];
+    res[ 2] = a[b[ 2]];
+    res[ 3] = a[b[ 3]];
+    res[ 4] = a[b[ 4]];
+    res[ 5] = a[b[ 5]];
+    res[ 6] = a[b[ 6]];
+    res[ 7] = a[b[ 7]];
+    res[ 8] = a[b[ 8]];
+    res[ 9] = a[b[ 9]];
+    res[10] = a[b[10]];
+    res[11] = a[b[11]];
+    res[12] = a[b[12]];
+    res[13] = a[b[13]];
+    res[14] = a[b[14]];
+    res[15] = a[b[15]];
+
+    return res;
+}
+
+#else
+
+#define ggml_int16x8x2_t  int16x8x2_t
+#define ggml_uint8x16x2_t uint8x16x2_t
+#define ggml_uint8x16x4_t uint8x16x4_t
+#define ggml_int8x16x2_t  int8x16x2_t
+#define ggml_int8x16x4_t  int8x16x4_t
+
+#define ggml_vld1q_s16_x2 vld1q_s16_x2
+#define ggml_vld1q_u8_x2  vld1q_u8_x2
+#define ggml_vld1q_u8_x4  vld1q_u8_x4
+#define ggml_vld1q_s8_x2  vld1q_s8_x2
+#define ggml_vld1q_s8_x4  vld1q_s8_x4
+#define ggml_vqtbl1q_s8   vqtbl1q_s8
+#define ggml_vqtbl1q_u8   vqtbl1q_u8
+
+#endif // !defined(__aarch64__)
+
+#if !defined(__ARM_FEATURE_DOTPROD)
+
+inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b) {
+    const int16x8_t p0 = vmull_s8(vget_low_s8 (a), vget_low_s8 (b));
+    const int16x8_t p1 = vmull_s8(vget_high_s8(a), vget_high_s8(b));
+
+    return vaddq_s32(acc, vaddq_s32(vpaddlq_s16(p0), vpaddlq_s16(p1)));
+}
+
+#else
+
+#define ggml_vdotq_s32(a, b, c) vdotq_s32(a, b, c)
+
+#endif // !defined(__ARM_FEATURE_DOTPROD)
+
+#endif // defined(__ARM_NEON)
+
+#ifdef __wasm_simd128__
+#include <wasm_simd128.h>
+#else
+#ifdef __POWER9_VECTOR__
+#include <altivec.h>
+#undef bool
+#define bool _Bool
+#else
+#if defined(_MSC_VER) || defined(__MINGW32__)
+#include <intrin.h>
+#else
+#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) || defined(__SSE3__) || defined(__SSE__)
+#if !defined(__riscv)
+#include <immintrin.h>
+#endif
+#endif
+#endif
+#endif
+#endif
+
+#ifdef __riscv_v_intrinsic
+#include <riscv_vector.h>
+#endif
+
+#if defined(__loongarch64)
+#if defined(__loongarch_asx)
+#include <lasxintrin.h>
+#endif
+#if defined(__loongarch_sx)
+#include <lsxintrin.h>
+#endif
+#endif
+
+#if defined(__loongarch_asx)
+
+typedef union {
+    int32_t i;
+    float f;
+} ft_union;
+
+/* float type data load instructions */
+static __m128 __lsx_vreplfr2vr_s(float val) {
+    ft_union fi_tmpval = {.f = val};
+    return (__m128)__lsx_vreplgr2vr_w(fi_tmpval.i);
+}
+
+static __m256 __lasx_xvreplfr2vr_s(float val) {
+    ft_union fi_tmpval = {.f = val};
+    return (__m256)__lasx_xvreplgr2vr_w(fi_tmpval.i);
+}
+#endif
+
+#ifdef __cplusplus
+}
+#endif

ggml/src/ggml-cpu/ggml-cpu-quants.h

@@ -0,0 +1,63 @@
+#pragma once
+
+#define GGML_COMMON_DECL_C
+#include "ggml-common.h"
+
+#include "ggml.h"
+
+// GGML CPU internal header
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Quantization
+void quantize_row_q4_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q4_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q5_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q5_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+
+void quantize_row_q2_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q3_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q4_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q5_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q6_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+
+void quantize_row_tq1_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_tq2_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+
+void quantize_row_iq4_nl (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_iq4_xs (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+
+// Dot product
+void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
+void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
+void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
+void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq2_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq2_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq1_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq1_m_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq4_nl_q8_0 (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq4_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq3_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
+#ifdef __cplusplus
+}
+#endif

ggml/src/ggml-cpu/ggml-cpu.cpp

@@ -0,0 +1,575 @@
+#include "ggml-backend.h"
+#include "ggml-backend-impl.h"
+#include "ggml-cpu.h"
+#include "ggml-impl.h"
+#include <cctype>
+#include <string>
+#include <vector>
+
+#if defined(__APPLE__)
+#include <sys/types.h>
+#include <sys/sysctl.h>
+#endif
+
+#if defined(_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
+    #define NOMINMAX
+#endif
+#include <windows.h>
+#endif
+
+// ggml-backend interface
+
+#ifdef GGML_USE_CPU_HBM
+
+// buffer type HBM
+
+#include <hbwmalloc.h>
+
+static const char * ggml_backend_cpu_hbm_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
+    return "CPU_HBM";
+
+    GGML_UNUSED(buft);
+}
+
+static void ggml_backend_cpu_hbm_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    hbw_free(buffer->context);
+}
+
+static ggml_backend_buffer_t ggml_backend_cpu_hbm_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    void * ptr;
+    int result = hbw_posix_memalign(&ptr, ggml_backend_cpu_buffer_type_get_alignment(buft), size);
+    if (result != 0) {
+        GGML_LOG_ERROR("failed to allocate HBM buffer of size %zu\n", size);
+        return NULL;
+    }
+
+    ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);
+    buffer->buft = buft;
+    buffer->iface.free_buffer = ggml_backend_cpu_hbm_buffer_free_buffer;
+
+    return buffer;
+}
+
+ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void) {
+    static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type_hbm = {
+        /* .iface    = */ {
+            /* .get_name         = */ ggml_backend_cpu_hbm_buffer_type_get_name,
+            /* .alloc_buffer     = */ ggml_backend_cpu_hbm_buffer_type_alloc_buffer,
+            /* .get_alignment    = */ ggml_backend_cpu_buffer_type_get_alignment,
+            /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
+            /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
+            /* .is_host          = */ ggml_backend_cpu_buffer_type_is_host,
+        },
+        /* .context  = */ NULL,
+    };
+
+    return &ggml_backend_cpu_buffer_type_hbm;
+}
+#endif
+
+static ggml_backend_buffer_type_t * ggml_backend_cpu_get_extra_bufts(ggml_backend_dev_t device) {
+    static ggml_backend_buffer_type_t bufts[] = {
+#ifdef GGML_USE_CPU_HBM
+        ggml_backend_cpu_hbm_buffer_type(),
+#endif
+        NULL
+    };
+
+    return bufts;
+
+    GGML_UNUSED(device);
+}
+
+// CPU backend - backend (stream)
+
+struct ggml_backend_cpu_context {
+    int                 n_threads;
+    ggml_threadpool_t   threadpool;
+
+    uint8_t *           work_data;
+    size_t              work_size;
+
+    ggml_abort_callback abort_callback;
+    void *              abort_callback_data;
+};
+
+static const char * ggml_backend_cpu_get_name(ggml_backend_t backend) {
+    return "CPU";
+
+    GGML_UNUSED(backend);
+}
+
+static void ggml_backend_cpu_free(ggml_backend_t backend) {
+    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
+    delete[] cpu_ctx->work_data;
+    delete cpu_ctx;
+    delete backend;
+}
+
+struct ggml_backend_plan_cpu {
+    struct ggml_cplan cplan;
+    struct ggml_cgraph cgraph;
+};
+
+static ggml_backend_graph_plan_t ggml_backend_cpu_graph_plan_create(ggml_backend_t backend, const struct ggml_cgraph * cgraph) {
+    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
+
+    struct ggml_backend_plan_cpu * cpu_plan = new ggml_backend_plan_cpu;
+
+    cpu_plan->cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads, cpu_ctx->threadpool);
+    cpu_plan->cgraph = *cgraph; // FIXME: deep copy
+
+    if (cpu_plan->cplan.work_size > 0) {
+        cpu_plan->cplan.work_data = new uint8_t[cpu_plan->cplan.work_size];
+        if (cpu_plan->cplan.work_data == NULL) {
+            delete cpu_plan;
+            return NULL;
+        }
+    }
+
+    cpu_plan->cplan.abort_callback      = cpu_ctx->abort_callback;
+    cpu_plan->cplan.abort_callback_data = cpu_ctx->abort_callback_data;
+
+    return cpu_plan;
+}
+
+static void ggml_backend_cpu_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
+    struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
+
+    delete[] cpu_plan->cplan.work_data;
+    delete cpu_plan;
+
+    GGML_UNUSED(backend);
+}
+
+static enum ggml_status ggml_backend_cpu_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
+    struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
+
+    return ggml_graph_compute(&cpu_plan->cgraph, &cpu_plan->cplan);
+
+    GGML_UNUSED(backend);
+}
+
+static enum ggml_status ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
+
+    struct ggml_cplan cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads, cpu_ctx->threadpool);
+
+    if (cpu_ctx->work_size < cplan.work_size) {
+        delete[] cpu_ctx->work_data;
+        cpu_ctx->work_data = new uint8_t[cplan.work_size];
+        if (cpu_ctx->work_data == NULL) {
+            cpu_ctx->work_size = 0;
+            return GGML_STATUS_ALLOC_FAILED;
+        }
+        cpu_ctx->work_size = cplan.work_size;
+    }
+    cplan.work_data = (uint8_t *)cpu_ctx->work_data;
+
+    cplan.abort_callback      = cpu_ctx->abort_callback;
+    cplan.abort_callback_data = cpu_ctx->abort_callback_data;
+
+    return ggml_graph_compute(cgraph, &cplan);
+}
+
+static const struct ggml_backend_i ggml_backend_cpu_i = {
+    /* .get_name                = */ ggml_backend_cpu_get_name,
+    /* .free                    = */ ggml_backend_cpu_free,
+    /* .set_tensor_async        = */ NULL,
+    /* .get_tensor_async        = */ NULL,
+    /* .cpy_tensor_async        = */ NULL,
+    /* .synchronize             = */ NULL,
+    /* .graph_plan_create       = */ ggml_backend_cpu_graph_plan_create,
+    /* .graph_plan_free         = */ ggml_backend_cpu_graph_plan_free,
+    /* .graph_plan_update       = */ NULL,
+    /* .graph_plan_compute      = */ ggml_backend_cpu_graph_plan_compute,
+    /* .graph_compute           = */ ggml_backend_cpu_graph_compute,
+    /* .event_record            = */ NULL,
+    /* .event_wait              = */ NULL,
+};
+
+static ggml_guid_t ggml_backend_cpu_guid(void) {
+    static ggml_guid guid = { 0xaa, 0x67, 0xc7, 0x43, 0x96, 0xe6, 0xa3, 0x8a, 0xe3, 0xaf, 0xea, 0x92, 0x36, 0xbc, 0xfc, 0x89 };
+    return &guid;
+}
+
+ggml_backend_t ggml_backend_cpu_init(void) {
+    // initialize CPU backend now to avoid slowing the first graph computation
+    ggml_cpu_init();
+
+    struct ggml_backend_cpu_context * ctx = new ggml_backend_cpu_context;
+    if (ctx == NULL) {
+        return NULL;
+    }
+
+    ctx->n_threads           = GGML_DEFAULT_N_THREADS;
+    ctx->threadpool          = NULL;
+    ctx->work_data           = NULL;
+    ctx->work_size           = 0;
+    ctx->abort_callback      = NULL;
+    ctx->abort_callback_data = NULL;
+
+    ggml_backend_t cpu_backend = new ggml_backend {
+        /* .guid      = */ ggml_backend_cpu_guid(),
+        /* .interface = */ ggml_backend_cpu_i,
+        /* .device    = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
+        /* .context   = */ ctx,
+    };
+
+    if (cpu_backend == NULL) {
+        delete ctx;
+        return NULL;
+    }
+
+    return cpu_backend;
+}
+
+bool ggml_backend_is_cpu(ggml_backend_t backend) {
+    return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_cpu_guid());
+}
+
+void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads) {
+    GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
+
+    struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
+    ctx->n_threads = n_threads;
+}
+
+void ggml_backend_cpu_set_threadpool(ggml_backend_t backend_cpu, ggml_threadpool_t threadpool) {
+    GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
+
+    struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
+
+    if (ctx->threadpool && ctx->threadpool != threadpool) {
+        // already had a different threadpool, pause/suspend it before switching
+        ggml_threadpool_pause(ctx->threadpool);
+    }
+    ctx->threadpool = threadpool;
+}
+
+void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data) {
+    GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
+
+    struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
+    ctx->abort_callback = abort_callback;
+    ctx->abort_callback_data = abort_callback_data;
+}
+
+// CPU backend - device
+
+struct ggml_backend_cpu_device_context {
+    std::string description = "CPU";
+
+    ggml_backend_cpu_device_context() {
+#ifdef __APPLE__
+        size_t len = 0;
+        if (!sysctlbyname("machdep.cpu.brand_string", NULL, &len, NULL, 0)) {
+            description.resize(len);
+            sysctlbyname("machdep.cpu.brand_string", &description[0], &len, NULL, 0); // NOLINT
+        }
+#elif defined(__linux__)
+        FILE * f = fopen("/proc/cpuinfo", "r");
+        if (f) {
+            char buf[1024];
+            while (fgets(buf, sizeof(buf), f)) {
+                if (strncmp(buf, "model name", 10) == 0) {
+                    char * p = strchr(buf, ':');
+                    if (p) {
+                        p++;
+                        while (std::isspace(*p)) {
+                            p++;
+                        }
+                        while (std::isspace(p[strlen(p) - 1])) {
+                            p[strlen(p) - 1] = '\0';
+                        }
+                        description = p;
+                        break;
+                    }
+                }
+            }
+            fclose(f);
+        }
+#elif defined(_WIN32)
+        HKEY hKey;
+        if (RegOpenKeyEx(HKEY_LOCAL_MACHINE,
+                        TEXT("HARDWARE\\DESCRIPTION\\System\\CentralProcessor\\0"),
+                        0,
+                        KEY_READ,
+                        &hKey) == ERROR_SUCCESS) {
+            DWORD cpu_brand_size = 0;
+            if (RegQueryValueExA(hKey,
+                                TEXT("ProcessorNameString"),
+                                NULL,
+                                NULL,
+                                NULL,
+                                &cpu_brand_size) == ERROR_SUCCESS) {
+                description.resize(cpu_brand_size);
+                if (RegQueryValueExA(hKey,
+                                    TEXT("ProcessorNameString"),
+                                    NULL,
+                                    NULL,
+                                    (LPBYTE)&description[0], // NOLINT
+                                    &cpu_brand_size) == ERROR_SUCCESS) {
+                    if (description.find('\0') != std::string::npos) {
+                        description.resize(description.find('\0'));
+                    }
+                }
+            }
+            RegCloseKey(hKey);
+        }
+#endif
+    }
+};
+
+static const char * ggml_backend_cpu_device_get_name(ggml_backend_dev_t dev) {
+    return "CPU";
+
+    GGML_UNUSED(dev);
+}
+
+static const char * ggml_backend_cpu_device_get_description(ggml_backend_dev_t dev) {
+    struct ggml_backend_cpu_device_context * ctx = (struct ggml_backend_cpu_device_context *)dev->context;
+
+    return ctx->description.c_str();
+}
+
+static void ggml_backend_cpu_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
+    // TODO
+    *free = 0;
+    *total = 0;
+
+    GGML_UNUSED(dev);
+}
+
+static enum ggml_backend_dev_type ggml_backend_cpu_device_get_type(ggml_backend_dev_t dev) {
+    return GGML_BACKEND_DEVICE_TYPE_CPU;
+
+    GGML_UNUSED(dev);
+}
+
+static void ggml_backend_cpu_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) {
+    props->name        = ggml_backend_cpu_device_get_name(dev);
+    props->description = ggml_backend_cpu_device_get_description(dev);
+    props->type        = ggml_backend_cpu_device_get_type(dev);
+    ggml_backend_cpu_device_get_memory(dev, &props->memory_free, &props->memory_total);
+    props->caps = {
+        /* .async                 = */ false,
+        /* .host_buffer           = */ false,
+        /* .buffer_from_host_ptr  = */ true,
+        /* .events                = */ false,
+    };
+}
+
+static ggml_backend_t ggml_backend_cpu_device_init_backend(ggml_backend_dev_t dev, const char * params) {
+    return ggml_backend_cpu_init();
+
+    GGML_UNUSED(dev);
+    GGML_UNUSED(params);
+}
+
+static ggml_backend_buffer_type_t ggml_backend_cpu_device_get_buffer_type(ggml_backend_dev_t dev) {
+    return ggml_backend_cpu_buffer_type();
+
+    GGML_UNUSED(dev);
+}
+
+static ggml_backend_buffer_t ggml_backend_cpu_device_buffer_from_host_ptr(ggml_backend_dev_t dev, void * ptr, size_t size, size_t max_tensor_size) {
+    return ggml_backend_cpu_buffer_from_ptr(ptr, size);
+
+    GGML_UNUSED(dev);
+    GGML_UNUSED(max_tensor_size);
+}
+
+static bool ggml_backend_cpu_device_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) {
+    switch (op->op) {
+        case GGML_OP_CPY:
+            return
+                op->type != GGML_TYPE_IQ2_XXS &&
+                op->type != GGML_TYPE_IQ2_XS  &&
+                op->type != GGML_TYPE_IQ1_S   &&
+                op->type != GGML_TYPE_IQ1_M; // missing type_traits.from_float
+        case GGML_OP_MUL_MAT:
+            return op->src[1]->type == GGML_TYPE_F32;// FIXME || op->src[1]->type == ggml_get_type_traits(op->src[0]->type)->vec_dot_type;
+        case GGML_OP_ROPE_BACK:
+            return op->src[2] == NULL && (op->op_params[2] & 4) == 0;
+        case GGML_OP_IM2COL_BACK:
+            return op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32;
+        case GGML_OP_OUT_PROD:
+            return (op->src[0]->type == GGML_TYPE_F32 || ggml_is_quantized(op->src[0]->type)) && op->src[1]->type == GGML_TYPE_F32;
+        default:
+            return true;
+    }
+
+    GGML_UNUSED(dev);
+}
+
+static bool ggml_backend_cpu_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
+    return ggml_backend_buft_is_host(buft);
+
+    GGML_UNUSED(dev);
+}
+
+static const struct ggml_backend_device_i ggml_backend_cpu_device_i = {
+    /* .get_name             = */ ggml_backend_cpu_device_get_name,
+    /* .get_description      = */ ggml_backend_cpu_device_get_description,
+    /* .get_memory           = */ ggml_backend_cpu_device_get_memory,
+    /* .get_type             = */ ggml_backend_cpu_device_get_type,
+    /* .get_props            = */ ggml_backend_cpu_device_get_props,
+    /* .init_backend         = */ ggml_backend_cpu_device_init_backend,
+    /* .get_buffer_type      = */ ggml_backend_cpu_device_get_buffer_type,
+    /* .get_host_buffer_type = */ NULL,
+    /* .buffer_from_host_ptr = */ ggml_backend_cpu_device_buffer_from_host_ptr,
+    /* .supports_op          = */ ggml_backend_cpu_device_supports_op,
+    /* .supports_buft        = */ ggml_backend_cpu_device_supports_buft,
+    /* .offload_op           = */ NULL,
+    /* .event_new            = */ NULL,
+    /* .event_free           = */ NULL,
+    /* .event_synchronize    = */ NULL,
+};
+
+// CPU backend - backend (reg)
+
+static const char * ggml_backend_cpu_reg_get_name(ggml_backend_reg_t reg) {
+    return "CPU";
+
+    GGML_UNUSED(reg);
+}
+
+static size_t ggml_backend_cpu_reg_get_device_count(ggml_backend_reg_t reg) {
+    return 1;
+
+    GGML_UNUSED(reg);
+}
+
+static ggml_backend_dev_t ggml_backend_cpu_reg_get_device(ggml_backend_reg_t reg, size_t index) {
+    GGML_ASSERT(index == 0);
+
+    static ggml_backend_cpu_device_context ctx;
+    static ggml_backend_device ggml_backend_cpu_device = {
+        /* .iface   = */ ggml_backend_cpu_device_i,
+        /* .reg     = */ reg,
+        /* .context = */ &ctx,
+    };
+
+    return &ggml_backend_cpu_device;
+}
+
+struct ggml_backend_feature {
+    const char * name;
+    const char * value;
+};
+
+// Not used yet
+// This is intended to replace the the ggml_cpu_has_* functions when loading the CPU backend dynamically,
+// and additionally to allow other backends to expose their own list of features that applications can query using the same API.
+static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t reg) {
+    static std::vector<ggml_backend_feature> features = []() {
+        std::vector<ggml_backend_feature> features;
+        if (ggml_cpu_has_sse3()) {
+            features.push_back({ "SSE3", "1" });
+        }
+        if (ggml_cpu_has_ssse3()) {
+            features.push_back({ "SSSE3", "1" });
+        }
+        if (ggml_cpu_has_avx()) {
+            features.push_back({ "AVX", "1" });
+        }
+        if (ggml_cpu_has_avx2()) {
+            features.push_back({ "AVX2", "1" });
+        }
+        if (ggml_cpu_has_f16c()) {
+            features.push_back({ "F16C", "1" });
+        }
+        if (ggml_cpu_has_fma()) {
+            features.push_back({ "FMA", "1" });
+        }
+        if (ggml_cpu_has_avx_vnni()) {
+            features.push_back({ "AVX_VNNI", "1" });
+        }
+        if (ggml_cpu_has_avx512()) {
+            features.push_back({ "AVX512", "1" });
+        }
+        if (ggml_cpu_has_avx512_vbmi()) {
+            features.push_back({ "AVX512_VBMI", "1" });
+        }
+        if (ggml_cpu_has_avx512_vnni()) {
+            features.push_back({ "AVX512_VNNI", "1" });
+        }
+        if (ggml_cpu_has_avx512_bf16()) {
+            features.push_back({ "AVX512_BF16", "1" });
+        }
+        if (ggml_cpu_has_amx_int8()) {
+            features.push_back({ "AMX_INT8", "1" });
+        }
+        if (ggml_cpu_has_neon()) {
+            features.push_back({ "NEON", "1" });
+        }
+        if (ggml_cpu_has_arm_fma()) {
+            features.push_back({ "ARM_FMA", "1" });
+        }
+        if (ggml_cpu_has_fp16_va()) {
+            features.push_back({ "FP16_VA", "1" });
+        }
+        if (ggml_cpu_has_matmul_int8()) {
+            features.push_back({ "MATMUL_INT8", "1" });
+        }
+        if (ggml_cpu_has_sve()) {
+            features.push_back({ "SVE", "1" });
+        }
+        if (ggml_cpu_get_sve_cnt() > 0) {
+            static std::string sve_cnt = std::to_string(ggml_cpu_get_sve_cnt());
+            features.push_back({ "SVE_CNT", sve_cnt.c_str() });
+        }
+        if (ggml_cpu_has_riscv_v()) {
+            features.push_back({ "RISCV_V", "1" });
+        }
+        if (ggml_cpu_has_vsx()) {
+            features.push_back({ "VSX", "1" });
+        }
+        if (ggml_cpu_has_wasm_simd()) {
+            features.push_back({ "WASM_SIMD", "1" });
+        }
+        if (ggml_cpu_has_llamafile()) {
+            features.push_back({ "LLAMAFILE", "1" });
+        }
+
+        features.push_back({ nullptr, nullptr });
+
+        return features;
+    }();
+
+    return features.data();
+
+    GGML_UNUSED(reg);
+}
+
+static void * ggml_backend_cpu_get_proc_address(ggml_backend_reg_t reg, const char * name) {
+    if (strcmp(name, "ggml_backend_set_n_threads") == 0) {
+        return (void *)ggml_backend_cpu_set_n_threads;
+    }
+    if (strcmp(name, "ggml_backend_dev_get_extra_bufts") == 0) {
+        return (void *)ggml_backend_cpu_get_extra_bufts;
+    }
+
+    return NULL;
+
+    GGML_UNUSED(reg);
+}
+
+static const struct ggml_backend_reg_i ggml_backend_cpu_reg_i = {
+    /* .get_name         = */ ggml_backend_cpu_reg_get_name,
+    /* .get_device_count = */ ggml_backend_cpu_reg_get_device_count,
+    /* .get_device       = */ ggml_backend_cpu_reg_get_device,
+    /* .get_proc_address = */ ggml_backend_cpu_get_proc_address,
+};
+
+ggml_backend_reg_t ggml_backend_cpu_reg(void) {
+    static struct ggml_backend_reg ggml_backend_cpu_reg = {
+        /* .iface   = */ ggml_backend_cpu_reg_i,
+        /* .context = */ NULL,
+    };
+
+    return &ggml_backend_cpu_reg;
+}

