mv dpct definition from folder dpct to ggml-sycl.h
This commit is contained in:
jianyuzh
Meng, Hengyu
parent
c67c2ab228
commit
c3c5b20ac5
11 changed files with 2808 additions and 36517 deletions
File diff suppressed because it is too large
Load diff
|
|
@ -1,92 +0,0 @@
|
||||||
#pragma once
|
|
||||||
|
|
||||||
#include "ggml.h"
|
|
||||||
|
|
||||||
#ifdef __cplusplus
|
|
||||||
extern "C" {
|
|
||||||
#endif
|
|
||||||
|
|
||||||
struct ggml_backend;
|
|
||||||
struct ggml_backend_buffer;
|
|
||||||
struct ggml_backend_buffer_type;
|
|
||||||
|
|
||||||
//
|
|
||||||
// Legacy API
|
|
||||||
//
|
|
||||||
|
|
||||||
typedef struct ggml_allocr * ggml_allocr_t;
|
|
||||||
|
|
||||||
// initialize allocator for use with CPU backend only
|
|
||||||
GGML_API ggml_allocr_t ggml_allocr_new(void * data, size_t size, size_t alignment);
|
|
||||||
GGML_API ggml_allocr_t ggml_allocr_new_measure(size_t alignment);
|
|
||||||
|
|
||||||
// initialize allocator for use with ggml-backend
|
|
||||||
GGML_API ggml_allocr_t ggml_allocr_new_from_buffer(struct ggml_backend_buffer * buffer);
|
|
||||||
GGML_API ggml_allocr_t ggml_allocr_new_from_backend(struct ggml_backend * backend, size_t size); // allocates an owned buffer
|
|
||||||
GGML_API ggml_allocr_t ggml_allocr_new_measure_from_backend(struct ggml_backend * backend);
|
|
||||||
|
|
||||||
GGML_API struct ggml_backend_buffer * ggml_allocr_get_buffer(ggml_allocr_t alloc);
|
|
||||||
|
|
||||||
// tell the allocator to parse nodes following the order described in the list
|
|
||||||
// you should call this if your graph are optimized to execute out-of-order
|
|
||||||
GGML_API void ggml_allocr_set_parse_seq(ggml_allocr_t alloc, const int * list, int n);
|
|
||||||
|
|
||||||
GGML_API void ggml_allocr_free (ggml_allocr_t alloc);
|
|
||||||
GGML_API bool ggml_allocr_is_measure (ggml_allocr_t alloc);
|
|
||||||
GGML_API void ggml_allocr_reset (ggml_allocr_t alloc);
|
|
||||||
GGML_API void ggml_allocr_alloc (ggml_allocr_t alloc, struct ggml_tensor * tensor);
|
|
||||||
GGML_API size_t ggml_allocr_max_size (ggml_allocr_t alloc);
|
|
||||||
|
|
||||||
GGML_API size_t ggml_allocr_alloc_graph(ggml_allocr_t alloc, struct ggml_cgraph * graph);
|
|
||||||
|
|
||||||
//
|
|
||||||
// ggml-backend v2 API
|
|
||||||
//
|
|
||||||
|
|
||||||
// Separate tensor and graph allocator objects
|
|
||||||
// This is necessary for multi-backend allocation because the graph allocator needs to use multiple tensor allocators
|
|
||||||
// The original API is kept as a wrapper around the new API
|
|
||||||
|
|
||||||
// Tensor allocator
|
|
||||||
typedef struct ggml_tallocr * ggml_tallocr_t;
|
|
||||||
|
|
||||||
GGML_API ggml_tallocr_t ggml_tallocr_new(void * data, size_t size, size_t alignment);
|
|
||||||
GGML_API ggml_tallocr_t ggml_tallocr_new_measure(size_t alignment);
|
|
||||||
GGML_API ggml_tallocr_t ggml_tallocr_new_from_buffer(struct ggml_backend_buffer * buffer);
|
|
||||||
GGML_API ggml_tallocr_t ggml_tallocr_new_from_backend(struct ggml_backend * backend, size_t size); // allocates an owned buffer
|
|
||||||
GGML_API ggml_tallocr_t ggml_tallocr_new_measure_from_backend(struct ggml_backend * backend);
|
|
||||||
|
|
||||||
GGML_API struct ggml_backend_buffer * ggml_tallocr_get_buffer(ggml_tallocr_t talloc);
|
|
||||||
|
|
||||||
GGML_API void ggml_tallocr_free (ggml_tallocr_t talloc);