ggml/src/ggml-cpu/llamafile/sgemm.cpp

@@ -0,0 +1,1884 @@
+// Copyright 2024 Mozilla Foundation
+//
+// Permission is hereby granted, free of charge, to any person obtaining
+// a copy of this software and associated documentation files (the
+// "Software"), to deal in the Software without restriction, including
+// without limitation the rights to use, copy, modify, merge, publish,
+// distribute, sublicense, and/or sell copies of the Software, and to
+// permit persons to whom the Software is furnished to do so, subject to
+// the following conditions:
+//
+// The above copyright notice and this permission notice shall be
+// included in all copies or substantial portions of the Software.
+//
+// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+// EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+// MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+// NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+// BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+// ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+// CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+// SOFTWARE.
+
+//
+//                   _   _          ___ _      _   ___
+//                  | |_(_)_ _ _  _| _ ) |    /_\ / __|
+//                  |  _| | ' \ || | _ \ |__ / _ \\__ \.
+//                   \__|_|_||_\_, |___/____/_/ \_\___/
+//                             |__/
+//
+//                    BASIC LINEAR ALGEBRA SUBPROGRAMS
+//
+//
+// This file implements multithreaded CPU matrix multiplication for the
+// common contiguous use case C = Aᵀ * B. These kernels are designed to
+// have excellent performance[1] for matrices that fit in the CPU cache
+// without imposing any overhead such as cache filling or malloc calls.
+//
+// This implementation does not guarantee any upper bound with rounding
+// errors, which grow along with k. Our goal's to maximally exploit the
+// hardware for performance, and then use whatever resources remain for
+// improving numerical accuracy.
+//
+// [1] J. Tunney, ‘LLaMA Now Goes Faster on CPUs’, Mar. 2024. [Online].
+//     Available: https://justine.lol/matmul/. [Accessed: 29-Mar-2024].
+
+#if defined(__GNUC__)
+#pragma GCC diagnostic ignored "-Wpedantic"
+#pragma GCC diagnostic ignored "-Wignored-attributes"
+#endif
+
+#include "sgemm.h"
+#include "ggml-impl.h"
+#include "ggml-cpu-impl.h"
+#include "ggml-quants.h"
+
+#ifdef _MSC_VER
+#define NOINLINE __declspec(noinline)
+#else
+#define NOINLINE __attribute__((__noinline__))
+#endif
+
+#if defined(__ARM_NEON) || defined(__AVX512F__)
+#define VECTOR_REGISTERS 32
+#else
+#define VECTOR_REGISTERS 16
+#endif
+
+#define MM256_SET_M128I(a, b) _mm256_insertf128_si256(_mm256_castsi128_si256(b), (a), 1)
+
+namespace {
+
+inline float unhalf(ggml_fp16_t d) {
+    return GGML_FP16_TO_FP32(d);
+}
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+// VECTORIZED ARITHMETIC OPERATIONS
+
+#if defined(__SSE__) || defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
+inline __m128 add(__m128 x, __m128 y) { return _mm_add_ps(x, y); }
+inline __m128 sub(__m128 x, __m128 y) { return _mm_sub_ps(x, y); }
+inline __m128 mul(__m128 x, __m128 y) { return _mm_mul_ps(x, y); }
+#endif  // __SSE__
+
+#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
+inline __m256 add(__m256 x, __m256 y) { return _mm256_add_ps(x, y); }
+inline __m256 sub(__m256 x, __m256 y) { return _mm256_sub_ps(x, y); }
+inline __m256 mul(__m256 x, __m256 y) { return _mm256_mul_ps(x, y); }
+#endif // __AVX__
+
+#if defined(__AVX512F__)
+inline __m512 add(__m512 x, __m512 y) { return _mm512_add_ps(x, y); }
+inline __m512 sub(__m512 x, __m512 y) { return _mm512_sub_ps(x, y); }
+inline __m512 mul(__m512 x, __m512 y) { return _mm512_mul_ps(x, y); }
+#endif // __AVX512F__
+
+#if defined(__ARM_NEON)
+inline float32x4_t add(float32x4_t x, float32x4_t y) { return vaddq_f32(x, y); }
+inline float32x4_t sub(float32x4_t x, float32x4_t y) { return vsubq_f32(x, y); }
+inline float32x4_t mul(float32x4_t x, float32x4_t y) { return vmulq_f32(x, y); }
+#endif // __ARM_NEON
+
+#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
+inline float16x8_t add(float16x8_t x, float16x8_t y) { return vaddq_f16(x, y); }
+inline float16x8_t sub(float16x8_t x, float16x8_t y) { return vsubq_f16(x, y); }
+inline float16x8_t mul(float16x8_t x, float16x8_t y) { return vmulq_f16(x, y); }
+#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+
+#if defined(__MMA__)
+typedef vector unsigned char vec_t;
+typedef __vector_quad acc_t;
+#endif
+////////////////////////////////////////////////////////////////////////////////////////////////////
+// VECTORIZED FUSED MULTIPLY ADD
+
+/**
+ * Computes a * b + c.
+ */
+template <typename T, typename U>
+inline U madd(T a, T b, U c) {
+    return add(mul(a, b), c);
+}
+
+#if defined(__FMA__)
+#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
+template <>
+inline __m256 madd(__m256 a, __m256 b, __m256 c) {
+    return _mm256_fmadd_ps(a, b, c);
+}
+#endif
+#if defined(__AVX512F__)
+template <>
+inline __m512 madd(__m512 a, __m512 b, __m512 c) {
+    return _mm512_fmadd_ps(a, b, c);
+}
+#endif
+#endif
+
+#if defined(__ARM_FEATURE_FMA)
+template <>
+inline float32x4_t madd(float32x4_t a, float32x4_t b, float32x4_t c) {
+    return vfmaq_f32(c, b, a);
+}
+#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && !defined(_MSC_VER)
+template <>
+inline float16x8_t madd(float16x8_t a, float16x8_t b, float16x8_t c) {
+    return vfmaq_f16(c, b, a);
+}
+#endif
+#endif
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+// VECTORIZED HORIZONTAL SUM
+
+#if defined(__ARM_NEON)
+inline float hsum(float32x4_t x) {
+    return vaddvq_f32(x);
+}
+#endif // __ARM_NEON
+
+#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && !defined(_MSC_VER)
+inline float hsum(float16x8_t x) {
+    return vaddvq_f32(vaddq_f32(vcvt_f32_f16(vget_low_f16(x)),
+                                vcvt_f32_f16(vget_high_f16(x))));
+}
+#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+
+#if defined(__SSE__) || defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
+inline float hsum(__m128 x) {
+#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
+    x = _mm_add_ps(x, _mm_movehl_ps(x, x));
+    x = _mm_add_ss(x, _mm_movehdup_ps(x));
+#else
+    __m128 t;
+    t = _mm_shuffle_ps(x, x, _MM_SHUFFLE(2, 3, 0, 1));
+    x = _mm_add_ps(x, t);
+    t = _mm_movehl_ps(t, x);
+    x = _mm_add_ss(x, t);
+#endif
+    return _mm_cvtss_f32(x);
+}
+#endif
+
+#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
+inline float hsum(__m256 x) {
+    return hsum(_mm_add_ps(_mm256_extractf128_ps(x, 1),
+                           _mm256_castps256_ps128(x)));
+}
+#endif // __AVX__
+
+#if defined(__AVX512F__)
+inline float hsum(__m512 x) {
+    return _mm512_reduce_add_ps(x);
+}
+#endif // __AVX512F__
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+// VECTORIZED MEMORY LOADING
+
+template <typename T, typename U> T load(const U *);
+
+#if defined(__ARM_NEON)
+template <> inline float32x4_t load(const float *p) {
+    return vld1q_f32(p);
+}
+#if !defined(_MSC_VER)
+template <> inline float16x8_t load(const ggml_fp16_t *p) {
+    return vld1q_f16((const float16_t *)p);
+}
+template <> inline float32x4_t load(const ggml_fp16_t *p) {
+    return vcvt_f32_f16(vld1_f16((const float16_t *)p));
+}
+#endif // _MSC_VER
+#endif // __ARM_NEON
+
+#if defined(__SSE__) || defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
+template <> inline __m128 load(const float *p) {
+    return _mm_loadu_ps(p);
+}
+#endif  // __SSE__
+
+#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
+template <> inline __m256 load(const float *p) {
+    return _mm256_loadu_ps(p);
+}
+#endif // __AVX__
+
+#if defined(__F16C__)
+template <> inline __m256 load(const ggml_fp16_t *p) {
+    return _mm256_cvtph_ps(_mm_loadu_si128((const __m128i *)p));
+}
+#endif // __F16C__
+
+#if defined(__AVX512F__)
+template <> inline __m512 load(const float *p) {
+    return _mm512_loadu_ps(p);
+}
+template <> inline __m512 load(const ggml_fp16_t *p) {
+    return _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)p));
+}
+#endif // __AVX512F__
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+// CONSTANTS
+
+#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
+static const int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
+static const __m128i iq4nlt = _mm_loadu_si128((const __m128i *) kvalues_iq4nl);
+#endif
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+// FLOATING POINT MATRIX MULTIPLICATION
+
+template <int KN, typename D, typename V, typename TA, typename TB, typename TC>
+class tinyBLAS {
+  public:
+    tinyBLAS(int64_t k,
+             const TA *A, int64_t lda,
+             const TB *B, int64_t ldb,
+             TC *C, int64_t ldc,
+             int ith, int nth)
+        : A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) {
+    }
+
+    void matmul(int64_t m, int64_t n) {
+        mnpack(0, m, 0, n);
+    }
+
+  private:
+    NOINLINE void mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+        int64_t mc, nc, mp, np;
+        switch ((MIN(m - m0, 5) << 4) | MIN(n - n0, 5)) {
+#if VECTOR_REGISTERS == 32
+        case 0x55:
+            mc = 5;
+            nc = 5;
+            gemm<5, 5>(m0, m, n0, n);
+            break;
+        case 0x45:
+            mc = 4;
+            nc = 5;
+            gemm<4, 5>(m0, m, n0, n);
+            break;
+        case 0x54:
+            mc = 5;
+            nc = 4;
+            gemm<5, 4>(m0, m, n0, n);
+            break;
+        case 0x44:
+            mc = 4;
+            nc = 4;
+            gemm<4, 4>(m0, m, n0, n);
+            break;
+        case 0x53:
+            mc = 5;
+            nc = 3;
+            gemm<5, 3>(m0, m, n0, n);
+            break;
+        case 0x35:
+            mc = 3;
+            nc = 5;
+            gemm<3, 5>(m0, m, n0, n);
+            break;
+        case 0x43:
+            mc = 4;
+            nc = 3;
+            gemm<4, 3>(m0, m, n0, n);
+            break;
+#else
+        case 0x55:
+        case 0x54:
+        case 0x53:
+        case 0x45:
+        case 0x44:
+        case 0x43:
+            mc = 4;
+            nc = 3;
+            gemm<4, 3>(m0, m, n0, n);
+            break;
+        case 0x35:
+#endif
+        case 0x34:
+            mc = 3;
+            nc = 4;
+            gemm<3, 4>(m0, m, n0, n);
+            break;
+        case 0x52:
+            mc = 5;
+            nc = 2;
+            gemm<5, 2>(m0, m, n0, n);
+            break;
+        case 0x33:
+            mc = 3;
+            nc = 3;
+            gemm<3, 3>(m0, m, n0, n);
+            break;
+        case 0x25:
+            mc = 2;
+            nc = 5;
+            gemm<2, 5>(m0, m, n0, n);
+            break;
+        case 0x42:
+            mc = 4;
+            nc = 2;
+            gemm<4, 2>(m0, m, n0, n);
+            break;
+        case 0x24:
+            mc = 2;
+            nc = 4;
+            gemm<2, 4>(m0, m, n0, n);
+            break;
+        case 0x32:
+            mc = 3;
+            nc = 2;
+            gemm<3, 2>(m0, m, n0, n);
+            break;
+        case 0x23:
+            mc = 2;
+            nc = 3;
+            gemm<2, 3>(m0, m, n0, n);
+            break;
+        case 0x51:
+            mc = 5;
+            nc = 1;
+            gemm<5, 1>(m0, m, n0, n);
+            break;
+        case 0x41:
+            mc = 4;
+            nc = 1;
+            gemm<4, 1>(m0, m, n0, n);
+            break;
+        case 0x22:
+            mc = 2;
+            nc = 2;
+            gemm<2, 2>(m0, m, n0, n);
+            break;
+        case 0x15:
+            mc = 1;
+            nc = 5;
+            gemm<1, 5>(m0, m, n0, n);
+            break;
+        case 0x14:
+            mc = 1;
+            nc = 4;
+            gemm<1, 4>(m0, m, n0, n);
+            break;
+        case 0x31:
+            mc = 3;
+            nc = 1;
+            gemm<3, 1>(m0, m, n0, n);
+            break;
+        case 0x13:
+            mc = 1;
+            nc = 3;
+            gemm<1, 3>(m0, m, n0, n);
+            break;
+        case 0x21:
+            mc = 2;
+            nc = 1;
+            gemm<2, 1>(m0, m, n0, n);
+            break;
+        case 0x12:
+            mc = 1;
+            nc = 2;
+            gemm<1, 2>(m0, m, n0, n);
+            break;
+        case 0x11:
+            mc = 1;
+            nc = 1;
+            gemm<1, 1>(m0, m, n0, n);
+            break;
+        default:
+            return;
+        }
+        mp = m0 + (m - m0) / mc * mc;
+        np = n0 + (n - n0) / nc * nc;
+        mnpack(mp, m, n0, np);
+        mnpack(m0, m, np, n);
+    }
+
+    template <int RM, int RN>
+    NOINLINE void gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+        int64_t ytiles = (m - m0) / RM;
+        int64_t xtiles = (n - n0) / RN;
+        int64_t tiles = xtiles * ytiles;
+        int64_t duty = (tiles + nth - 1) / nth;
+        int64_t start = duty * ith;
+        int64_t end = start + duty;
+        if (end > tiles)
+            end = tiles;
+        for (int64_t job = start; job < end; ++job) {
+            int64_t ii = m0 + job / xtiles * RM;
+            int64_t jj = n0 + job % xtiles * RN;
+            D Cv[RN][RM] = {};
+            for (int64_t l = 0; l < k; l += KN)
+                for (int64_t j = 0; j < RN; ++j)
+                    for (int64_t i = 0; i < RM; ++i)
+                        Cv[j][i] = madd(load<V>(A + lda * (ii + i) + l),
+                                        load<V>(B + ldb * (jj + j) + l),
+                                        Cv[j][i]);
+            for (int64_t j = 0; j < RN; ++j)
+                for (int64_t i = 0; i < RM; ++i)
+                    C[ldc * (jj + j) + (ii + i)] = hsum(Cv[j][i]);
+        }
+    }
+
+    const TA *const A;
+    const TB *const B;
+    TC *const C;
+    const int64_t k;
+    const int64_t lda;
+    const int64_t ldb;
+    const int64_t ldc;
+    const int ith;
+    const int nth;
+};
+
+//////////////////////////////////////////////////////////////////////////////////////////
+// QUANT ZERO MATRIX MULTIPLICATION
+
+#if defined(__ARM_FEATURE_DOTPROD)
+template <typename TA>
+class tinyBLAS_Q0_ARM {
+  public:
+    tinyBLAS_Q0_ARM(int64_t k,
+                    const TA *A, int64_t lda,
+                    const block_q8_0 *B, int64_t ldb,
+                    float *C, int64_t ldc,
+                    int ith, int nth)
+        : A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) {
+    }
+
+    void matmul(int64_t m, int64_t n) {
+        mnpack(0, m, 0, n);
+    }
+
+  private:
+    NOINLINE void mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+        int64_t mc, nc, mp, np;
+        switch ((MIN(m - m0, 3) << 4) | MIN(n - n0, 3ll)) {
+        case 0x33:
+            mc = 3;
+            nc = 3;
+            gemm<3, 3>(m0, m, n0, n);
+            break;
+        case 0x32:
+            mc = 3;
+            nc = 2;
+            gemm<3, 2>(m0, m, n0, n);
+            break;
+        case 0x23:
+            mc = 2;
+            nc = 3;
+            gemm<2, 3>(m0, m, n0, n);
+            break;
+        case 0x22:
+            mc = 2;
+            nc = 2;
+            gemm<2, 2>(m0, m, n0, n);
+            break;
+        case 0x31:
+            mc = 3;
+            nc = 1;
+            gemm<3, 1>(m0, m, n0, n);
+            break;
+        case 0x13:
+            mc = 1;
+            nc = 3;
+            gemm<1, 3>(m0, m, n0, n);
+            break;
+        case 0x21:
+            mc = 2;
+            nc = 1;
+            gemm<2, 1>(m0, m, n0, n);
+            break;
+        case 0x12:
+            mc = 1;
+            nc = 2;
+            gemm<1, 2>(m0, m, n0, n);
+            break;
+        case 0x11:
+            mc = 1;
+            nc = 1;
+            gemm<1, 1>(m0, m, n0, n);
+            break;
+        default:
+            return;
+        }
+        mp = m0 + (m - m0) / mc * mc;
+        np = n0 + (n - n0) / nc * nc;
+        mnpack(mp, m, n0, np);
+        mnpack(m0, m, np, n);
+    }
+
+    template <int RM, int RN>
+    NOINLINE void gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+        int64_t ytiles = (m - m0) / RM;
+        int64_t xtiles = (n - n0) / RN;
+        int64_t tiles = xtiles * ytiles;
+        int64_t duty = (tiles + nth - 1) / nth;
+        int64_t start = duty * ith;
+        int64_t end = start + duty;
+        if (end > tiles)
+            end = tiles;
+        for (int64_t job = start; job < end; ++job) {
+            int64_t ii = m0 + job / xtiles * RM;
+            int64_t jj = n0 + job % xtiles * RN;
+            float32x4_t Cv[RN][RM] = {};
+            for (int64_t l = 0; l < k; ++l)
+                for (int64_t j = 0; j < RN; ++j)
+                    for (int64_t i = 0; i < RM; ++i)
+                        Cv[j][i] = vmlaq_n_f32(Cv[j][i],
+                                               vcvtq_f32_s32(vdotq_s32(
+                                                   vdotq_s32(vdupq_n_s32(0),
+                                                             load_lo(A + lda * (ii + i) + l),
+                                                             load_lo(B + ldb * (jj + j) + l)),
+                                                   load_hi(A + lda * (ii + i) + l),
+                                                   load_hi(B + ldb * (jj + j) + l))),
+                                               unhalf(A[lda * (ii + i) + l].d) *
+                                               unhalf(B[ldb * (jj + j) + l].d));
+            for (int64_t j = 0; j < RN; ++j)
+                for (int64_t i = 0; i < RM; ++i)
+                    C[ldc * (jj + j) + (ii + i)] = hsum(Cv[j][i]);
+        }
+    }
+
+    inline int8x16_t load_lo(const block_q8_0 *b) {
+        return vld1q_s8(b->qs);
+    }
+
+    inline int8x16_t load_hi(const block_q8_0 *b) {
+        return vld1q_s8(b->qs + 16);
+    }
+
+    inline int8x16_t load_lo(const block_q4_0 *b) {
+        return vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vld1q_u8(b->qs),
+                                                     vdupq_n_u8(0x0f))),
+                        vdupq_n_s8(0x8));
+    }
+
+    inline int8x16_t load_hi(const block_q4_0 *b) {
+        return vsubq_s8(vreinterpretq_s8_u8(vshrq_n_u8(vld1q_u8(b->qs), 4)),
+                        vdupq_n_s8(0x8));
+    }
+
+    const TA *const A;
+    const block_q8_0 *const B;
+    float *const C;
+    const int64_t k;
+    const int64_t lda;
+    const int64_t ldb;
+    const int64_t ldc;
+    const int ith;
+    const int nth;
+};
+#endif // __ARM_FEATURE_DOTPROD
+
+#if defined(__AVX2__) || defined(__AVX512F__) || defined(__AVX__)
+template <typename TA, typename TB, typename TC>
+class tinyBLAS_Q0_AVX {
+  public:
+    tinyBLAS_Q0_AVX(int64_t k,
+                    const TA *A, int64_t lda,
+                    const TB *B, int64_t ldb,
+                    TC *C, int64_t ldc,
+                    int ith, int nth)
+        : A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) {
+    }
+
+    void matmul(int64_t m, int64_t n) {
+        mnpack(0, m, 0, n);
+    }
+
+  private:
+    void mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+        int64_t mc, nc, mp, np;
+        switch ((MIN(m - m0, 4) << 4) | MIN(n - n0, 4)) {
+#if VECTOR_REGISTERS == 32
+        case 0x44:
+            mc = 4;
+            nc = 4;
+#if defined(__AVX2__) && defined(__F16C__)
+            gemm4xN<4>(m0, m, n0, n);
+#else
+            gemm<4, 4>(m0, m, n0, n);
+#endif
+            break;
+        case 0x43:
+            mc = 4;
+            nc = 3;
+#if defined(__AVX2__) && defined(__F16C__)
+            gemm4xN<3>(m0, m, n0, n);
+#else
+            gemm<4, 3>(m0, m, n0, n);
+#endif
+            break;
+        case 0x34:
+            mc = 3;
+            nc = 4;
+#if defined(__AVX2__) && defined(__F16C__)
+            gemmMx4<3>(m0, m, n0, n);
+#else
+            gemm<3, 4>(m0, m, n0, n);
+#endif
+            break;
+        case 0x33:
+            mc = 3;
+            nc = 3;
+            gemm<3, 3>(m0, m, n0, n);
+            break;
+        case 0x42:
+            mc = 4;
+            nc = 2;
+#if defined(__AVX2__) && defined(__F16C__)
+            gemm4xN<2>(m0, m, n0, n);
+#else
+            gemm<4, 2>(m0, m, n0, n);
+#endif
+            break;
+        case 0x24:
+            mc = 2;
+            nc = 4;
+#if defined(__AVX2__) && defined(__F16C__)
+            gemmMx4<2>(m0, m, n0, n);
+#else
+            gemm<2, 4>(m0, m, n0, n);
+#endif
+            break;
+#else
+        case 0x44:
+        case 0x43:
+        case 0x42:
+            mc = 4;
+            nc = 2;
+#if defined(__AVX2__) && defined(__F16C__)
+            gemm4xN<2>(m0, m, n0, n);
+#else
+            gemm<4, 2>(m0, m, n0, n);
+#endif
+            break;
+        case 0x34:
+        case 0x24:
+            mc = 2;
+            nc = 4;
+#if defined(__AVX2__) && defined(__F16C__)
+            gemmMx4<2>(m0, m, n0, n);
+#else
+            gemm<2, 4>(m0, m, n0, n);
+#endif
+            break;
+        case 0x33:
+#endif
+        case 0x32:
+            mc = 3;
+            nc = 2;
+            gemm<3, 2>(m0, m, n0, n);
+            break;
+        case 0x23:
+            mc = 2;
+            nc = 3;
+            gemm<2, 3>(m0, m, n0, n);
+            break;
+        case 0x41:
+            mc = 4;
+            nc = 1;
+#if defined(__AVX2__) && defined(__F16C__)
+            gemm4xN<1>(m0, m, n0, n);
+#else
+            gemm<4, 1>(m0, m, n0, n);
+#endif
+            break;
+        case 0x22:
+            mc = 2;
+            nc = 2;
+            gemm<2, 2>(m0, m, n0, n);
+            break;
+        case 0x14:
+            mc = 1;
+            nc = 4;
+#if defined(__AVX2__) && defined(__F16C__)
+            gemmMx4<1>(m0, m, n0, n);
+#else
+            gemm<1, 4>(m0, m, n0, n);
+#endif
+            break;
+        case 0x31:
+            mc = 3;
+            nc = 1;
+            gemm<3, 1>(m0, m, n0, n);
+            break;
+        case 0x13:
+            mc = 1;
+            nc = 3;
+            gemm<1, 3>(m0, m, n0, n);
+            break;
+        case 0x21:
+            mc = 2;
+            nc = 1;
+            gemm<2, 1>(m0, m, n0, n);
+            break;
+        case 0x12:
+            mc = 1;
+            nc = 2;
+            gemm<1, 2>(m0, m, n0, n);
+            break;
+        case 0x11:
+            mc = 1;
+            nc = 1;
+            gemm<1, 1>(m0, m, n0, n);
+            break;
+        default:
+            return;
+        }
+        mp = m0 + (m - m0) / mc * mc;
+        np = n0 + (n - n0) / nc * nc;
+        mnpack(mp, m, n0, np);
+        mnpack(m0, m, np, n);
+    }
+
+#if defined(__AVX2__) && defined(__F16C__)
+// Templated functions for gemm of dimensions 4xN
+    template <int RN>
+    NOINLINE void gemm4xN(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+        int64_t ytiles = (m - m0) / 4;
+        int64_t xtiles = (n - n0) / RN;
+        int64_t tiles = xtiles * ytiles;
+        int64_t duty = (tiles + nth - 1) / nth;
+        int64_t start = duty * ith;
+        int64_t end = start + duty;
+        if (end > tiles)
+            end = tiles;
+        for (int64_t job = start; job < end; ++job) {
+            int64_t ii = m0 + job / xtiles * 4;
+            int64_t jj = n0 + job % xtiles * RN;
+            __m256 Cv[RN][4] = {};
+            for (int64_t l = 0; l < k; ++l) {
+                uint64_t a_delta = ((uint64_t)A[lda * (ii + 3) + l].d << 48) | ((uint64_t)A[lda * (ii + 2) + l].d << 32) | ((uint64_t)A[lda * (ii + 1) + l].d << 16) | (A[lda * (ii + 0) + l].d);
+                // Convert delta values for four blocks to float values
+                __m128 da = _mm_cvtph_ps(_mm_set_epi64x(0, a_delta));
+                __m256i avec0 = load(A + lda * (ii + 0) + l);
+                __m256i avec1 = load(A + lda * (ii + 1) + l);
+                __m256i avec2 = load(A + lda * (ii + 2) + l);
+                __m256i avec3 = load(A + lda * (ii + 3) + l);
+                for (int64_t j = 0; j < RN; ++j) {
+                        __m128 db = _mm_set1_ps(unhalf(B[ldb * (jj + j) + l].d));
+                        // Computation of product of delta values for four blocks and replicate it across 256 bit lane
+                        __m256 dvec =  _mm256_castps128_ps256(_mm_mul_ps(da, db));
+                        dvec = _mm256_permute2f128_ps(dvec ,dvec, 0);
+                        // Computation of dot product and multiplication with appropriate delta value products
+                        Cv[j][0] = madd(_mm256_shuffle_ps(dvec, dvec, 0),
+                                    updot(_mm256_sign_epi8(avec0, avec0),
+                                          _mm256_sign_epi8(load(B + ldb * (jj + j) + l), avec0)),
+                                    Cv[j][0]);
+                        Cv[j][1] = madd(_mm256_shuffle_ps(dvec, dvec, 85),
+                                    updot(_mm256_sign_epi8(avec1, avec1),
+                                            _mm256_sign_epi8(load(B + ldb * (jj + j) + l), avec1)),
+                                    Cv[j][1]);
+                        Cv[j][2] = madd(_mm256_shuffle_ps(dvec, dvec, 170),
+                                    updot(_mm256_sign_epi8(avec2, avec2),
+                                            _mm256_sign_epi8(load(B + ldb * (jj + j) + l), avec2)),
+                                    Cv[j][2]);
+                        Cv[j][3] = madd(_mm256_shuffle_ps(dvec, dvec, 255),
+                                    updot(_mm256_sign_epi8(avec3, avec3),
+                                            _mm256_sign_epi8(load(B + ldb * (jj + j) + l), avec3)),
+                                    Cv[j][3]);
+                }
+            }
+
+            for (int64_t j = 0; j < RN; ++j)
+                for (int64_t i = 0; i < 4; ++i)
+                    C[ldc * (jj + j) + (ii + i)] = hsum(Cv[j][i]);
+        }
+    }
+
+    // Templated functions for gemm of dimensions Mx4
+    template <int RM>
+    NOINLINE void gemmMx4(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+        int64_t ytiles = (m - m0) / RM;
+        int64_t xtiles = (n - n0) / 4;
+        int64_t tiles = xtiles * ytiles;
+        int64_t duty = (tiles + nth - 1) / nth;
+        int64_t start = duty * ith;
+        int64_t end = start + duty;
+        if (end > tiles)
+            end = tiles;
+        for (int64_t job = start; job < end; ++job) {
+            int64_t ii = m0 + job / xtiles * RM;
+            int64_t jj = n0 + job % xtiles * 4;
+            __m256 Cv[4][RM] = {};
+            for (int64_t l = 0; l < k; ++l) {
+                uint64_t b_delta = ((uint64_t)B[ldb * (jj + 3) + l].d << 48) | ((uint64_t)B[ldb * (jj + 2) + l].d << 32) | ((uint64_t)B[ldb * (jj + 1) + l].d << 16) | (B[ldb * (jj + 0) + l].d);
+                // Convert delta values for four blocks to float values
+                __m128 db = _mm_cvtph_ps(_mm_set_epi64x(0, b_delta));
+                __m256i bvec0 = load(B + ldb * (jj + 0) + l);
+                __m256i bvec1 = load(B + ldb * (jj + 1) + l);
+                __m256i bvec2 = load(B + ldb * (jj + 2) + l);
+                __m256i bvec3 = load(B + ldb * (jj + 3) + l);
+                for (int64_t i = 0; i < RM; ++i) {
+                    __m128 da = _mm_set1_ps(unhalf((A[lda * (ii + i) + l].d)));
+                    // Computation of product of delta values for four blocks and replicate it across 256 bit lane
+                    __m256 dvec =  _mm256_castps128_ps256(_mm_mul_ps(da, db));
+                    dvec = _mm256_permute2f128_ps(dvec ,dvec, 0);
+                    // Computation of dot product and multiplication with appropriate delta value products
+                    Cv[0][i] = madd(_mm256_shuffle_ps(dvec, dvec, 0),
+                                    updot(_mm256_sign_epi8(load(A + lda * (ii + i) + l),
+                                                            load(A + lda * (ii + i) + l)),
+                                            _mm256_sign_epi8(bvec0, load(A + lda * (ii + i) + l))),
+                                    Cv[0][i]);
+                    Cv[1][i] = madd(_mm256_shuffle_ps(dvec, dvec, 85),
+                                    updot(_mm256_sign_epi8(load(A + lda * (ii + i) + l),
+                                                            load(A + lda * (ii + i) + l)),
+                                            _mm256_sign_epi8(bvec1, load(A + lda * (ii + i) + l))),
+                                    Cv[1][i]);
+                    Cv[2][i] = madd(_mm256_shuffle_ps(dvec, dvec, 170),
+                                    updot(_mm256_sign_epi8(load(A + lda * (ii + i) + l),
+                                                            load(A + lda * (ii + i) + l)),
+                                            _mm256_sign_epi8(bvec2, load(A + lda * (ii + i) + l))),
+                                    Cv[2][i]);
+                    Cv[3][i] = madd(_mm256_shuffle_ps(dvec, dvec, 255),
+                                    updot(_mm256_sign_epi8(load(A + lda * (ii + i) + l),
+                                                            load(A + lda * (ii + i) + l)),
+                                            _mm256_sign_epi8(bvec3, load(A + lda * (ii + i) + l))),
+                                    Cv[3][i]);
+                }
+            }
+            for (int64_t j = 0; j < 4; ++j)
+                for (int64_t i = 0; i < RM; ++i)
+                    C[ldc * (jj + j) + (ii + i)] = hsum(Cv[j][i]);
+        }
+    }
+#endif
+
+    template <int RM, int RN>
+    NOINLINE void gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+        int64_t ytiles = (m - m0) / RM;
+        int64_t xtiles = (n - n0) / RN;
+        int64_t tiles = xtiles * ytiles;
+        int64_t duty = (tiles + nth - 1) / nth;
+        int64_t start = duty * ith;
+        int64_t end = start + duty;
+        if (end > tiles)
+            end = tiles;
+        for (int64_t job = start; job < end; ++job) {
+            int64_t ii = m0 + job / xtiles * RM;
+            int64_t jj = n0 + job % xtiles * RN;
+            __m256 Cv[RN][RM] = {};
+            for (int64_t l = 0; l < k; ++l)
+                for (int64_t j = 0; j < RN; ++j)
+                    for (int64_t i = 0; i < RM; ++i) {
+#if defined(__AVX2__)
+                        __m256 udTmp = updot(_mm256_sign_epi8(load(A + lda * (ii + i) + l),
+                                                              load(A + lda * (ii + i) + l)),
+                                             _mm256_sign_epi8(load(B + ldb * (jj + j) + l),
+                                                              load(A + lda * (ii + i) + l)));
+#else
+                        __m128i ali0 = load0(A + lda * (ii + i) + l);
+                        __m128i ali1 = load1(A + lda * (ii + i) + l);
+                        __m128i blj0 = load0(B + ldb * (jj + j) + l);
+                        __m128i blj1 = load1(B + ldb * (jj + j) + l);
+
+                        __m128i sepAA0 = _mm_sign_epi8(ali0, ali0);
+                        __m128i sepAA1 = _mm_sign_epi8(ali1, ali1);
+                        __m128i sepBA0 = _mm_sign_epi8(blj0, ali0);
+                        __m128i sepBA1 = _mm_sign_epi8(blj1, ali1);
+
+                        // updot
+                        const __m128i oneFill = _mm_set1_epi16(1);
+                        __m128i mad0 = _mm_maddubs_epi16(sepAA0, sepBA0);
+                        __m128i mad1 = _mm_maddubs_epi16(sepAA1, sepBA1);
+                        __m256 udTmp = _mm256_cvtepi32_ps(MM256_SET_M128I(_mm_madd_epi16(oneFill, mad1), _mm_madd_epi16(oneFill, mad0)));
+#endif
+                        Cv[j][i] = madd(_mm256_set1_ps(unhalf(A[lda * (ii + i) + l].d) *
+                                                       unhalf(B[ldb * (jj + j) + l].d)),
+                                                       udTmp,
+                                                       Cv[j][i]);
+                    }
+            for (int64_t j = 0; j < RN; ++j)
+                for (int64_t i = 0; i < RM; ++i)
+                    C[ldc * (jj + j) + (ii + i)] = hsum(Cv[j][i]);
+        }
+    }
+
+    inline __m256i load(const block_q8_0 *b) {
+        return _mm256_loadu_si256((const __m256i *)b->qs);
+    }
+
+    inline __m128i load0(const block_q8_0 *b) {
+        return _mm_loadu_si128((const __m128i *)b->qs);
+    }
+
+    inline __m128i load1(const block_q8_0 *b) {
+        return _mm_loadu_si128(((const __m128i *)b->qs) + 1);
+    }
+
+    inline __m256i load(const block_q4_0 *b) {
+        return _mm256_sub_epi8(denibble(b->qs), _mm256_set1_epi8(8));
+    }
+
+    inline __m128i load0(const block_q4_0 *b) {
+        const __m128i x = _mm_loadu_si128((const __m128i *)(b->qs));
+        return _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), x), _mm_set1_epi8(8));
+    }
+
+    inline __m128i load1(const block_q4_0 *b) {
+        const __m128i x = _mm_loadu_si128((const __m128i *)(b->qs));
+        return _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), _mm_srli_epi16(x, 4)), _mm_set1_epi8(8));
+    }
+
+    inline __m256i load(const block_q5_0 *b) {
+        return _mm256_or_si256(denibble(b->qs), bittobyte(b->qh));
+    }
+
+    inline __m128i load0(const block_q5_0* b) {
+        const __m128i x = _mm_loadu_si128((const __m128i *)(b->qs));
+        uint32_t x32;
+        memcpy(&x32, b->qh, sizeof(uint32_t));
+        __m128i qxl = _mm_and_si128(_mm_set1_epi8(15), x);
+        __m128i bytesl = _mm_cmpeq_epi8(_mm_set1_epi64x(-1),
+                                        _mm_or_si128(_mm_set1_epi64x(0x7fbfdfeff7fbfdfe),
+                                                     _mm_shuffle_epi8(_mm_set1_epi32(x32),
+                                                                      _mm_set_epi64x(0x0101010101010101, 0x0000000000000000))));
+        bytesl = _mm_andnot_si128(bytesl, _mm_set1_epi8((char)0xF0));
+        return _mm_or_si128(qxl, bytesl);
+    }
+
+    inline __m128i load1(const block_q5_0* b) {
+        const __m128i x = _mm_loadu_si128((const __m128i *)(b->qs));
+        uint32_t x32;
+        memcpy(&x32, b->qh, sizeof(uint32_t));
+        __m128i qxh = _mm_and_si128(_mm_set1_epi8(15), _mm_srli_epi16(x, 4));
+        __m128i bytesh = _mm_cmpeq_epi8(_mm_set1_epi64x(-1),
+                                        _mm_or_si128(_mm_set1_epi64x(0x7fbfdfeff7fbfdfe),