|
|
||||||
GGML_API bool ggml_tallocr_is_measure (ggml_tallocr_t talloc);
|
|
||||||
GGML_API void ggml_tallocr_reset (ggml_tallocr_t talloc);
|
|
||||||
GGML_API void ggml_tallocr_alloc (ggml_tallocr_t talloc, struct ggml_tensor * tensor);
|
|
||||||
GGML_API size_t ggml_tallocr_max_size (ggml_tallocr_t talloc);
|
|
||||||
|
|
||||||
|
|
||||||
// Graph allocator
|
|
||||||
typedef struct ggml_gallocr * ggml_gallocr_t;
|
|
||||||
|
|
||||||
GGML_API ggml_gallocr_t ggml_gallocr_new(void);
|
|
||||||
GGML_API void ggml_gallocr_free(ggml_gallocr_t galloc);
|
|
||||||
|
|
||||||
GGML_API void ggml_gallocr_set_parse_seq(ggml_gallocr_t galloc, const int * list, int n);
|
|
||||||
GGML_API size_t ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, ggml_tallocr_t talloc, struct ggml_cgraph * graph);
|
|
||||||
|
|
||||||
// Allocate tensors from the allocators given by the hash table
|
|
||||||
GGML_API void ggml_gallocr_alloc_graph_n(
|
|
||||||
ggml_gallocr_t galloc,
|
|
||||||
struct ggml_cgraph * graph,
|
|
||||||
struct ggml_hash_set hash_set,
|
|
||||||
ggml_tallocr_t * hash_node_talloc);
|
|
||||||
|
|
||||||
|
|
||||||
// Utils
|
|
||||||
// Create a buffer and allocate all the tensors in a ggml_context
|
|
||||||
GGML_API struct ggml_backend_buffer * ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, struct ggml_backend_buffer_type * buft);
|
|
||||||
GGML_API struct ggml_backend_buffer * ggml_backend_alloc_ctx_tensors(struct ggml_context * ctx, struct ggml_backend * backend);
|
|
||||||
|
|
||||||
#ifdef __cplusplus
|
|
||||||
}
|
|
||||||
#endif
|
|
||||||
|
|
@ -1,116 +0,0 @@
|
||||||
#pragma once
|
|
||||||
|
|
||||||
// ggml-backend internal header
|
|
||||||
|
|
||||||
#include "ggml-backend.h"
|
|
||||||
|
|
||||||
#ifdef __cplusplus
|
|
||||||
extern "C" {
|
|
||||||
#endif
|
|
||||||
|
|
||||||
//
|
|
||||||
// Backend buffer
|
|
||||||
//
|
|
||||||
|
|
||||||
// buffer type
|
|
||||||
typedef void * ggml_backend_buffer_type_context_t;
|
|
||||||
|
|
||||||
struct ggml_backend_buffer_type_i {
|
|
||||||
ggml_backend_buffer_t (*alloc_buffer) (ggml_backend_buffer_type_t buft, size_t size);
|
|
||||||
size_t (*get_alignment) (ggml_backend_buffer_type_t buft); // tensor alignment
|
|
||||||
size_t (*get_alloc_size) (ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor); // data size needed to allocate the tensor, including padding
|
|
||||||
bool (*supports_backend)(ggml_backend_buffer_type_t buft, ggml_backend_t backend); // check if the buffer type is usable by the backend
|
|
||||||
// check if tensor data is in host memory
|
|
||||||
// should be equivalent to supports_backend(buft, ggml_backend_cpu_init())
|
|
||||||
bool (*is_host) (ggml_backend_buffer_type_t buft);
|
|
||||||
};
|
|
||||||
|
|
||||||
struct ggml_backend_buffer_type {
|
|
||||||
struct ggml_backend_buffer_type_i iface;
|
|
||||||
ggml_backend_buffer_type_context_t context;
|
|
||||||
};
|
|
||||||
|
|
||||||
// buffer
|
|
||||||
typedef void * ggml_backend_buffer_context_t;
|
|
||||||
|
|
||||||
struct ggml_backend_buffer_i {
|
|
||||||
void (*free_buffer) (ggml_backend_buffer_t buffer);
|
|
||||||
//void (*reset) (ggml_backend_buffer_t buffer); // reset any internal state due to tensor initialization, such as tensor extras
|
|
||||||
void * (*get_base) (ggml_backend_buffer_t buffer);
|
|
||||||