+                                                     _mm_shuffle_epi8(_mm_set1_epi32(x32),
+                                                                      _mm_set_epi64x(0x0303030303030303, 0x0202020202020202))));
+        bytesh = _mm_andnot_si128(bytesh, _mm_set1_epi8((char)0xF0));
+        return _mm_or_si128(qxh, bytesh);
+    }
+
+    inline __m256i load(const block_iq4_nl *b) {
+        return MM256_SET_M128I(load1(b), load0(b));
+    }
+
+    inline __m128i load0(const block_iq4_nl *b) {
+        const __m128i x = _mm_loadu_si128((const __m128i *)(b->qs));
+        return _mm_shuffle_epi8(iq4nlt, _mm_and_si128(_mm_set1_epi8(15), x));
+    }
+
+    inline __m128i load1(const block_iq4_nl *b) {
+        const __m128i x = _mm_loadu_si128((const __m128i *)(b->qs));
+        return _mm_shuffle_epi8(iq4nlt, _mm_and_si128(_mm_set1_epi8(15), _mm_srli_epi16(x, 4)));
+    }
+
+    inline __m256 updot(__m256i u, __m256i s) {
+        __m256i res;
+#if defined(__AVXVNNI__) || (defined(__AVX512VNNI__) && defined(__AVX512VL__))
+        res = _mm256_dpbusd_epi32(_mm256_setzero_si256(), u, s);
+#else
+        res = _mm256_madd_epi16(_mm256_set1_epi16(1), _mm256_maddubs_epi16(u, s));
+#endif
+        return _mm256_cvtepi32_ps(res);
+    }
+
+    static inline __m256i denibble(const uint8_t *p) {
+        __m128i x = _mm_loadu_si128((const __m128i *)p);
+        return _mm256_and_si256(_mm256_set1_epi8(15),
+                                _mm256_insertf128_si256(_mm256_castsi128_si256(x),
+                                                        _mm_srli_epi16(x, 4), 1));
+    }
+
+    static inline __m256i bittobyte(const uint8_t *p) {
+        uint32_t x32;
+        memcpy(&x32, p, sizeof(uint32_t));
+        __m256i bytes = _mm256_cmpeq_epi8(_mm256_set1_epi64x(-1),
+                                          _mm256_or_si256(_mm256_set1_epi64x(0x7fbfdfeff7fbfdfe),
+                                                          _mm256_shuffle_epi8(_mm256_set1_epi32(x32),
+                                                                              _mm256_set_epi64x(0x0303030303030303, 0x0202020202020202,
+                                                                                                0x0101010101010101, 0x0000000000000000))));
+        return _mm256_andnot_si256(bytes, _mm256_set1_epi8((char)0xF0));
+    }
+
+    const TA *const A;
+    const TB *const B;
+    TC *const C;
+    const int64_t k;
+    const int64_t lda;
+    const int64_t ldb;
+    const int64_t ldc;
+    const int ith;
+    const int nth;
+};
+#endif // __AVX__
+
+//PPC Implementation
+#if defined(__MMA__)
+
+#define SAVE_ACC(ACC, ii, jj) \
+   __builtin_mma_disassemble_acc(vec_C, ACC); \
+   for (int I = 0; I < 4; I++) { \
+      for (int J = 0; J < 4; J++) { \
+         *((float*)(C+ii+((jj+J)*ldc)+I)) = *((float*)&vec_C[I]+J); \
+      } \
+   } \
+
+template <typename TA, typename TB, typename TC>
+class tinyBLAS_PPC {
+  public:
+    tinyBLAS_PPC(int64_t k,
+                const TA *A, int64_t lda,
+                const TB *B, int64_t ldb,
+                TC *C, int64_t ldc,
+                int ith, int nth)
+        : A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) {
+    }
+
+    void matmul(int64_t m, int64_t n) {
+       mnpack(0, m, 0, n);
+    }
+
+  private:
+
+    void (tinyBLAS_PPC::*kernel)(int64_t, int64_t);
+
+    void READ_BLOCK(const float* a, int64_t lda, int rows, int cols, float* vec) {
+        int64_t i, j;
+        float *aoffset = NULL, *boffset = NULL;
+        float *aoffset1 = NULL, *aoffset2 = NULL, *aoffset3 = NULL, *aoffset4 = NULL;
+        float *aoffset5 = NULL, *aoffset6 = NULL, *aoffset7 = NULL, *aoffset8 = NULL;
+
+        aoffset = const_cast<float*>(a);
+        boffset = vec;
+        j = (rows >> 3);
+        if (j > 0) {
+            do {
+                aoffset1 = aoffset;
+                aoffset2 = aoffset1 + lda;
+                aoffset3 = aoffset2 + lda;
+                aoffset4 = aoffset3 + lda;
+                aoffset5 = aoffset4 + lda;
+                aoffset6 = aoffset5 + lda;
+                aoffset7 = aoffset6 + lda;
+                aoffset8 = aoffset7 + lda;
+                aoffset += 8 * lda;
+                i = (cols >> 3);
+                if (i > 0) {
+                    __vector_pair C1, C2, C3, C4, C5, C6, C7, C8;
+                    vector float c1[2], c2[2], c3[2], c4[2], c5[2], c6[2], c7[2], c8[2];
+                    vector float t1, t2, t3, t4, t5, t6, t7, t8;
+                    do {
+                        C1 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset1);
+                        C2 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset2);
+                        C3 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset3);
+                        C4 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset4);
+                        C5 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset5);
+                        C6 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset6);
+                        C7 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset7);
+                        C8 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset8);
+                        __builtin_vsx_disassemble_pair(c1, &C1);
+                        __builtin_vsx_disassemble_pair(c2, &C2);
+                        __builtin_vsx_disassemble_pair(c3, &C3);
+                        __builtin_vsx_disassemble_pair(c4, &C4);
+                        __builtin_vsx_disassemble_pair(c5, &C5);
+                        __builtin_vsx_disassemble_pair(c6, &C6);
+                        __builtin_vsx_disassemble_pair(c7, &C7);
+                        __builtin_vsx_disassemble_pair(c8, &C8);
+
+                        t1 = vec_mergeh(c1[0], c2[0]);
+                        t2 = vec_mergeh(c3[0], c4[0]);
+                        t3 = vec_mergeh(c5[0], c6[0]);
+                        t4 = vec_mergeh(c7[0], c8[0]);
+                        t5 = vec_xxpermdi(t1, t2, 0);
+                        t6 = vec_xxpermdi(t3, t4, 0);
+                        t7 = vec_xxpermdi(t1, t2, 3);
+                        t8 = vec_xxpermdi(t3, t4, 3);
+                        vec_xst(t5, 0, boffset);
+                        vec_xst(t6, 0, boffset+4);
+                        vec_xst(t7, 0, boffset+8);
+                        vec_xst(t8, 0, boffset+12);
+
+                        t1 = vec_mergel(c1[0], c2[0]);
+                        t2 = vec_mergel(c3[0], c4[0]);
+                        t3 = vec_mergel(c5[0], c6[0]);
+                        t4 = vec_mergel(c7[0], c8[0]);
+                        t5 = vec_xxpermdi(t1, t2, 0);
+                        t6 = vec_xxpermdi(t3, t4, 0);
+                        t7 = vec_xxpermdi(t1, t2, 3);
+                        t8 = vec_xxpermdi(t3, t4, 3);
+                        vec_xst(t5, 0, boffset+16);
+                        vec_xst(t6, 0, boffset+20);
+                        vec_xst(t7, 0, boffset+24);
+                        vec_xst(t8, 0, boffset+28);
+
+                        t1 = vec_mergeh(c1[1], c2[1]);
+                        t2 = vec_mergeh(c3[1], c4[1]);
+                        t3 = vec_mergeh(c5[1], c6[1]);
+                        t4 = vec_mergeh(c7[1], c8[1]);
+                        t5 = vec_xxpermdi(t1, t2, 0);
+                        t6 = vec_xxpermdi(t3, t4, 0);
+                        t7 = vec_xxpermdi(t1, t2, 3);
+                        t8 = vec_xxpermdi(t3, t4, 3);
+                        vec_xst(t5, 0, boffset+32);
+                        vec_xst(t6, 0, boffset+36);
+                        vec_xst(t7, 0, boffset+40);
+                        vec_xst(t8, 0, boffset+44);
+
+                        t1 = vec_mergel(c1[1], c2[1]);
+                        t2 = vec_mergel(c3[1], c4[1]);
+                        t3 = vec_mergel(c5[1], c6[1]);
+                        t4 = vec_mergel(c7[1], c8[1]);
+                        t5 = vec_xxpermdi(t1, t2, 0);
+                        t6 = vec_xxpermdi(t3, t4, 0);
+                        t7 = vec_xxpermdi(t1, t2, 3);
+                        t8 = vec_xxpermdi(t3, t4, 3);
+                        vec_xst(t5, 0, boffset+48);
+                        vec_xst(t6, 0, boffset+52);
+                        vec_xst(t7, 0, boffset+56);
+                        vec_xst(t8, 0, boffset+60);
+
+                        aoffset1 += 8*lda;
+                        aoffset2 += 8*lda;
+                        aoffset3 += 8*lda;
+                        aoffset4 += 8*lda;
+                        boffset += 64;
+                        i--;
+                    } while(i > 0);
+                }
+                if (cols & 4) {
+                    vector float c1, c2, c3, c4, c5, c6, c7, c8;
+                    vector float t1, t2, t3, t4, t5, t6, t7, t8;
+                    c1 = vec_xl(0, aoffset1);
+                    c2 = vec_xl(0, aoffset2);
+                    c3 = vec_xl(0, aoffset3);
+                    c4 = vec_xl(0, aoffset4);
+                    c5 = vec_xl(0, aoffset5);
+                    c6 = vec_xl(0, aoffset6);
+                    c7 = vec_xl(0, aoffset7);
+                    c8 = vec_xl(0, aoffset8);
+
+                    t1 = vec_mergeh(c1, c2);
+                    t2 = vec_mergeh(c3, c4);
+                    t3 = vec_mergeh(c5, c6);
+                    t4 = vec_mergeh(c7, c8);
+                    t5 = vec_xxpermdi(t1, t2, 0);
+                    t6 = vec_xxpermdi(t3, t4, 0);
+                    t7 = vec_xxpermdi(t1, t2, 3);
+                    t8 = vec_xxpermdi(t3, t4, 3);
+                    vec_xst(t5, 0, boffset);
+                    vec_xst(t6, 0, boffset+4);
+                    vec_xst(t7, 0, boffset+8);
+                    vec_xst(t8, 0, boffset+12);
+
+                    t1 = vec_mergel(c1, c2);
+                    t2 = vec_mergel(c3, c4);
+                    t3 = vec_mergel(c5, c6);
+                    t4 = vec_mergel(c7, c8);
+                    t5 = vec_xxpermdi(t1, t2, 0);
+                    t6 = vec_xxpermdi(t3, t4, 0);
+                    t7 = vec_xxpermdi(t1, t2, 3);
+                    t8 = vec_xxpermdi(t3, t4, 3);
+                    vec_xst(t5, 0, boffset+16);
+                    vec_xst(t6, 0, boffset+20);
+                    vec_xst(t7, 0, boffset+24);
+                    vec_xst(t8, 0, boffset+28);
+                }
+            j--;
+            } while(j > 0);
+        }
+
+        if (rows & 4) {
+            aoffset1 = aoffset;
+            aoffset2 = aoffset1 + lda;
+            aoffset3 = aoffset2 + lda;
+            aoffset4 = aoffset3 + lda;
+            aoffset += 4 * lda;
+            i = (cols >> 3);
+            if (i > 0) {
+                __vector_pair C1, C2, C3, C4;
+                vector float c1[2], c2[2], c3[2], c4[2];
+                vector float t1, t2, t3, t4, t5, t6, t7, t8;
+                do {
+                    C1 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset1);
+                    C2 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset2);
+                    C3 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset3);
+                    C4 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset4);
+                    __builtin_vsx_disassemble_pair(c1, &C1);
+                    __builtin_vsx_disassemble_pair(c2, &C2);
+                    __builtin_vsx_disassemble_pair(c3, &C3);
+                    __builtin_vsx_disassemble_pair(c4, &C4);
+
+                    t1 = vec_mergeh(c1[0], c2[0]);
+                    t2 = vec_mergeh(c3[0], c4[0]);
+                    t3 = vec_mergel(c1[0], c2[0]);
+                    t4 = vec_mergel(c3[0], c4[0]);
+                    t5 = vec_xxpermdi(t1, t2, 0);
+                    t6 = vec_xxpermdi(t1, t2, 3);
+                    t7 = vec_xxpermdi(t3, t4, 0);
+                    t8 = vec_xxpermdi(t3, t4, 3);
+                    vec_xst(t5, 0, boffset);
+                    vec_xst(t6, 0, boffset+4);
+                    vec_xst(t7, 0, boffset+8);
+                    vec_xst(t8, 0, boffset+12);
+
+                    t1 = vec_mergeh(c1[1], c2[1]);
+                    t2 = vec_mergeh(c3[1], c4[1]);
+                    t3 = vec_mergel(c1[1], c2[1]);
+                    t4 = vec_mergel(c3[1], c4[1]);
+                    t5 = vec_xxpermdi(t1, t2, 0);
+                    t6 = vec_xxpermdi(t1, t2, 3);
+                    t7 = vec_xxpermdi(t3, t4, 0);
+                    t8 = vec_xxpermdi(t3, t4, 3);
+                    vec_xst(t5, 0, boffset+16);
+                    vec_xst(t6, 0, boffset+20);
+                    vec_xst(t7, 0, boffset+24);
+                    vec_xst(t8, 0, boffset+28);
+
+                    aoffset1 += 8*lda;
+                    aoffset2 += 8*lda;
+                    aoffset3 += 8*lda;
+                    aoffset4 += 8*lda;
+                    boffset += 32;
+                    i--;
+                } while(i > 0);
+            }
+
+            if (cols & 4) {
+                vector float c1, c2, c3, c4;
+                vector float t1, t2, t3, t4;
+                c1 = vec_xl(0, aoffset1);
+                c2 = vec_xl(0, aoffset2);
+                c3 = vec_xl(0, aoffset3);
+                c4 = vec_xl(0, aoffset4);
+
+                t1 = vec_mergeh(c1, c2);
+                t2 = vec_mergeh(c3, c4);
+                t3 = vec_xxpermdi(t1, t2, 0);
+                t4 = vec_xxpermdi(t1, t2, 3);
+                vec_xst(t3, 0, boffset);
+                vec_xst(t4, 0, boffset+4);
+
+                t1 = vec_mergel(c1, c2);
+                t2 = vec_mergel(c3, c4);
+                t3 = vec_xxpermdi(t1, t2, 0);
+                t4 = vec_xxpermdi(t1, t2, 3);
+                vec_xst(t3, 0, boffset+8);
+                vec_xst(t4, 0, boffset+12);
+            }
+        }
+        if (rows & 3) {
+            aoffset1 = aoffset;
+            aoffset2 = aoffset1 + lda;
+            aoffset3 = aoffset2 + lda;
+            if (cols & 4) {
+                vector float c1, c2, c3, c4 = {0};
+                vector float t1, t2, t3, t4;
+                c1 = vec_xl(0, aoffset1);
+                c2 = vec_xl(0, aoffset2);
+                c3 = vec_xl(0, aoffset3);
+
+                t1 = vec_mergeh(c1, c2);
+                t2 = vec_mergeh(c3, c4);
+                t3 = vec_xxpermdi(t1, t2, 0);
+                t4 = vec_xxpermdi(t1, t2, 3);
+                vec_xst(t3, 0, boffset);
+                vec_xst(t4, 0, boffset+4);
+
+                t1 = vec_mergel(c1, c2);
+                t2 = vec_mergel(c3, c4);
+                t3 = vec_xxpermdi(t1, t2, 0);
+                t4 = vec_xxpermdi(t1, t2, 3);
+                vec_xst(t3, 0, boffset+8);
+                vec_xst(t4, 0, boffset+12);
+            }
+        }
+    }
+
+    void KERNEL_4x4(int64_t ii, int64_t jj) {
+        vec_t vec_A[4], vec_B[4], vec_C[4];
+        acc_t acc_0;
+        __builtin_mma_xxsetaccz(&acc_0);
+        for (int l = 0; l < k; l+=4) {
+            READ_BLOCK(A+(ii*lda)+l, lda, 4, 4, (float*)vec_A);
+            READ_BLOCK(B+(jj*ldb)+l, ldb, 4, 4, (float*)vec_B);
+            __builtin_mma_xvf32gerpp(&acc_0, vec_A[0], vec_B[0]);
+            __builtin_mma_xvf32gerpp(&acc_0, vec_A[1], vec_B[1]);
+            __builtin_mma_xvf32gerpp(&acc_0, vec_A[2], vec_B[2]);
+            __builtin_mma_xvf32gerpp(&acc_0, vec_A[3], vec_B[3]);
+        }
+        SAVE_ACC(&acc_0, ii, jj);
+    }
+
+    void KERNEL_4x8(int64_t ii, int64_t jj) {
+        vec_t vec_A[4], vec_B[8], vec_C[4];
+        acc_t acc_0, acc_1;
+        __builtin_mma_xxsetaccz(&acc_0);
+        __builtin_mma_xxsetaccz(&acc_1);
+        for (int64_t l = 0; l < k; l+=4) {
+            READ_BLOCK(A+(ii*lda)+l, lda, 4, 4, (float*)vec_A);
+            READ_BLOCK(B+(jj*ldb)+l, ldb, 8, 4, (float*)vec_B);
+            __builtin_mma_xvf32gerpp(&acc_0, vec_A[0], (vec_t)vec_B[0]);
+            __builtin_mma_xvf32gerpp(&acc_1, vec_A[0], (vec_t)vec_B[1]);
+            __builtin_mma_xvf32gerpp(&acc_0, vec_A[1], (vec_t)vec_B[2]);
+            __builtin_mma_xvf32gerpp(&acc_1, vec_A[1], (vec_t)vec_B[3]);
+            __builtin_mma_xvf32gerpp(&acc_0, vec_A[2], (vec_t)vec_B[4]);
+            __builtin_mma_xvf32gerpp(&acc_1, vec_A[2], (vec_t)vec_B[5]);
+            __builtin_mma_xvf32gerpp(&acc_0, vec_A[3], (vec_t)vec_B[6]);
+            __builtin_mma_xvf32gerpp(&acc_1, vec_A[3], (vec_t)vec_B[7]);
+        }
+        SAVE_ACC(&acc_0, ii, jj);
+        SAVE_ACC(&acc_1, ii, jj+4);
+    }
+
+    void KERNEL_8x4(int64_t ii, int64_t jj) {
+        vec_t vec_A[8], vec_B[4], vec_C[4];
+        acc_t acc_0, acc_1;
+        __builtin_mma_xxsetaccz(&acc_0);
+        __builtin_mma_xxsetaccz(&acc_1);
+        for (int64_t l = 0; l < k; l+=4) {
+            READ_BLOCK(A+(ii*lda)+l, lda, 8, 4, (float*)vec_A);
+            READ_BLOCK(B+(jj*ldb)+l, ldb, 4, 4, (float*)vec_B);
+            __builtin_mma_xvf32gerpp(&acc_0, (vec_t)vec_A[0], vec_B[0]);
+            __builtin_mma_xvf32gerpp(&acc_1, (vec_t)vec_A[1], vec_B[0]);
+            __builtin_mma_xvf32gerpp(&acc_0, (vec_t)vec_A[2], vec_B[1]);
+            __builtin_mma_xvf32gerpp(&acc_1, (vec_t)vec_A[3], vec_B[1]);