void (*init_tensor) (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
|
|
||||||
void (*set_tensor) (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
|
|
||||||
void (*get_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size);
|
|
||||||
// (optional) copy tensor between different buffer-type, allow for single-copy tranfers
|
|
||||||
void (*cpy_tensor_from)(ggml_backend_buffer_t buffer, struct ggml_tensor * src, struct ggml_tensor * dst);
|
|
||||||
void (*cpy_tensor_to) (ggml_backend_buffer_t buffer, struct ggml_tensor * src, struct ggml_tensor * dst);
|
|
||||||
void (*clear) (ggml_backend_buffer_t buffer, uint8_t value);
|
|
||||||
};
|
|
||||||
|
|
||||||
struct ggml_backend_buffer {
|
|
||||||
struct ggml_backend_buffer_i iface;
|
|
||||||
ggml_backend_buffer_type_t buft;
|
|
||||||
ggml_backend_buffer_context_t context;
|
|
||||||
size_t size;
|
|
||||||
};
|
|
||||||
|
|
||||||
ggml_backend_buffer_t ggml_backend_buffer_init(
|
|
||||||
ggml_backend_buffer_type_t buft,
|
|
||||||
struct ggml_backend_buffer_i iface,
|
|
||||||
ggml_backend_buffer_context_t context,
|
|
||||||
size_t size);
|
|
||||||
|
|
||||||
|
|
||||||
//
|
|
||||||
// Backend
|
|
||||||
//
|
|
||||||
|
|
||||||
typedef void * ggml_backend_context_t;
|
|
||||||
|
|
||||||
struct ggml_backend_i {
|
|
||||||
const char * (*get_name)(ggml_backend_t backend);
|
|
||||||
|
|
||||||
void (*free)(ggml_backend_t backend);
|
|
||||||
|
|
||||||
// buffer allocation
|
|
||||||
ggml_backend_buffer_type_t (*get_default_buffer_type)(ggml_backend_t backend);
|
|
||||||
|
|
||||||
// (optional) asynchroneous tensor data access
|
|
||||||
void (*set_tensor_async)(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
|
|
||||||
void (*get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size);
|
|
||||||
|
|
||||||
// (optional) asynchroneous tensor copy
|
|
||||||
void (*cpy_tensor_from_async)(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
|
|
||||||
void (*cpy_tensor_to_async) (ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
|
|
||||||
|
|
||||||
void (*synchronize)(ggml_backend_t backend);
|
|
||||||
|
|
||||||
// compute graph with a plan
|
|
||||||
ggml_backend_graph_plan_t (*graph_plan_create) (ggml_backend_t backend, struct ggml_cgraph * cgraph);
|
|
||||||
void (*graph_plan_free) (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
|
|
||||||
void (*graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
|
|
||||||
|
|
||||||
// compute graph without a plan
|
|
||||||
void (*graph_compute)(ggml_backend_t backend, struct ggml_cgraph * cgraph);
|
|
||||||
|
|
||||||
// check if the backend supports an operation
|
|
||||||
bool (*supports_op)(ggml_backend_t backend, const struct ggml_tensor * op);
|
|
||||||
};
|
|
||||||
|
|
||||||
struct ggml_backend {
|
|
||||||
struct ggml_backend_i iface;
|
|
||||||
|
|
||||||
ggml_backend_context_t context;
|
|
||||||
};
|
|
||||||
|
|
||||||
|
|
||||||
//
|
|
||||||
// Backend registry
|
|
||||||
//
|
|
||||||
|
|
||||||
typedef ggml_backend_t (*ggml_backend_init_fn)(const char * params, void * user_data);
|
|
||||||
|
|
||||||
void ggml_backend_register(const char * name, ggml_backend_init_fn init_fn, ggml_backend_buffer_type_t default_buffer_type, void * user_data);
|
|
||||||
|
|
||||||
#ifdef __cplusplus
|
|
||||||
}
|
|
||||||
#endif
|
|
||||||
|
|
@ -1,188 +0,0 @@
|
||||||
#pragma once
|
|
||||||
|
|
||||||
#include "ggml.h"
|
|
||||||
#include "ggml-alloc.h"
|
|
||||||
|
|
||||||
#ifdef __cplusplus
|
|
||||||
extern "C" {
|
|
||||||
#endif
|
|
||||||
|
|
||||||
typedef struct ggml_backend_buffer_type * ggml_backend_buffer_type_t;
|
|
||||||
typedef struct ggml_backend_buffer * ggml_backend_buffer_t;
|
|
||||||
typedef struct ggml_backend * ggml_backend_t;
|
|
||||||
typedef void * ggml_backend_graph_plan_t;
|
|
||||||
|
|
||||||
//
|
|
||||||
// Backend buffer
|
|
||||||
//
|
|
||||||
|
|
||||||
// buffer type
|
|
||||||
GGML_API ggml_backend_buffer_t ggml_backend_buft_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size);
|
|
||||||
GGML_API size_t ggml_backend_buft_get_alignment (ggml_backend_buffer_type_t buft);
|
|
||||||
GGML_API size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor);
|
|
||||||
GGML_API bool ggml_backend_buft_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend);
|
|
||||||
GGML_API bool ggml_backend_buft_is_host (ggml_backend_buffer_type_t buft);
|
|
||||||
|
|
||||||
// buffer
|
|
||||||
GGML_API void ggml_backend_buffer_free (ggml_backend_buffer_t buffer);
|
|
||||||
GGML_API void * ggml_backend_buffer_get_base (ggml_backend_buffer_t buffer);
|
|
||||||
GGML_API size_t ggml_backend_buffer_get_size (ggml_backend_buffer_t buffer);
|
|
||||||
GGML_API void ggml_backend_buffer_init_tensor (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
|
|
||||||
GGML_API size_t ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
|
|
||||||
GGML_API size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
|
|
||||||
GGML_API void ggml_backend_buffer_clear (ggml_backend_buffer_t buffer, uint8_t value);
|
|
||||||
GGML_API bool ggml_backend_buffer_is_host (ggml_backend_buffer_t buffer);
|
|
||||||
GGML_API ggml_backend_buffer_type_t ggml_backend_buffer_type(ggml_backend_buffer_t buffer);
|
|
||||||
|
|
||||||
//
|
|
||||||
// Backend
|
|
||||||
//
|
|
||||||
|
|
||||||
|
|
||||||
GGML_API const char * ggml_backend_name(ggml_backend_t backend);
|
|
||||||
GGML_API void ggml_backend_free(ggml_backend_t backend);
|
|
||||||
|
|
||||||
GGML_API ggml_backend_buffer_type_t ggml_backend_get_default_buffer_type(ggml_backend_t backend);
|
|
||||||
GGML_API ggml_backend_buffer_t ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size);
|
|
||||||
GGML_API size_t ggml_backend_get_alignment(ggml_backend_t backend);
|
|
||||||
|
|
||||||
GGML_API void ggml_backend_tensor_set_async(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
|
|
||||||
GGML_API void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size);
|
|
||||||
|
|
||||||
GGML_API void ggml_backend_tensor_set( struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
|
|
||||||
GGML_API void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size);
|
|
||||||
|
|
||||||
GGML_API void ggml_backend_synchronize(ggml_backend_t backend);
|
|
||||||
|
|
||||||
GGML_API ggml_backend_graph_plan_t ggml_backend_graph_plan_create (ggml_backend_t backend, struct ggml_cgraph * cgraph);
|
|
||||||
|
|
||||||
GGML_API void ggml_backend_graph_plan_free (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
|
|
||||||
GGML_API void ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
|
|
||||||
GGML_API void ggml_backend_graph_compute (ggml_backend_t backend, struct ggml_cgraph * cgraph);
|
|
||||||