+            __builtin_mma_xvf32gerpp(&acc_0, (vec_t)vec_A[4], vec_B[2]);
+            __builtin_mma_xvf32gerpp(&acc_1, (vec_t)vec_A[5], vec_B[2]);
+            __builtin_mma_xvf32gerpp(&acc_0, (vec_t)vec_A[6], vec_B[3]);
+            __builtin_mma_xvf32gerpp(&acc_1, (vec_t)vec_A[7], vec_B[3]);
+        }
+        SAVE_ACC(&acc_0, ii, jj);
+        SAVE_ACC(&acc_1, ii+4, jj);
+    }
+
+    void KERNEL_8x8(int64_t ii, int64_t jj) {
+        vec_t vec_A[16], vec_B[16], vec_C[4];
+        acc_t acc_0, acc_1, acc_2, acc_3;
+        __builtin_mma_xxsetaccz(&acc_0);
+        __builtin_mma_xxsetaccz(&acc_1);
+        __builtin_mma_xxsetaccz(&acc_2);
+        __builtin_mma_xxsetaccz(&acc_3);
+        for (int l = 0; l < k; l+=8) {
+            READ_BLOCK(A+(ii*lda)+l, lda, 8, 8, (float*)vec_A);
+            READ_BLOCK(B+(jj*ldb)+l, ldb, 8, 8, (float*)vec_B);
+            for(int x = 0; x < 16; x+=2) {
+                __builtin_mma_xvf32gerpp(&acc_0, (vec_t)vec_A[x], vec_B[x]);
+                __builtin_mma_xvf32gerpp(&acc_1, (vec_t)vec_A[x], vec_B[x+1]);
+                __builtin_mma_xvf32gerpp(&acc_2, (vec_t)vec_A[x+1], vec_B[x]);
+                __builtin_mma_xvf32gerpp(&acc_3, (vec_t)vec_A[x+1], vec_B[x+1]);
+            }
+        }
+        SAVE_ACC(&acc_0, ii, jj);
+        SAVE_ACC(&acc_1, ii, jj+4);
+        SAVE_ACC(&acc_2, ii+4, jj);
+        SAVE_ACC(&acc_3, ii+4, jj+4);
+    }
+
+    void mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+        int64_t mc, nc, mp, np;
+        int m_rem = MIN(m - m0, 16);
+        int n_rem = MIN(n - n0, 16);
+        if (m_rem >= 16 && n_rem >= 8) {
+            mc = 8;
+            nc = 8;
+            gemm<8,8>(m0, m, n0, n);
+        } else if(m_rem >= 8 && n_rem >= 16) {
+            mc = 8;
+            nc = 8;
+            gemm<8,8>(m0, m, n0, n);
+        } else if (m_rem >= 8 && n_rem >= 8) {
+            mc = 8;
+            nc = 8;
+            gemm<8,8>(m0, m, n0, n);
+        } else if (m_rem >= 4 && n_rem >= 8) {
+            mc = 4;
+            nc = 8;
+            gemm<4,8>(m0, m, n0, n);
+        } else if (m_rem >= 8 && n_rem >= 4) {
+            mc = 8;
+            nc = 4;
+            gemm<8,4>(m0, m, n0, n);
+        } else if (m_rem >= 4 && n_rem >= 4) {
+            mc = 4;
+            nc = 4;
+            gemm<4,4>(m0, m, n0, n);
+        } else if ((m_rem < 4) && (n_rem > 4)) {
+            nc = 4;
+            switch(m_rem) {
+                case 1:
+                    mc = 1;
+                    gemm_small(m0, m, n0, n, mc, nc);
+                    break;
+                case 2:
+                    mc = 2;
+                    gemm_small(m0, m, n0, n, mc, nc);
+                    break;
+                case 3:
+                    mc = 3;
+                    gemm_small(m0, m, n0, n, mc, nc);
+                    break;
+                default:
+                    return;
+            }
+        } else if ((m_rem > 4) && (n_rem < 4)) {
+            mc = 4;
+            switch(n_rem) {
+                case 1:
+                    nc = 1;
+                    gemm_small(m0, m, n0, n, mc, nc);
+                    break;
+                case 2:
+                    nc = 2;
+                    gemm_small(m0, m, n0, n, mc, nc);
+                    break;
+                case 3:
+                    nc = 3;
+                    gemm_small(m0, m, n0, n, mc, nc);
+                    break;
+                default:
+                    return;
+            }
+        } else {
+            switch((m_rem << 4) | n_rem) {
+                case 0x43:
+                    mc = 4;
+                    nc = 3;
+                    gemm_small(m0, m, n0, n, mc, nc);
+                    break;
+                case 0x42:
+                    mc = 4;
+                    nc = 2;
+                    gemm_small(m0, m, n0, n, mc, nc);
+                    break;
+                case 0x41:
+                    mc = 4;
+                    nc = 1;
+                    gemm_small(m0, m, n0, n, mc, nc);
+                    break;
+                case 0x34:
+                    mc = 3;
+                    nc = 4;
+                    gemm_small(m0, m, n0, n, mc, nc);
+                    break;
+                case 0x33:
+                    mc = 3;
+                    nc = 3;
+                    gemm_small(m0, m, n0, n, mc, nc);
+                    break;
+                case 0x32:
+                    mc = 3;
+                    nc = 2;
+                    gemm_small(m0, m, n0, n, mc, nc);
+                    break;
+                case 0x31:
+                    mc = 3;
+                    nc = 1;
+                    gemm_small(m0, m, n0, n, mc, nc);
+                    break;
+                case 0x24:
+                    mc = 2;
+                    nc = 4;
+                    gemm_small(m0, m, n0, n, mc, nc);
+                    break;
+                case 0x23:
+                    mc = 2;
+                    nc = 3;
+                    gemm_small(m0, m, n0, n, mc, nc);
+                    break;
+                case 0x22:
+                    mc = 2;
+                    nc = 2;
+                    gemm_small(m0, m, n0, n, mc, nc);
+                    break;
+                case 0x21:
+                    mc = 2;
+                    nc = 1;
+                    gemm_small(m0, m, n0, n, mc, nc);
+                    break;
+                case 0x14:
+                    mc = 1;
+                    nc = 4;
+                    gemm_small(m0, m, n0, n, mc, nc);
+                    break;
+                case 0x13:
+                    mc = 1;
+                    nc = 3;
+                    gemm_small(m0, m, n0, n, mc, nc);
+                    break;
+                case 0x12:
+                    mc = 1;
+                    nc = 2;
+                    gemm_small(m0, m, n0, n, mc, nc);
+                    break;
+                case 0x11:
+                    mc = 1;
+                    nc = 1;
+                    gemm_small(m0, m, n0, n, mc, nc);
+                    break;
+                default:
+                    return;
+            }
+        }
+        mp = m0 + (m - m0) / mc * mc;
+        np = n0 + (n - n0) / nc * nc;
+        mnpack(mp, m, n0, np);
+        mnpack(m0, m, np, n);
+    }
+
+     void gemm_small(int64_t m0, int64_t m, int64_t n0, int64_t n, int RM, int RN) {
+        int64_t ytiles = (m - m0) / RM;
+        int64_t xtiles = (n - n0) / RN;
+        int64_t tiles = xtiles * ytiles;
+        int64_t duty = (tiles + nth - 1) / nth;
+        int64_t start = duty * ith;
+        int64_t end = start + duty;
+        if (end > tiles)
+            end = tiles;
+        for (int64_t job = start; job < end; ++job) {
+            int64_t ii = m0 + job / xtiles * RM;
+            int64_t jj = n0 + job % xtiles * RN;
+            vec_t vec_C[4];
+            acc_t acc_0;
+            __builtin_mma_xxsetaccz(&acc_0);
+            vec_t vec_A[4], vec_B[4];
+            for (int l=0; l<k; l+=4) {
+                if (RN >= 4 && RM == 1) {
+                    float* a = const_cast<float*>(A+(ii)*lda+l);
+                    READ_BLOCK(B+(jj*ldb)+l, ldb, 4, 4, (float*)vec_B);
+                    vec_A[0] = (vec_t)vec_xl(0,a);
+                    vec_A[1] = (vec_t)vec_splats(*((float*)&vec_A+1));
+                    vec_A[2] = (vec_t)vec_splats(*((float*)&vec_A+2));
+                    vec_A[3] = (vec_t)vec_splats(*((float*)&vec_A+3));
+                } else {
+                    READ_BLOCK(A+(ii*lda)+l, lda, RM, 4, (float*)vec_A);
+                    READ_BLOCK(B+(jj*ldb)+l, ldb, RN, 4, (float*)vec_B);
+                }
+                __builtin_mma_xvf32gerpp(&acc_0, vec_A[0], vec_B[0]);
+                __builtin_mma_xvf32gerpp(&acc_0, vec_A[1], vec_B[1]);
+                __builtin_mma_xvf32gerpp(&acc_0, vec_A[2], vec_B[2]);
+                __builtin_mma_xvf32gerpp(&acc_0, vec_A[3], vec_B[3]);
+            }
+            __builtin_mma_disassemble_acc(vec_C, &acc_0);
+            for (int I = 0; I < RM; I++) {
+                for (int J = 0; J < RN; J++) {
+                    *((float*)(C+ii+((jj+J)*ldc)+I)) = *((float*)&vec_C[I]+J);
+                }
+            }
+       }
+    }
+
+    template <int RM, int RN>
+    NOINLINE void gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+        int64_t ytiles = (m - m0) / RM;
+        int64_t xtiles = (n - n0) / RN;
+        int64_t tiles = xtiles * ytiles;
+        int64_t duty = (tiles + nth - 1) / nth;
+        int64_t start = duty * ith;
+        int64_t end = start + duty;
+        if (RM == 4 && RN == 4) {
+            kernel = &tinyBLAS_PPC::KERNEL_4x4;
+        } else if (RM == 4 && RN == 8) {
+            kernel = &tinyBLAS_PPC::KERNEL_4x8;
+        } else if (RM == 8 && RN == 4) {
+            kernel = &tinyBLAS_PPC::KERNEL_8x4;
+        } else if (RM == 8 && RN == 8) {
+            kernel = &tinyBLAS_PPC::KERNEL_8x8;
+        }
+        if (end > tiles)
+            end = tiles;
+        for (int64_t job = start; job < end; ++job) {
+            int64_t ii = m0 + job / xtiles * RM;
+            int64_t jj = n0 + job % xtiles * RN;
+            (this->*kernel)(ii, jj);
+        }
+    }
+
+    const TA *const A;
+    const TB *const B;
+    TC *C;
+    TA *At;
+    TB *Bt;
+    const int64_t k;
+    const int64_t lda;
+    const int64_t ldb;
+    const int64_t ldc;
+    const int ith;
+    const int nth;
+};
+#endif
+} // namespace
+
+/**
+ * Performs optimized matrix multiplication on CPU.
+ *
+ * This subroutine may compute C = Aᵀ * B with column major ordering.
+ * Despite its name, this isn't a generalized implementation. Work is
+ * only performed when a handwritten kernel is written and available.
+ * Otherwise the caller should fall back to a general matmul routine.
+ *
+ * For example, for single-threaded single-precision GEMM you can say
+ *
+ *     llamafile_sgemm(m, n, k, A, lda, B, ldb, C, ldc,
+ *                     0, 1,
+ *                     GGML_TYPE_F32, GGML_TYPE_F32, GGML_TYPE_F32);
+ *
+ * @param m is rows in `A` and `C`
+ * @param n is cols in `B` and `C`
+ * @param k is cols in `A` and rows in `B`
+ * @param A is first input matrix (always transposed)
+ * @param lda is row stride of `A`
+ * @param B is second input matrix (never transposed)
+ * @param ldb is row stride of `B`
+ * @param C is input/output array of output matrices
+ * @param ldc is row stride of `C`
+ * @param ith is thread id (must be less than `nth`)
+ * @param nth is number of threads (must be greater than zero)
+ * @param Atype is GGML data type of `A`
+ * @param Btype is GGML data type of `B`
+ * @param Ctype is GGML data type of `C`
+ * @return true if this function was able to service the matmul request
+ */
+bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda, const void *B, int64_t ldb, void *C,
+                     int64_t ldc, int ith, int nth, int Atype, int Btype, int Ctype) {
+
+    assert(m >= 0);
+    assert(n >= 0);
+    assert(k >= 0);
+    assert(lda >= k);
+    assert(ldb >= k);
+    assert(ldc >= m);
+    assert(nth > 0);
+    assert(ith < nth);
+
+    // only enable sgemm for prompt processing
+    if (n < 2)
+        return false;
+
+    if (Ctype != GGML_TYPE_F32)
+        return false;
+
+    switch (Atype) {
+
+    case GGML_TYPE_F32: {
+        if (Btype != GGML_TYPE_F32)
+            return false;
+#if defined(__AVX512F__)
+        if (k % 16)
+            return false;
+        tinyBLAS<16, __m512, __m512, float, float, float> tb{
+            k, (const float *)A, lda,
+            (const float *)B, ldb,
+            (float *)C, ldc,
+            ith, nth};
+        tb.matmul(m, n);
+        return true;
+#elif defined(__AVX__) || defined(__AVX2__)
+        if (k % 8)
+            return false;
+        tinyBLAS<8, __m256, __m256, float, float, float> tb{
+            k, (const float *)A, lda,
+            (const float *)B, ldb,
+            (float *)C, ldc,
+            ith, nth};
+        tb.matmul(m, n);
+        return true;
+#elif defined(__ARM_NEON)
+        if (n < 4)
+            return false;
+        if (k % 4)
+            return false;
+        tinyBLAS<4, float32x4_t, float32x4_t, float, float, float> tb{
+            k, (const float *)A, lda,
+            (const float *)B, ldb,
+            (float *)C, ldc,
+            ith, nth};
+        tb.matmul(m, n);
+        return true;
+#elif defined(__MMA__)
+        if (k % 8)
+            return false;
+        tinyBLAS_PPC<float, float, float> tb{
+            k, (const float *)A, lda,
+            (const float *)B, ldb,
+            (float *)C, ldc,
+            ith, nth};
+        tb.matmul(m, n);
+        return true;
+#else
+        return false;
+#endif
+    }
+
+    case GGML_TYPE_F16: {
+#if defined(__AVX512F__)
+        if (k % 16)
+            return false;
+        if (Btype != GGML_TYPE_F32)
+            return false;
+        tinyBLAS<16, __m512, __m512, ggml_fp16_t, float, float> tb{
+            k, (const ggml_fp16_t *)A, lda,
+            (const float *)B, ldb,
+            (float *)C, ldc,
+            ith, nth};
+        tb.matmul(m, n);
+        return true;
+#elif (defined(__AVX__) || defined(__AVX2__)) && defined(__F16C__)
+        if (k % 8)
+            return false;
+        if (Btype != GGML_TYPE_F32)
+            return false;
+        tinyBLAS<8, __m256, __m256, ggml_fp16_t, float, float> tb{
+            k, (const ggml_fp16_t *)A, lda,
+            (const float *)B, ldb,
+            (float *)C, ldc,
+            ith, nth};
+        tb.matmul(m, n);
+        return true;
+#elif defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && !defined(_MSC_VER)
+        if (n < 8)
+            return false;
+        if (k % 8)
+            return false;
+        if (Btype != GGML_TYPE_F16)
+            return false;
+        tinyBLAS<8, float16x8_t, float16x8_t, ggml_fp16_t, ggml_fp16_t, float> tb{
+            k, (const ggml_fp16_t *)A, lda,
+            (const ggml_fp16_t *)B, ldb,
+            (float *)C, ldc,
+            ith, nth};
+        tb.matmul(m, n);
+        return true;
+#elif defined(__ARM_NEON) && !defined(_MSC_VER)
+        if (k % 4)
+            return false;
+        if (Btype != GGML_TYPE_F32)
+            return false;
+        tinyBLAS<4, float32x4_t, float32x4_t, ggml_fp16_t, float, float> tb{
+            k, (const ggml_fp16_t *)A, lda,
+            (const float *)B, ldb,
+            (float *)C, ldc,
+            ith, nth};
+        tb.matmul(m, n);
+        return true;
+#else
+        return false;
+#endif
+    }
+
+    case GGML_TYPE_Q8_0: {
+        if (Btype != GGML_TYPE_Q8_0)
+           return false;
+#if defined(__AVX2__) || defined(__AVX512F__) || defined(__AVX__)
+        tinyBLAS_Q0_AVX<block_q8_0, block_q8_0, float> tb{
+            k, (const block_q8_0 *)A, lda,
+            (const block_q8_0 *)B, ldb,
+            (float *)C, ldc,
+            ith, nth};
+        tb.matmul(m, n);
+        return true;
+#elif defined(__ARM_FEATURE_DOTPROD)
+        tinyBLAS_Q0_ARM<block_q8_0> tb{
+            k, (const block_q8_0 *)A, lda,
+            (const block_q8_0 *)B, ldb,
+            (float *)C, ldc,
+            ith, nth};
+        tb.matmul(m, n);
+        return true;
+#else
+        return false;
+#endif
+    }
+
+    case GGML_TYPE_Q4_0: {
+        if (Btype != GGML_TYPE_Q8_0)
+            return false;
+#if defined(__AVX2__) || defined(__AVX512F__) || defined(__AVX__)
+        tinyBLAS_Q0_AVX<block_q4_0, block_q8_0, float> tb{
+            k, (const block_q4_0 *)A, lda,
+            (const block_q8_0 *)B, ldb,
+            (float *)C, ldc,
+            ith, nth};
+        tb.matmul(m, n);
+        return true;
+#elif defined(__ARM_FEATURE_DOTPROD)
+        tinyBLAS_Q0_ARM<block_q4_0> tb{
+            k, (const block_q4_0 *)A, lda,
+            (const block_q8_0 *)B, ldb,
+            (float *)C, ldc,
+            ith, nth};
+        tb.matmul(m, n);
+        return true;
+#else
+        return false;
+#endif
+    }
+
+    case GGML_TYPE_Q5_0: {
+        if (Btype != GGML_TYPE_Q8_0)
+            return false;
+#if defined(__AVX2__) || defined(__AVX512F__) || defined(__AVX__)
+        tinyBLAS_Q0_AVX<block_q5_0, block_q8_0, float> tb{
+            k, (const block_q5_0 *)A, lda,
+            (const block_q8_0 *)B, ldb,
+            (float *)C, ldc,
+            ith, nth};
+        tb.matmul(m, n);
+        return true;
+#else
+        return false;
+#endif
+    }
+
+    case GGML_TYPE_IQ4_NL: {
+        if (Btype != GGML_TYPE_Q8_0)
+            return false;
+#if defined(__AVX2__) || defined(__AVX512F__) || defined(__AVX__)
+        tinyBLAS_Q0_AVX<block_iq4_nl, block_q8_0, float> tb{
+            k, (const block_iq4_nl *)A, lda,
+            (const block_q8_0 *)B, ldb,
+            (float *)C, ldc,
+            ith, nth};
+        tb.matmul(m, n);
+        return true;
+#else
+        return false;
+#endif
+    }
+
+    default:
+        return false;
+    }
+
+    (void)m;
+    (void)n;
+    (void)k;
+    (void)A;
+    (void)lda;
+    (void)B;
+    (void)ldb;
+    (void)C;
+    (void)ldc;
+    (void)ith;
+    (void)nth;
+    (void)Atype;
+    (void)Btype;
+    (void)Ctype;
+}

ggml/src/ggml-cpu/llamafile/sgemm.h

@@ -0,0 +1,14 @@
+#pragma once
+#include <stdint.h>
+#include <stdbool.h>
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+bool llamafile_sgemm(int64_t, int64_t, int64_t, const void *, int64_t,
+                     const void *, int64_t, void *, int64_t, int, int,
+                     int, int, int);
+
+#ifdef __cplusplus
+}
+#endif

ggml/src/ggml-cuda/common.cuh

@@ -6,7 +6,7 @@
 #include <cstdint>
 #include <memory>
 
-#if defined(GGML_USE_HIPBLAS)
+#if defined(GGML_USE_HIP)
 #define GGML_COMMON_DECL_HIP
 #define GGML_COMMON_IMPL_HIP
 #else
@@ -26,13 +26,13 @@
 #include <string>
 #include <vector>
 
-#if defined(GGML_USE_HIPBLAS)
+#if defined(GGML_USE_HIP)
 #include "vendors/hip.h"
 #elif defined(GGML_USE_MUSA)
 #include "vendors/musa.h"
 #else
 #include "vendors/cuda.h"
-#endif // defined(GGML_USE_HIPBLAS)
+#endif // defined(GGML_USE_HIP)
 
 #define STRINGIZE_IMPL(...) #__VA_ARGS__
 #define STRINGIZE(...) STRINGIZE_IMPL(__VA_ARGS__)
@@ -97,7 +97,7 @@ void ggml_cuda_error(const char * stmt, const char * func, const char * file, in
 
 #define CUBLAS_CHECK(err) CUDA_CHECK_GEN(err, CUBLAS_STATUS_SUCCESS, cublas_get_error_str)
 
-#if !defined(GGML_USE_HIPBLAS)
+#if !defined(GGML_USE_HIP)
 static const char * cu_get_error_str(CUresult err) {
     const char * err_str;
     cuGetErrorString(err, &err_str);
@@ -120,21 +120,21 @@ typedef float dfloat; // dequantize float
 typedef float2 dfloat2;
 #endif // GGML_CUDA_F16
 
-#if (defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= CC_PASCAL
+#if (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= CC_PASCAL
 #define FP16_AVAILABLE
-#endif // (defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= CC_PASCAL
+#endif // (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= CC_PASCAL
 
 #if defined(FP16_AVAILABLE) && __CUDA_ARCH__ != 610
 #define FAST_FP16_AVAILABLE
 #endif // defined(FP16_AVAILABLE) && __CUDA_ARCH__ != 610
 
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_VOLTA
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_VOLTA
 #define FP16_MMA_AVAILABLE
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_VOLTA
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_VOLTA
 
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_TURING
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_TURING
 #define INT8_MMA_AVAILABLE
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_TURING
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_TURING
 
 #if !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ <= CC_QY1)
 #define FLASH_ATTN_AVAILABLE
@@ -156,14 +156,14 @@ static constexpr bool int8_mma_available(const int cc) {
 static __device__ void no_device_code(
     const char * file_name, const int line, const char * function_name, const int arch, const char * arch_list) {
 
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
     printf("%s:%d: ERROR: HIP kernel %s has no device code compatible with HIP arch %d.\n",
            file_name, line, function_name, arch);
     GGML_UNUSED(arch_list);
 #else
     printf("%s:%d: ERROR: CUDA kernel %s has no device code compatible with CUDA arch %d. ggml-cuda.cu was compiled for: %s\n",
            file_name, line, function_name, arch, arch_list);
-#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
     __trap();
 
     GGML_UNUSED(no_device_code); // suppress unused function warning
@@ -176,7 +176,7 @@ static __device__ void no_device_code(
 #endif // __CUDA_ARCH__
 
 static __device__ __forceinline__ int warp_reduce_sum(int x) {
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_AMPERE
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_AMPERE
     return __reduce_add_sync(0xffffffff, x);
 #else
 #pragma unroll
@@ -184,7 +184,7 @@ static __device__ __forceinline__ int warp_reduce_sum(int x) {
         x += __shfl_xor_sync(0xffffffff, x, mask, 32);
     }
     return x;
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_AMPERE
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_AMPERE
 }
 
 static __device__ __forceinline__ float warp_reduce_sum(float x) {
@@ -207,7 +207,7 @@ static __device__ __forceinline__ float2 warp_reduce_sum(float2 a) {
 static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) {
 #ifdef FP16_AVAILABLE
 
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 #pragma unroll
     for (int mask = 16; mask > 0; mask >>= 1) {
         const half2 a_other = __shfl_xor_sync(0xffffffff, a, mask, 32);
@@ -221,7 +221,7 @@ static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) {
         a = __hadd2(a, __shfl_xor_sync(0xffffffff, a, mask, 32));
     }
     return a;
-#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 
 #else
     NO_DEVICE_CODE;
@@ -240,11 +240,11 @@ static __device__ __forceinline__ float warp_reduce_max(float x) {
 static __device__ __forceinline__ half ggml_cuda_hmax(const half a, const half b) {
 #ifdef FP16_AVAILABLE
 
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
     return __float2half(fmaxf(__half2float(a), __half2float(b)));
 #else
     return __hmax(a, b);
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
 
 #else
    NO_DEVICE_CODE;
@@ -254,7 +254,7 @@ static __device__ __forceinline__ half ggml_cuda_hmax(const half a, const half b
 }
 
 static __device__ __forceinline__ half2 ggml_cuda_hmax2(const half2 a, const half2 b) {
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 
 #if CUDART_VERSION >= CUDART_HMAX
     return __hmax2(a, b);
@@ -269,11 +269,11 @@ static __device__ __forceinline__ half2 ggml_cuda_hmax2(const half2 a, const hal
     GGML_UNUSED(a);
     GGML_UNUSED(b);
     NO_DEVICE_CODE;
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 }
 
 static __device__ __forceinline__ half2 warp_reduce_max(half2 x) {
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
 #pragma unroll
    for (int mask = 16; mask > 0; mask >>= 1) {
        x = ggml_cuda_hmax2(x, __shfl_xor_sync(0xffffffff, x, mask, 32));
@@ -282,7 +282,7 @@ static __device__ __forceinline__ half2 warp_reduce_max(half2 x) {
 #else
    GGML_UNUSED(x);
    NO_DEVICE_CODE;
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
 }
 
 #if CUDART_VERSION < CUDART_HMASK
@@ -294,7 +294,7 @@ static __device__ __forceinline__ uint32_t __hgt2_mask(const half2 a, const half
 #endif // CUDART_VERSION < CUDART_HMASK
 
 static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, int c) {
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 #if defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || defined(RDNA2)
     c = __builtin_amdgcn_sdot4(a, b, c, false);
 #elif defined(RDNA3)
@@ -320,7 +320,7 @@ static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, i
 #endif
     return c;
 
-#else // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#else // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 
 #if __CUDA_ARCH__ >= MIN_CC_DP4A
     return __dp4a(a, b, c);
@@ -330,7 +330,7 @@ static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, i
     return c + a8[0]*b8[0] + a8[1]*b8[1] + a8[2]*b8[2] + a8[3]*b8[3];
 #endif // __CUDA_ARCH__ >= MIN_CC_DP4A
 