GGML_API bool ggml_backend_supports_op (ggml_backend_t backend, const struct ggml_tensor * op);
|
|
||||||
|
|
||||||
// tensor copy between different backends
|
|
||||||
GGML_API void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst);
|
|
||||||
GGML_API void ggml_backend_tensor_copy_async(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst); // automatic fallback to sync copy
|
|
||||||
|
|
||||||
//
|
|
||||||
// CPU backend
|
|
||||||
//
|
|
||||||
|
|
||||||
GGML_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);
|
|
||||||
|
|
||||||
// Create a backend buffer from an existing pointer
|
|
||||||
GGML_API ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size);
|
|
||||||
|
|
||||||
GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void);
|
|
||||||
|
|
||||||
#ifdef GGML_USE_CPU_HBM
|
|
||||||
GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void);
|
|
||||||
#endif
|
|
||||||
|
|
||||||
//
|
|
||||||
// Backend registry
|
|
||||||
//
|
|
||||||
|
|
||||||
// The backend registry is a registry of all the available backends, and allows initializing backends in a generic way
|
|
||||||
|
|
||||||
GGML_API size_t ggml_backend_reg_get_count(void);
|
|
||||||
GGML_API size_t ggml_backend_reg_find_by_name(const char * name);
|
|
||||||
GGML_API ggml_backend_t ggml_backend_reg_init_backend_from_str(const char * backend_str); // str is name[:params]
|
|
||||||
GGML_API const char * ggml_backend_reg_get_name(size_t i);
|
|
||||||
GGML_API ggml_backend_t ggml_backend_reg_init_backend(size_t i, const char * params); // params is backend-specific
|
|
||||||
GGML_API ggml_backend_buffer_type_t ggml_backend_reg_get_default_buffer_type(size_t i);
|
|
||||||
GGML_API ggml_backend_buffer_t ggml_backend_reg_alloc_buffer(size_t i, size_t size);
|
|
||||||
|
|
||||||
//
|
|
||||||
// Backend scheduler
|
|
||||||
//
|
|
||||||
|
|
||||||
// The backend scheduler allows for multiple backends to be used together
|
|
||||||
// Handles compute buffer allocation, assignment of tensors to backends, and copying of tensors between backends
|
|
||||||
// The backends are selected based on:
|
|
||||||
// - the backend that supports the operation
|
|
||||||
// - the location of the pre-allocated tensors (e.g. the weights)
|
|
||||||
/*
|
|
||||||
Example usage:
|
|
||||||
|
|
||||||
sched = ggml_backend_sched_new({backend_gpu, backend_gpu2, backend_cpu}, num_backends);
|
|
||||||
// sched is initialized with measure allocators and cannot be used until allocated with a measure graph
|
|
||||||
|
|
||||||
// initialize buffers from a measure graph
|
|
||||||
measure_graph = build_graph(sched); // use the allocr to allocate inputs as needed
|
|
||||||
|
|
||||||
// in build_graph:
|
|
||||||
build_graph(...) {
|
|
||||||
// allocating tensors in a specific backend (optional, recommended: pre-allocate inputs in a different buffer)
|
|
||||||
alloc_cpu = ggml_backend_sched_get_allocr(sched, backend_cpu);
|
|
||||||
ggml_allocr_alloc(alloc_cpu, tensor);
|
|
||||||
|
|
||||||
// manually assigning nodes to a backend (optional, shouldn't be needed in most cases)
|
|
||||||
struct ggml_tensor * node = ggml_mul_mat(ctx, ...);
|
|
||||||
ggml_backend_sched_set_node_backend(sched, node, backend_gpu);
|
|
||||||
}
|
|
||||||
|
|
||||||
// allocate backend buffers from measure graph
|
|
||||||
ggml_backend_sched_init_measure(sched, measure_graph);
|
|
||||||
|
|
||||||
// the scheduler is now ready to compute graphs
|
|
||||||
|
|
||||||
// compute
|
|
||||||
graph = build_graph(sched);
|
|
||||||
ggml_backend_sched_graph_compute(sched, graph);
|
|
||||||
*/
|
|
||||||
|
|
||||||
struct ggml_backend_sched;
|
|
||||||
typedef struct ggml_backend_sched * ggml_backend_sched_t;
|
|
||||||
|
|
||||||
// Initialize a backend scheduler
|
|
||||||
GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, int n_backends);
|
|
||||||
|
|
||||||
GGML_API void ggml_backend_sched_free(ggml_backend_sched_t sched);
|
|
||||||
|
|
||||||
// Initialize backend buffers from a measure graph
|
|
||||||
GGML_API void ggml_backend_sched_init_measure(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph);
|
|
||||||
|
|
||||||
GGML_API ggml_tallocr_t ggml_backend_sched_get_tallocr(ggml_backend_sched_t sched, ggml_backend_t backend);
|
|
||||||
GGML_API ggml_backend_buffer_t ggml_backend_sched_get_buffer (ggml_backend_sched_t sched, ggml_backend_t backend);
|
|
||||||
|
|
||||||
GGML_API void ggml_backend_sched_set_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend);
|
|
||||||
|
|
||||||
// Allocate a graph on the backend scheduler
|
|
||||||
GGML_API void ggml_backend_sched_graph_compute(
|
|
||||||
ggml_backend_sched_t sched,
|
|
||||||
struct ggml_cgraph * graph);
|
|
||||||
|
|
||||||
|
|
||||||
//
|
|
||||||
// Utils
|
|
||||||
//
|
|
||||||
|
|
||||||
struct ggml_backend_graph_copy {
|
|
||||||
ggml_backend_buffer_t buffer;
|
|
||||||
struct ggml_context * ctx_allocated;
|
|
||||||
struct ggml_context * ctx_unallocated;
|
|
||||||
struct ggml_cgraph * graph;
|
|
||||||
};
|
|
||||||
|
|
||||||
// Copy a graph to a different backend
|
|
||||||
GGML_API struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, struct ggml_cgraph * graph);
|
|
||||||
GGML_API void ggml_backend_graph_copy_free(struct ggml_backend_graph_copy copy);
|
|
||||||
|
|
||||||
typedef bool (*ggml_backend_eval_callback)(int node_index, struct ggml_tensor * t1, struct ggml_tensor * t2, void * user_data);
|
|
||||||
|
|
||||||
// Compare the output of two backends
|
|
||||||
GGML_API void ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data);
|
|
||||||
|
|
||||||
// Tensor initialization
|
|
||||||
GGML_API void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr);
|
|
||||||
GGML_API void ggml_backend_view_init(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
|
|
||||||
|
|
||||||
|
|
||||||
#ifdef __cplusplus
|
|
||||||
}
|
|
||||||
#endif
|
|
||||||
File diff suppressed because it is too large
Load diff
|
|
@ -1,64 +0,0 @@
|
||||||
#pragma once
|
|
||||||
|
|
||||||
#include "ggml.h"
|
|
||||||
#include "ggml-backend.h"
|
|
||||||
|
|
||||||
#ifdef GGML_USE_HIPBLAS
|
|
||||||
#define GGML_CUDA_NAME "ROCm"
|
|
||||||
#define GGML_CUBLAS_NAME "hipBLAS"
|
|
||||||
#else
|
|
||||||
#define GGML_CUDA_NAME "CUDA"
|
|
||||||
#define GGML_CUBLAS_NAME "cuBLAS"
|
|
||||||
#endif
|
|
||||||
|
|
||||||
#ifdef __cplusplus
|
|
||||||
extern "C" {
|
|
||||||
#endif
|
|
||||||
|
|
||||||
#define GGML_CUDA_MAX_DEVICES 16
|
|
||||||
|
|
||||||
// Always success. To check if CUDA is actually loaded, use `ggml_cublas_loaded`.
|
|
||||||
GGML_API void ggml_init_cublas(void);
|
|
||||||
|
|
||||||
// Returns `true` if there are available CUDA devices and cublas loads successfully; otherwise, it returns `false`.