-#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 }
 
 // TODO: move to ggml-common.h

ggml/src/ggml-cuda/fattn-common.cuh

@@ -517,9 +517,9 @@ constexpr __device__ dequantize_1_f32_t get_dequantize_1_f32(ggml_type type_V) {
 }
 
 template<int D, int parallel_blocks> // D == head size
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 __launch_bounds__(D, 1)
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 static __global__ void flash_attn_combine_results(
         const float  * __restrict__ VKQ_parts,
         const float2 * __restrict__ VKQ_meta,

ggml/src/ggml-cuda/fattn-tile-f16.cu

@@ -5,9 +5,9 @@
 #define FATTN_KQ_STRIDE_TILE_F16 64
 
 template<int D, int ncols, int nwarps, int parallel_blocks, bool use_logit_softcap> // D == head size
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 __launch_bounds__(nwarps*WARP_SIZE, 1)
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 static __global__ void flash_attn_tile_ext_f16(
         const char * __restrict__ Q,
         const char * __restrict__ K,

ggml/src/ggml-cuda/fattn-tile-f32.cu

@@ -5,9 +5,9 @@
 #define FATTN_KQ_STRIDE_TILE_F32 32
 
 template<int D, int ncols, int nwarps, int parallel_blocks, bool use_logit_softcap> // D == head size
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 __launch_bounds__(nwarps*WARP_SIZE, 1)
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 static __global__ void flash_attn_tile_ext_f32(
         const char * __restrict__ Q,
         const char * __restrict__ K,

ggml/src/ggml-cuda/fattn-vec-f16.cuh

@@ -2,9 +2,9 @@
 #include "fattn-common.cuh"
 
 template<int D, int ncols, int parallel_blocks, ggml_type type_K, ggml_type type_V, bool use_logit_softcap> // D == head size
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 __launch_bounds__(D, 1)
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 static __global__ void flash_attn_vec_ext_f16(
         const char * __restrict__ Q,
         const char * __restrict__ K,

ggml/src/ggml-cuda/fattn-vec-f32.cuh

@@ -2,9 +2,9 @@
 #include "fattn-common.cuh"
 
 template<int D, int ncols, int parallel_blocks, ggml_type type_K, ggml_type type_V, bool use_logit_softcap> // D == head size
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 __launch_bounds__(D, 1)
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 static __global__ void flash_attn_vec_ext_f32(
         const char * __restrict__ Q,
         const char * __restrict__ K,

ggml/src/ggml-cuda/fattn-wmma-f16.cuh

@@ -7,9 +7,9 @@
 
 // D == head size, VKQ_stride == num VKQ rows calculated in parallel:
 template<int D, int ncols, int nwarps, int VKQ_stride, int parallel_blocks, typename KQ_acc_t, bool use_logit_softcap>
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 __launch_bounds__(nwarps*WARP_SIZE, 1)
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 static __global__ void flash_attn_ext_f16(
         const char * __restrict__ Q,
         const char * __restrict__ K,

ggml/src/ggml-cuda/ggml/CMakeLists.txt

@@ -0,0 +1,165 @@
+cmake_minimum_required(VERSION 3.18)  # for CMAKE_CUDA_ARCHITECTURES
+
+find_package(CUDAToolkit)
+
+if (CUDAToolkit_FOUND)
+    message(STATUS "CUDA Toolkit found")
+
+    if (NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
+        # 52 == lowest CUDA 12 standard
+        # 60 == FP16 CUDA intrinsics
+        # 61 == integer CUDA intrinsics
+        # 70 == compute capability at which unrolling a loop in mul_mat_q kernels is faster
+        if (GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
+            set(CMAKE_CUDA_ARCHITECTURES "60;61;70;75")
+        else()
+            set(CMAKE_CUDA_ARCHITECTURES "52;61;70;75")
+            #set(CMAKE_CUDA_ARCHITECTURES "OFF") # use this to compile much faster, but only F16 models work
+        endif()
+    endif()
+    message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")
+
+    enable_language(CUDA)
+
+    file(GLOB   GGML_HEADERS_CUDA "*.cuh")
+    list(APPEND GGML_HEADERS_CUDA "../../include/ggml-cuda.h")
+
+    file(GLOB   GGML_SOURCES_CUDA "*.cu")
+    file(GLOB   SRCS "template-instances/fattn-wmma*.cu")
+    list(APPEND GGML_SOURCES_CUDA ${SRCS})
+    file(GLOB   SRCS "template-instances/mmq*.cu")
+    list(APPEND GGML_SOURCES_CUDA ${SRCS})
+
+    if (GGML_CUDA_FA_ALL_QUANTS)
+        file(GLOB   SRCS "template-instances/fattn-vec*.cu")
+        list(APPEND GGML_SOURCES_CUDA ${SRCS})
+        add_compile_definitions(GGML_CUDA_FA_ALL_QUANTS)
+    else()
+        file(GLOB   SRCS "template-instances/fattn-vec*q4_0-q4_0.cu")
+        list(APPEND GGML_SOURCES_CUDA ${SRCS})
+        file(GLOB   SRCS "template-instances/fattn-vec*q8_0-q8_0.cu")
+        list(APPEND GGML_SOURCES_CUDA ${SRCS})
+        file(GLOB   SRCS "template-instances/fattn-vec*f16-f16.cu")
+        list(APPEND GGML_SOURCES_CUDA ${SRCS})
+    endif()
+
+    add_library(ggml-cuda
+                ${GGML_HEADERS_CUDA}
+                ${GGML_SOURCES_CUDA}
+                )
+
+    target_link_libraries(ggml-cuda PRIVATE ggml-base)
+    target_include_directories(ggml-cuda PRIVATE . ..)
+
+    # TODO: change the definitions to this target only
+
+    add_compile_definitions(GGML_CUDA_DMMV_X=${GGML_CUDA_DMMV_X})
+    add_compile_definitions(GGML_CUDA_MMV_Y=${GGML_CUDA_MMV_Y})
+    add_compile_definitions(K_QUANTS_PER_ITERATION=${GGML_CUDA_KQUANTS_ITER})
+    add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${GGML_CUDA_PEER_MAX_BATCH_SIZE})
+
+    if (GGML_CUDA_GRAPHS)
+        add_compile_definitions(GGML_CUDA_USE_GRAPHS)
+    endif()
+
+    if (GGML_CUDA_FORCE_DMMV)
+        add_compile_definitions(GGML_CUDA_FORCE_DMMV)
+    endif()
+
+    if (GGML_CUDA_FORCE_MMQ)
+        add_compile_definitions(GGML_CUDA_FORCE_MMQ)
+    endif()
+
+    if (GGML_CUDA_FORCE_CUBLAS)
+        add_compile_definitions(GGML_CUDA_FORCE_CUBLAS)
+    endif()
+
+    if (GGML_CUDA_NO_VMM)
+        add_compile_definitions(GGML_CUDA_NO_VMM)
+    endif()
+
+    if (DEFINED GGML_CUDA_DMMV_Y)
+        add_compile_definitions(GGML_CUDA_MMV_Y=${GGML_CUDA_DMMV_Y}) # for backwards compatibility
+    endif()
+
+    if (GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
+        add_compile_definitions(GGML_CUDA_F16)
+    endif()
+
+    if (GGML_CUDA_NO_PEER_COPY)
+        add_compile_definitions(GGML_CUDA_NO_PEER_COPY)
+    endif()
+
+    if (GGML_STATIC)
+        if (WIN32)
+            # As of 12.3.1 CUDA Toolkit for Windows does not offer a static cublas library
+            target_link_libraries(ggml-cuda PRIVATE CUDA::cudart_static CUDA::cublas CUDA::cublasLt)
+        else ()
+            target_link_libraries(ggml-cuda PRIVATE  CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
+        endif()
+    else()
+        target_link_libraries(ggml-cuda PRIVATE CUDA::cudart CUDA::cublas CUDA::cublasLt)
+    endif()
+
+    if (GGML_CUDA_NO_VMM)
+        # No VMM requested, no need to link directly with the cuda driver lib (libcuda.so)
+    else()
+        target_link_libraries(ggml-cuda PRIVATE CUDA::cuda_driver)
+    endif()
+
+    set(CUDA_CXX_FLAGS "")
+
+    set(CUDA_FLAGS -use_fast_math)
+
+    if (GGML_FATAL_WARNINGS)
+        list(APPEND CUDA_FLAGS -Werror all-warnings)
+    endif()
+
+    if (GGML_ALL_WARNINGS AND NOT MSVC)
+        set(NVCC_CMD ${CMAKE_CUDA_COMPILER} .c)
+        if (NOT CMAKE_CUDA_HOST_COMPILER STREQUAL "")
+            list(APPEND NVCC_CMD -ccbin ${CMAKE_CUDA_HOST_COMPILER})
+        endif()
+
+        execute_process(
+            COMMAND ${NVCC_CMD} -Xcompiler --version
+            OUTPUT_VARIABLE CUDA_CCFULLVER
+            ERROR_QUIET
+        )
+
+        if (NOT CUDA_CCFULLVER MATCHES clang)
+            set(CUDA_CCID "GNU")
+            execute_process(
+                COMMAND ${NVCC_CMD} -Xcompiler "-dumpfullversion -dumpversion"
+                OUTPUT_VARIABLE CUDA_CCVER
+                ERROR_QUIET
+            )
+        else()
+            if (CUDA_CCFULLVER MATCHES Apple)
+                set(CUDA_CCID "AppleClang")
+            else()
+                set(CUDA_CCID "Clang")
+            endif()
+            string(REGEX REPLACE "^.* version ([0-9.]*).*$" "\\1" CUDA_CCVER ${CUDA_CCFULLVER})
+        endif()
+
+        message("-- CUDA host compiler is ${CUDA_CCID} ${CUDA_CCVER}")
+
+        get_flags(${CUDA_CCID} ${CUDA_CCVER})
+        list(APPEND CUDA_CXX_FLAGS ${CXX_FLAGS} ${GF_CXX_FLAGS})  # This is passed to -Xcompiler later
+    endif()
+
+    if (NOT MSVC)
+        list(APPEND CUDA_CXX_FLAGS -Wno-pedantic)
+    endif()
+
+    list(JOIN   CUDA_CXX_FLAGS " " CUDA_CXX_FLAGS_JOINED)  # pass host compiler flags as a single argument
+
+    if (NOT CUDA_CXX_FLAGS_JOINED STREQUAL "")
+        list(APPEND CUDA_FLAGS -Xcompiler ${CUDA_CXX_FLAGS_JOINED})
+    endif()
+
+    add_compile_options("$<$<COMPILE_LANGUAGE:CUDA>:${CUDA_FLAGS}>")
+else()
+    message(FATAL_ERROR "CUDA Toolkit not found")
+endif()

ggml/src/ggml-cuda/mmq.cuh

@@ -100,9 +100,9 @@ static constexpr __device__ int get_mmq_x_max_device() {
     return 128;
 #else // INT8_MMA_AVAILABLE
 
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
     return 128;
-#else // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#else // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 
 #if __CUDA_ARCH__ >= CC_VOLTA
 #ifdef GGML_CUDA_FORCE_MMQ
@@ -115,7 +115,7 @@ static constexpr __device__ int get_mmq_x_max_device() {
     return 64;
 #endif // __CUDA_ARCH__ >= CC_VOLTA
 
-#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 #endif // INT8_MMA_AVAILABLE
 }
 
@@ -124,7 +124,7 @@ static constexpr int get_mmq_y_host(const int cc) {
 }
 
 static constexpr __device__ int get_mmq_y_device() {
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 #if defined(RDNA1)
     return 64;
 #else
@@ -136,7 +136,7 @@ static constexpr __device__ int get_mmq_y_device() {
 #else
     return 64;
 #endif // __CUDA_ARCH__ >= CC_VOLTA
-#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 }
 
 #define MMQ_DP4A_TXS_Q4_0    tile_x_sizes{mmq_y*WARP_SIZE   + mmq_y, mmq_y*WARP_SIZE/QI4_0   + mmq_y/QI4_0,     0}
@@ -2569,7 +2569,7 @@ static __device__ void mul_mat_q_process_tile(
 // The mul_mat_q kernel implements "stream-k" work partitioning as described in https://arxiv.org/abs/2301.03598
 
 template <ggml_type type, int mmq_x, int nwarps, bool need_check>
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 #if defined(RDNA3) || defined(RDNA2)
     __launch_bounds__(WARP_SIZE*nwarps, 2)
 #endif // defined(RDNA3) || defined(RDNA2)
@@ -2579,7 +2579,7 @@ template <ggml_type type, int mmq_x, int nwarps, bool need_check>
 #else
     __launch_bounds__(WARP_SIZE*nwarps, 2)
 #endif // __CUDA_ARCH__ >= CC_VOLTA
-#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 static __global__ void mul_mat_q(
     const char * __restrict__ x, const char * __restrict__ yc, float * __restrict__ dst, float * __restrict__ tmp_fixup,
     const int ne00, const int ne01, const int stride01, const int ne10, const int ne11, const int stride11, const int ne0) {
@@ -2594,7 +2594,7 @@ static __global__ void mul_mat_q(
     constexpr int mmq_y = get_mmq_y_device();
 
     // On AMD or old CUDA the performance with stream-k was worse, use conventional tiling instead:
-#if (defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ < CC_VOLTA
+#if (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ < CC_VOLTA
     {
         constexpr bool fixup = false;
         mul_mat_q_process_tile<type, mmq_x, nwarps, need_check, fixup>
@@ -2602,7 +2602,7 @@ static __global__ void mul_mat_q(
                 blockIdx.x, blockIdx.y, 0, ne00/qk);
         return;
     }
-#endif // (defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ < CC_VOLTA
+#endif // (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ < CC_VOLTA
 
     const     int64_t blocks_per_ne00 = ne00 / qk;
     constexpr int     blocks_per_iter = MMQ_ITER_K / qk;
@@ -2765,14 +2765,14 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a
 
     const int shmem = mmq_get_shmem<type>(mmq_x, mmq_y, cc);
 
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
     static bool shmem_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
     if (!shmem_limit_raised[id]) {
         CUDA_CHECK(cudaFuncSetAttribute(mul_mat_q<type, mmq_x, MMQ_NWARPS, false>, cudaFuncAttributeMaxDynamicSharedMemorySize, shmem));
         CUDA_CHECK(cudaFuncSetAttribute(mul_mat_q<type, mmq_x, MMQ_NWARPS, true>,  cudaFuncAttributeMaxDynamicSharedMemorySize, shmem));
         shmem_limit_raised[id] = true;
     }
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 
     const int nty = (args.ne01 + mmq_y - 1) / mmq_y;
     const int ntx = (args.ne11 + mmq_x - 1) / mmq_x;

ggml/src/ggml-cuda/mmvq.cu

@@ -48,10 +48,10 @@ static constexpr __device__ int get_vdr_mmvq(ggml_type type) {
 }
 
 template <ggml_type type, int ncols_y>
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 // tell the compiler to use as many registers as it wants, see nwarps definition below
 __launch_bounds__((ncols_y <= 4 ? 4 : 2)*WARP_SIZE, 1)
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 static __global__ void mul_mat_vec_q(
     const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
     const int ncols_x, const int nrows_x, const int nrows_y, const int nrows_dst) {
@@ -62,13 +62,13 @@ static __global__ void mul_mat_vec_q(
 
     constexpr vec_dot_q_cuda_t vec_dot_q_cuda = get_vec_dot_q_cuda(type);
 
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) && (defined(RDNA2) || defined(RDNA3))
+#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && (defined(RDNA2) || defined(RDNA3))
     constexpr int nwarps              = 1;
     constexpr int rows_per_cuda_block = 1;
 #else
     constexpr int nwarps              = ncols_y <= 4 ? 4 : 2;
     constexpr int rows_per_cuda_block = ncols_y == 1 ? 1 : 2;
-#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) && !defined(RDNA2) && !defined(RDNA3)
+#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && !defined(RDNA2) && !defined(RDNA3)
 
     const     int tid = WARP_SIZE*threadIdx.y + threadIdx.x;
     const     int row0 = rows_per_cuda_block*blockIdx.x;

ggml/src/ggml-cuda/sum.cu

@@ -1,6 +1,6 @@
-#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11700
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11700
 #define USE_CUB
-#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11700
+#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11700
 
 #ifdef USE_CUB
 // On Windows CUB uses libraries with variables called CC_PASCAL which conflict with the define in common.cuh.

ggml/src/ggml-hip/CMakeLists.txt

@@ -0,0 +1,113 @@
+if (NOT EXISTS $ENV{ROCM_PATH})
+    if (NOT EXISTS /opt/rocm)
+        set(ROCM_PATH /usr)
+    else()
+        set(ROCM_PATH /opt/rocm)
+    endif()
+else()
+    set(ROCM_PATH $ENV{ROCM_PATH})
+endif()
+
+list(APPEND CMAKE_PREFIX_PATH  ${ROCM_PATH})
+list(APPEND CMAKE_PREFIX_PATH "${ROCM_PATH}/lib64/cmake")
+
+# CMake on Windows doesn't support the HIP language yet
+if (WIN32)
+    set(CXX_IS_HIPCC TRUE)
+else()
+    string(REGEX MATCH "hipcc(\.bat)?$" CXX_IS_HIPCC "${CMAKE_CXX_COMPILER}")
+endif()
+
+if (CXX_IS_HIPCC)
+    if (LINUX)
+        if (NOT ${CMAKE_CXX_COMPILER_ID} MATCHES "Clang")
+            message(WARNING "Only LLVM is supported for HIP, hint: CXX=/opt/rocm/llvm/bin/clang++")
+        endif()
+
+        message(WARNING "Setting hipcc as the C++ compiler is legacy behavior."
+                " Prefer setting the HIP compiler directly. See README for details.")
+    endif()
+else()
+    # Forward AMDGPU_TARGETS to CMAKE_HIP_ARCHITECTURES.
+    if (AMDGPU_TARGETS AND NOT CMAKE_HIP_ARCHITECTURES)
+        set(CMAKE_HIP_ARCHITECTURES ${AMDGPU_TARGETS})
+    endif()
+    cmake_minimum_required(VERSION 3.21)
+    enable_language(HIP)
+endif()
+
+find_package(hip     REQUIRED)
+find_package(hipblas REQUIRED)
+find_package(rocblas REQUIRED)
+
+message(STATUS "HIP and hipBLAS found")
+
+file(GLOB   GGML_HEADERS_ROCM "../ggml-cuda/*.cuh")
+list(APPEND GGML_HEADERS_ROCM "../../include/ggml-cuda.h")
+
+file(GLOB   GGML_SOURCES_ROCM "../ggml-cuda/*.cu")
+file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-wmma*.cu")
+list(APPEND GGML_SOURCES_ROCM ${SRCS})
+file(GLOB   SRCS "../ggml-cuda/template-instances/mmq*.cu")
+list(APPEND GGML_SOURCES_ROCM ${SRCS})
+
+if (GGML_CUDA_FA_ALL_QUANTS)
+    file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-vec*.cu")
+    list(APPEND GGML_SOURCES_ROCM ${SRCS})
+    add_compile_definitions(GGML_CUDA_FA_ALL_QUANTS)
+else()
+    file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-vec*q4_0-q4_0.cu")
+    list(APPEND GGML_SOURCES_ROCM ${SRCS})
+    file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-vec*q8_0-q8_0.cu")
+    list(APPEND GGML_SOURCES_ROCM ${SRCS})
+    file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-vec*f16-f16.cu")
+    list(APPEND GGML_SOURCES_ROCM ${SRCS})
+endif()
+
+add_library(ggml-hip
+            ${GGML_HEADERS_ROCM}
+            ${GGML_SOURCES_ROCM})
+
+target_link_libraries(ggml-hip PRIVATE ggml-base)
+target_include_directories(ggml-hip PRIVATE . ..)
+
+# TODO: do not use CUDA definitions for HIP
+target_compile_definitions(ggml PUBLIC GGML_USE_CUDA)
+
+add_compile_definitions(GGML_USE_HIP)
+add_compile_definitions(GGML_CUDA_DMMV_X=${GGML_CUDA_DMMV_X})
+add_compile_definitions(GGML_CUDA_MMV_Y=${GGML_CUDA_MMV_Y})
+add_compile_definitions(K_QUANTS_PER_ITERATION=${GGML_CUDA_KQUANTS_ITER})
+
+if (GGML_HIP_UMA)
+    add_compile_definitions(GGML_HIP_UMA)
+endif()
+
+if (GGML_CUDA_FORCE_DMMV)
+    add_compile_definitions(GGML_CUDA_FORCE_DMMV)
+endif()
+
+if (GGML_CUDA_FORCE_MMQ)
+    add_compile_definitions(GGML_CUDA_FORCE_MMQ)
+endif()
+
+if (GGML_CUDA_FORCE_CUBLAS)
+    add_compile_definitions(GGML_CUDA_FORCE_CUBLAS)
+endif()
+
+if (GGML_CUDA_NO_PEER_COPY)
+    add_compile_definitions(GGML_CUDA_NO_PEER_COPY)
+endif()
+
+if (CXX_IS_HIPCC)
+    set_source_files_properties(${GGML_SOURCES_ROCM} PROPERTIES LANGUAGE CXX)
+    target_link_libraries(ggml-hip PRIVATE hip::device)
+else()
+    set_source_files_properties(${GGML_SOURCES_ROCM} PROPERTIES LANGUAGE HIP)
+endif()
+
+if (GGML_STATIC)
+    message(FATAL_ERROR "Static linking not supported for HIP/ROCm")
+endif()
+
+target_link_libraries(ggml-hip PRIVATE ggml-base hip::host roc::rocblas roc::hipblas)

ggml/src/ggml-impl.h

@@ -3,13 +3,29 @@
 // GGML internal header
 
 #include "ggml.h"
-
 #include <assert.h>
+#include <math.h>
 #include <stdlib.h> // load `stdlib.h` before other headers to work around MinGW bug: https://sourceforge.net/p/mingw-w64/bugs/192/
 #include <stdbool.h>
 #include <stdint.h>
 #include <string.h>
 
+#ifdef __ARM_FEATURE_SVE
+#include <arm_sve.h>
+#endif // __ARM_FEATURE_SVE
+
+#if defined(__ARM_NEON)
+// if YCM cannot find <arm_neon.h>, make a symbolic link to it, for example:
+//
+//   $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/
+//
+#include <arm_neon.h>
+#endif
+
+#if defined(__F16C__)
+#include <immintrin.h>
+#endif
+
 #ifdef __cplusplus
 extern "C" {
 #endif
@@ -28,13 +44,13 @@ extern "C" {
 // if C99 - static_assert is noop
 // ref: https://stackoverflow.com/a/53923785/4039976
 #ifndef __cplusplus
-#ifndef static_assert
-#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L)
-#define static_assert(cond, msg) _Static_assert(cond, msg)
-#else
-#define static_assert(cond, msg) struct global_scope_noop_trick
-#endif
-#endif
+    #ifndef static_assert
+        #if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L)
+            #define static_assert(cond, msg) _Static_assert(cond, msg)
+        #else
+            #define static_assert(cond, msg) struct global_scope_noop_trick
+        #endif
+    #endif
 #endif
 
 static inline int ggml_up32(int n) {
@@ -120,14 +136,12 @@ struct ggml_map_custom1_op_params {
     void             * userdata;
 };
 
-
 struct ggml_map_custom2_op_params {
     ggml_custom2_op_t   fun;
     int                 n_tasks;
     void              * userdata;
 };
 
-
 struct ggml_map_custom3_op_params {
     ggml_custom3_op_t fun;
     int n_tasks;
@@ -287,9 +301,249 @@ struct ggml_cgraph ggml_graph_view(struct ggml_cgraph * cgraph, int i0, int i1);
 void * ggml_aligned_malloc(size_t size);
 void ggml_aligned_free(void * ptr, size_t size);
 
-// TODO: move to threading file
-void ggml_critical_section_start(void);
-void ggml_critical_section_end(void);
+// FP16 to FP32 conversion
+
+#if defined(__ARM_NEON)
+    #ifdef _MSC_VER
+        typedef uint16_t ggml_fp16_internal_t;
+    #else
+        typedef __fp16 ggml_fp16_internal_t;
+    #endif
+#endif
+
+#if defined(__ARM_NEON) && !defined(_MSC_VER)
+    #define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
+    #define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
+
+    #define GGML_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
+
+    static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
+        ggml_fp16_internal_t tmp;
+        memcpy(&tmp, &h, sizeof(ggml_fp16_t));
+        return (float)tmp;
+    }
+
+    static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
+        ggml_fp16_t res;
+        ggml_fp16_internal_t tmp = f;
+        memcpy(&res, &tmp, sizeof(ggml_fp16_t));
+        return res;
+    }
+
+#elif defined(__F16C__)
+
+    #ifdef _MSC_VER
+        #define GGML_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x)))
+        #define GGML_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0)
+    #else
+        #define GGML_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x)
+        #define GGML_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0)
+    #endif
+
+#elif defined(__POWER9_VECTOR__)
+
+    #define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
+    #define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
+    /* the inline asm below is about 12% faster than the lookup method */
+    #define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x)
+    #define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
+
+    static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
+        register float f;
+        register double d;
+        __asm__(
+            "mtfprd %0,%2\n"
+            "xscvhpdp %0,%0\n"
+            "frsp %1,%0\n" :
+            /* temp */ "=d"(d),
+            /* out */  "=f"(f):
+            /* in */   "r"(h));
+        return f;
+    }
+
+    static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
+        register double d;
+        register ggml_fp16_t r;
+        __asm__( /* xscvdphp can work on double or single precision */
+            "xscvdphp %0,%2\n"
+            "mffprd %1,%0\n" :
+            /* temp */ "=d"(d),
+            /* out */  "=r"(r):
+            /* in */   "f"(f));
+        return r;
+    }
+
+#else
+
+    // FP16 <-> FP32
+    // ref: https://github.com/Maratyszcza/FP16
+
+    static inline float fp32_from_bits(uint32_t w) {
+        union {
+            uint32_t as_bits;
+            float as_value;
+        } fp32;
+        fp32.as_bits = w;
+        return fp32.as_value;
+    }
+
+    static inline uint32_t fp32_to_bits(float f) {
+        union {
+            float as_value;
+            uint32_t as_bits;
+        } fp32;
+        fp32.as_value = f;
+        return fp32.as_bits;
+    }
+
+    static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
+        const uint32_t w = (uint32_t) h << 16;
+        const uint32_t sign = w & UINT32_C(0x80000000);
+        const uint32_t two_w = w + w;
+
+        const uint32_t exp_offset = UINT32_C(0xE0) << 23;
+    #if (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)) && (!defined(__cplusplus) || __cplusplus >= 201703L)
+        const float exp_scale = 0x1.0p-112f;
+    #else
+        const float exp_scale = fp32_from_bits(UINT32_C(0x7800000));
+    #endif
+        const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale;
+
+        const uint32_t magic_mask = UINT32_C(126) << 23;
+        const float magic_bias = 0.5f;
+        const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias;
+
+        const uint32_t denormalized_cutoff = UINT32_C(1) << 27;
+        const uint32_t result = sign |
+            (two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value));
+        return fp32_from_bits(result);
+    }
+
+    static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
+    #if (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)) && (!defined(__cplusplus) || __cplusplus >= 201703L)
+        const float scale_to_inf = 0x1.0p+112f;
+        const float scale_to_zero = 0x1.0p-110f;
+    #else
+        const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000));
+        const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000));
+    #endif
+        float base = (fabsf(f) * scale_to_inf) * scale_to_zero;
+
+        const uint32_t w = fp32_to_bits(f);
+        const uint32_t shl1_w = w + w;
+        const uint32_t sign = w & UINT32_C(0x80000000);
+        uint32_t bias = shl1_w & UINT32_C(0xFF000000);
+        if (bias < UINT32_C(0x71000000)) {
+            bias = UINT32_C(0x71000000);
+        }
+
+        base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base;
+        const uint32_t bits = fp32_to_bits(base);
+        const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00);
+        const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF);
+        const uint32_t nonsign = exp_bits + mantissa_bits;
+        return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign);
+    }
+
+    #define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
+    #define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
+
+#endif // defined(__ARM_NEON) && (!defined(__MSC_VER)
+
+// precomputed f32 table for f16 (256 KB)
+// defined in ggml.c, initialized in ggml_init()
+GGML_API float ggml_table_f32_f16[1 << 16];
+
+// On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
+// so we define GGML_FP16_TO_FP32 and GGML_FP32_TO_FP16 elsewhere for NEON.
+// This is also true for POWER9.
+#if !defined(GGML_FP16_TO_FP32)
+inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
+    uint16_t s;
+    memcpy(&s, &f, sizeof(uint16_t));
+    return ggml_table_f32_f16[s];
+}
+
+#define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x)
+#endif
+
+#if !defined(GGML_FP32_TO_FP16)
+#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
+#endif
+
+/**
+ * Converts brain16 to float32.
+ *
+ * The bfloat16 floating point format has the following structure:
+ *
+ *       ┌sign
+ *       │
+ *       │   ┌exponent
+ *       │   │
+ *       │   │      ┌mantissa
+ *       │   │      │
+ *       │┌──┴───┐┌─┴───┐
+ *     0b0000000000000000 brain16
+ *
+ * Since bf16 has the same number of exponent bits as a 32bit float,
+ * encoding and decoding numbers becomes relatively straightforward.
+ *
+ *       ┌sign
+ *       │
+ *       │   ┌exponent
+ *       │   │
+ *       │   │      ┌mantissa
+ *       │   │      │
+ *       │┌──┴───┐┌─┴───────────────────┐
+ *     0b00000000000000000000000000000000 IEEE binary32
+ *
+ * For comparison, the standard fp16 format has fewer exponent bits.
+ *
+ *       ┌sign
+ *       │
+ *       │  ┌exponent
+ *       │  │
+ *       │  │    ┌mantissa
+ *       │  │    │
+ *       │┌─┴─┐┌─┴──────┐
+ *     0b0000000000000000 IEEE binary16
+ *
+ * @see IEEE 754-2008
+ */
+static inline float ggml_compute_bf16_to_fp32(ggml_bf16_t h) {
+    union {
+        float f;
+        uint32_t i;
+    } u;
+    u.i = (uint32_t)h.bits << 16;
+    return u.f;
+}
+
+/**
+ * Converts float32 to brain16.
+ *
+ * This is binary identical with Google Brain float conversion.
+ * Floats shall round to nearest even, and NANs shall be quiet.
+ * Subnormals aren't flushed to zero, except perhaps when used.
+ * This code should vectorize nicely if using modern compilers.
+ */
+static inline ggml_bf16_t ggml_compute_fp32_to_bf16(float s) {
+    ggml_bf16_t h;
+    union {
+        float f;
+        uint32_t i;
+    } u;
+    u.f = s;
+    if ((u.i & 0x7fffffff) > 0x7f800000) { /* nan */
+        h.bits = (u.i >> 16) | 64; /* force to quiet */
+        return h;
+    }
+    h.bits = (u.i + (0x7fff + ((u.i >> 16) & 1))) >> 16;
+    return h;
+}
+
+#define GGML_FP32_TO_BF16(x) ggml_compute_fp32_to_bf16(x)
+#define GGML_BF16_TO_FP32(x) ggml_compute_bf16_to_fp32(x)
 
 #ifdef __cplusplus
 }

ggml/src/ggml-kompute/CMakeLists.txt

@@ -0,0 +1,162 @@
+
+find_package(Vulkan COMPONENTS glslc REQUIRED)
+find_program(glslc_executable NAMES glslc HINTS Vulkan::glslc)
+
+if (NOT glslc_executable)
+    message(FATAL_ERROR "glslc not found")
+endif()
+
+add_library(ggml-kompute
+            ggml-kompute.cpp
+            ../../include/ggml-kompute.h
+            )
+
+target_link_libraries(ggml-kompute PRIVATE ggml-base kompute)
+target_include_directories(ggml-kompute PRIVATE . .. ${CMAKE_CURRENT_BINARY_DIR})
+
+add_compile_definitions(VULKAN_HPP_DISPATCH_LOADER_DYNAMIC=1)
+
+function(compile_shader)
+    set(options)
+    set(oneValueArgs)
+    set(multiValueArgs SOURCES)
+    cmake_parse_arguments(compile_shader "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
+    foreach(source ${compile_shader_SOURCES})
+        get_filename_component(filename ${source} NAME)
+        set(spv_file ${filename}.spv)
+        add_custom_command(
+            OUTPUT ${spv_file}
+            DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/${source}
+            ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/common.comp
+            ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_getrows.comp
+            ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n_pre.comp
+            ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n.comp
+            COMMAND ${glslc_executable} --target-env=vulkan1.2 -o ${spv_file} ${CMAKE_CURRENT_SOURCE_DIR}/${source}
+            COMMENT "Compiling ${source} to ${spv_file}"
+            )
+
+        get_filename_component(RAW_FILE_NAME ${spv_file} NAME)
+        set(FILE_NAME "shader${RAW_FILE_NAME}")
+        string(REPLACE ".comp.spv" ".h" HEADER_FILE ${FILE_NAME})
+        string(TOUPPER ${HEADER_FILE} HEADER_FILE_DEFINE)
+        string(REPLACE "." "_" HEADER_FILE_DEFINE "${HEADER_FILE_DEFINE}")
+        set(OUTPUT_HEADER_FILE "${HEADER_FILE}")
+        message(STATUS "${HEADER_FILE} generating ${HEADER_FILE_DEFINE}")
+        if(CMAKE_GENERATOR MATCHES "Visual Studio")
+            add_custom_command(
+                OUTPUT ${OUTPUT_HEADER_FILE}
+                COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
+                COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+                COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+                COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
+                COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
+                COMMAND ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
+                COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
+                COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+                DEPENDS ${spv_file} xxd
+                COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd"
+                )
+        else()
+            add_custom_command(
+                OUTPUT ${OUTPUT_HEADER_FILE}
+                COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
+                COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+                COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+                COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
+                COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
+                COMMAND ${CMAKE_BINARY_DIR}/bin/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
+                COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
+                COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+                DEPENDS ${spv_file} xxd
+                COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/xxd"
+                )
+        endif()
+    endforeach()
+endfunction()
+
+if (EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/kompute/CMakeLists.txt")
+    message(STATUS "Kompute found")
+    set(KOMPUTE_OPT_LOG_LEVEL Error CACHE STRING "Kompute log level")
+    add_subdirectory(kompute)
+
+    # Compile our shaders
+    compile_shader(SOURCES
+        kompute-shaders/op_scale.comp
+        kompute-shaders/op_scale_8.comp
+        kompute-shaders/op_add.comp
+        kompute-shaders/op_addrow.comp
+        kompute-shaders/op_mul.comp
+        kompute-shaders/op_silu.comp
+        kompute-shaders/op_relu.comp
+        kompute-shaders/op_gelu.comp
+        kompute-shaders/op_softmax.comp
+        kompute-shaders/op_norm.comp
+        kompute-shaders/op_rmsnorm.comp
+        kompute-shaders/op_diagmask.comp
+        kompute-shaders/op_mul_mat_mat_f32.comp
+        kompute-shaders/op_mul_mat_f16.comp
+        kompute-shaders/op_mul_mat_q8_0.comp
+        kompute-shaders/op_mul_mat_q4_0.comp
+        kompute-shaders/op_mul_mat_q4_1.comp
+        kompute-shaders/op_mul_mat_q4_k.comp
+        kompute-shaders/op_mul_mat_q6_k.comp
+        kompute-shaders/op_getrows_f32.comp
+        kompute-shaders/op_getrows_f16.comp
+        kompute-shaders/op_getrows_q4_0.comp
+        kompute-shaders/op_getrows_q4_1.comp
+        kompute-shaders/op_getrows_q6_k.comp
+        kompute-shaders/op_rope_f16.comp
+        kompute-shaders/op_rope_f32.comp
+        kompute-shaders/op_cpy_f16_f16.comp
+        kompute-shaders/op_cpy_f16_f32.comp
+        kompute-shaders/op_cpy_f32_f16.comp
+        kompute-shaders/op_cpy_f32_f32.comp
+    )
+
+    # Create a custom target for our generated shaders
+    add_custom_target(generated_shaders DEPENDS
+        shaderop_scale.h
+        shaderop_scale_8.h
+        shaderop_add.h
+        shaderop_addrow.h
+        shaderop_mul.h
+        shaderop_silu.h
+        shaderop_relu.h
+        shaderop_gelu.h
+        shaderop_softmax.h
+        shaderop_norm.h
+        shaderop_rmsnorm.h
+        shaderop_diagmask.h
+        shaderop_mul_mat_mat_f32.h
+        shaderop_mul_mat_f16.h
+        shaderop_mul_mat_q8_0.h
+        shaderop_mul_mat_q4_0.h
+        shaderop_mul_mat_q4_1.h
+        shaderop_mul_mat_q4_k.h
+        shaderop_mul_mat_q6_k.h
+        shaderop_getrows_f32.h
+        shaderop_getrows_f16.h
+        shaderop_getrows_q4_0.h
+        shaderop_getrows_q4_1.h
+        shaderop_getrows_q6_k.h
+        shaderop_rope_f16.h
+        shaderop_rope_f32.h
+        shaderop_cpy_f16_f16.h
+        shaderop_cpy_f16_f32.h
+        shaderop_cpy_f32_f16.h
+        shaderop_cpy_f32_f32.h
+    )
+
+    # Create a custom command that depends on the generated_shaders
+    add_custom_command(
+        OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp
+        COMMAND ${CMAKE_COMMAND} -E touch ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp
+        DEPENDS generated_shaders
+        COMMENT "Ensuring shaders are generated before compiling ggml-kompute.cpp"
+    )
+
+    # Add the stamp to the main sources to ensure dependency tracking
+    target_sources(ggml-kompute PRIVATE ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp)
+else()
+    message(WARNING "Kompute not found")
+endif()