|
|
||||||
GGML_API bool ggml_cublas_loaded(void);
|
|
||||||
|
|
||||||
GGML_API void * ggml_cuda_host_malloc(size_t size);
|
|
||||||
GGML_API void ggml_cuda_host_free(void * ptr);
|
|
||||||
|
|
||||||
GGML_API bool ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
|
|
||||||
GGML_API void ggml_cuda_set_tensor_split(const float * tensor_split);
|
|
||||||
GGML_API void ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor);
|
|
||||||
GGML_API void ggml_cuda_free_data(struct ggml_tensor * tensor);
|
|
||||||
|
|
||||||
GGML_API void ggml_cuda_assign_buffers(struct ggml_tensor * tensor);
|
|
||||||
GGML_API void ggml_cuda_assign_buffers_no_scratch(struct ggml_tensor * tensor);
|
|
||||||
GGML_API void ggml_cuda_assign_buffers_force_inplace(struct ggml_tensor * tensor);
|
|
||||||
|
|
||||||
GGML_API void ggml_cuda_assign_buffers_no_alloc(struct ggml_tensor * tensor);
|
|
||||||
GGML_API void ggml_cuda_assign_scratch_offset(struct ggml_tensor * tensor, size_t offset);
|
|
||||||
GGML_API void ggml_cuda_copy_to_device(struct ggml_tensor * tensor);
|
|
||||||
|
|
||||||
GGML_API void ggml_cuda_set_main_device(int main_device);
|
|
||||||
GGML_API void ggml_cuda_set_mul_mat_q(bool mul_mat_q);
|
|
||||||
GGML_API void ggml_cuda_set_scratch_size(size_t scratch_size);
|
|
||||||
GGML_API void ggml_cuda_free_scratch(void);
|
|
||||||
GGML_API bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor);
|
|
||||||
|
|
||||||
GGML_API int ggml_cuda_get_device_count(void);
|
|
||||||
GGML_API void ggml_cuda_get_device_description(int device, char * description, size_t description_size);
|
|
||||||
|
|
||||||
// backend API
|
|
||||||
GGML_API ggml_backend_t ggml_backend_cuda_init(int device);
|
|
||||||
|
|
||||||
GGML_API bool ggml_backend_is_cuda(ggml_backend_t backend);
|
|
||||||
GGML_API int ggml_backend_cuda_get_device(ggml_backend_t backend);
|
|
||||||
|
|
||||||
GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device);
|
|
||||||
|
|
||||||
// pinned host buffer for use with CPU backend for faster copies between CPU and GPU
|
|
||||||
GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);
|
|
||||||
|
|
||||||
#ifdef __cplusplus
|
|
||||||
}
|
|
||||||
#endif
|
|
||||||
2253
dpcpp_out2/ggml.h
2253
dpcpp_out2/ggml.h
File diff suppressed because it is too large
Load diff
|
|
@ -1,100 +0,0 @@
|
||||||
---
|
|
||||||
MainSourceFile: '/ws1/jianyuzh/ws1/llama.cpp/llama.cpp_pub/dpcpp_out2/ggml.h'
|
|
||||||
Replacements:
|
|
||||||
- FilePath: '/ws1/jianyuzh/ws1/llama.cpp/llama.cpp_pub/ggml.h'
|
|
||||||
Offset: 7458
|
|
||||||
Length: 0
|
|
||||||
ReplacementText: "#define DPCT_PROFILING_ENABLED\n#include <sycl/sycl.hpp>\n#include <dpct/dpct.hpp>\n"
|
|
||||||
ConstantFlag: ''
|
|
||||||
ConstantOffset: 0
|
|
||||||
InitStr: ''
|
|
||||||
NewHostVarName: ''
|
|
||||||
BlockLevelFormatFlag: false
|
|
||||||
- FilePath: '/ws1/jianyuzh/ws1/llama.cpp/llama.cpp_pub/ggml.h'
|
|
||||||
Offset: 10556
|
|
||||||
Length: 10
|
|
||||||
ReplacementText: SYCL_LANGUAGE_VERSION
|
|
||||||
ConstantFlag: ''
|
|
||||||
ConstantOffset: 0
|
|
||||||
InitStr: ''
|
|
||||||
NewHostVarName: ''
|
|
||||||
BlockLevelFormatFlag: false
|
|
||||||
- FilePath: '/ws1/jianyuzh/ws1/llama.cpp/llama.cpp_pub/ggml.h'
|
|
||||||
Offset: 82284
|
|
||||||
Length: 0
|
|
||||||
ReplacementText: ' dpct_type_994041'
|
|
||||||
ConstantFlag: ''
|
|
||||||
ConstantOffset: 0
|
|
||||||
InitStr: ''
|
|
||||||
NewHostVarName: ''
|
|
||||||
BlockLevelFormatFlag: false
|
|
||||||
MainSourceFilesDigest:
|
|
||||||
- MainSourceFile: '/ws1/jianyuzh/ws1/llama.cpp/llama.cpp_pub/ggml.h'
|
|
||||||
Digest: a7f88ed7f3bbff01c9713ad58f5dac5b
|
|
||||||
DpctVersion: 18.0.0
|
|
||||||
MainHelperFileName: ''
|
|
||||||
USMLevel: ''
|
|
||||||
FeatureMap: {}
|
|
||||||
CompileTargets: {}
|
|
||||||
OptionMap:
|
|
||||||
AnalysisScopePath:
|
|
||||||
Value: '/ws1/jianyuzh/ws1/llama.cpp/llama.cpp_pub'
|
|
||||||
Specified: false
|
|
||||||
AsyncHandler:
|
|
||||||
Value: 'false'
|
|
||||||
Specified: false
|
|
||||||
CommentsEnabled:
|
|
||||||
Value: 'false'
|
|
||||||
Specified: false
|
|
||||||
CompilationsDir:
|
|
||||||
Value: ''
|
|
||||||
Specified: false
|
|
||||||
CtadEnabled:
|
|
||||||
Value: 'false'
|
|
||||||
Specified: false
|
|
||||||
EnablepProfiling:
|
|
||||||
Value: 'true'
|
|
||||||
Specified: true
|
|
||||||
ExperimentalFlag:
|
|
||||||
Value: '0'
|
|
||||||
Specified: false
|
|
||||||
ExplicitClNamespace:
|
|
||||||
Value: 'false'
|
|
||||||
Specified: false
|
|
||||||
ExplicitNamespace:
|
|
||||||
Value: '20'
|
|
||||||
Specified: false
|
|
||||||
ExtensionDDFlag:
|
|
||||||
Value: '0'
|
|
||||||
Specified: false
|
|
||||||
ExtensionDEFlag:
|
|
||||||
Value: '4294967295'
|
|
||||||
Specified: false
|
|
||||||
HelperFuncPreferenceFlag:
|
|
||||||
Value: '0'
|
|
||||||
Specified: false
|
|
||||||
NDRangeDim:
|
|
||||||
Value: '3'
|
|
||||||
Specified: false
|
|
||||||
NoDRYPattern:
|
|
||||||
Value: 'false'
|
|
||||||
Specified: false
|
|
||||||
NoUseGenericSpace:
|
|
||||||
Value: ''
|
|
||||||
Specified: true
|
|
||||||
OptimizeMigration:
|
|
||||||
Value: 'false'
|
|
||||||
Specified: false
|
|
||||||
ProcessAll:
|
|
||||||
Value: 'false'
|
|
||||||
Specified: false
|
|
||||||
RuleFile:
|
|
||||||
Value: ''
|
|
||||||
Specified: false
|
|
||||||
SyclNamedLambda:
|
|
||||||
Value: 'false'
|
|
||||||
Specified: false
|
|
||||||
UsmLevel:
|
|
||||||
Value: '1'
|
|
||||||
Specified: false
|
|
||||||
...
|
|
||||||
|
|
@ -19,10 +19,7 @@
|
||||||
|
|
||||||
#include <sycl/sycl.hpp>
|
#include <sycl/sycl.hpp>
|
||||||
#include <sycl/half_type.hpp>
|
#include <sycl/half_type.hpp>
|
||||||
// #include <dpct/dpct.hpp>
|
|
||||||
// #include <dpct/blas_utils.hpp>
|
|
||||||
// #include <dpct/lib_common_utils.hpp>
|
|
||||||
#include "dpct.hpp"
|
|
||||||
#include "ggml-sycl.h"
|
#include "ggml-sycl.h"
|
||||||
#include "ggml.h"
|
#include "ggml.h"
|
||||||
#include "ggml-backend-impl.h"
|
#include "ggml-backend-impl.h"
|
||||||
|
|
|
||||||
2807
ggml-sycl.h
2807
ggml-sycl.h
File diff suppressed because it is too large
Load diff
Loading…
Add table
Add a link
Reference in a new issue