ExternalVendorCode/whisper.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/whisper.cpp.git synced 2025-06-14 21:08:07 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

ggml : introduce GGML_CALL function annotation (llama/4850)

Browse Source 
This change makes it possible to build ggml-cuda.cu and ggml-metal.m as
independent dynamic shared objects, that may be conditionally linked at
runtime in a multiplatform binary. It introduces a GGML_CALL annotation
that documents which functions have a cyclic call relationship, between
the application code and GPU modules.

This change does nothing, unless the build defines -DGGML_MULTIPLATFORM
which causes back-references and function pointers to conform to MS ABI
which is supported by NVCC, ROCm, XCode, GCC and Clang across platforms
This commit is contained in:
Justine Tunney
2024-01-16 03:16:33 -08:00
committed by Georgi Gerganov
parent 61b9192f27
commit 138eaebead
9 changed files with 244 additions and 235 deletions
Download Patch File Download Diff File Expand all files Collapse all files

80
ggml-backend.c
View File

@ -19,7 +19,7 @@ const char * ggml_backend_buft_name(ggml_backend_buffer_type_t buft) {
return buft->iface.get_name(buft);
}
ggml_backend_buffer_t ggml_backend_buft_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
GGML_CALL ggml_backend_buffer_t ggml_backend_buft_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
return buft->iface.alloc_buffer(buft, size);
}
@ -27,7 +27,7 @@ size_t ggml_backend_buft_get_alignment(ggml_backend_buffer_type_t buft) {
return buft->iface.get_alignment(buft);
}
size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor) {
GGML_CALL size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor) {
// get_alloc_size is optional, defaults to ggml_nbytes
if (buft->iface.get_alloc_size) {
return buft->iface.get_alloc_size(buft, tensor);
@ -48,7 +48,7 @@ bool ggml_backend_buft_is_host(ggml_backend_buffer_type_t buft) {
// backend buffer
ggml_backend_buffer_t ggml_backend_buffer_init(
GGML_CALL 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,
@ -95,7 +95,7 @@ void * ggml_backend_buffer_get_base(ggml_backend_buffer_t buffer) {
return base;
}
void ggml_backend_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
GGML_CALL void ggml_backend_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
// init_tensor is optional
if (buffer->iface.init_tensor) {
buffer->iface.init_tensor(buffer, tensor);
@ -191,7 +191,7 @@ void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_ten
}
}
void ggml_backend_tensor_set(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
GGML_CALL void ggml_backend_tensor_set(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
@ -201,7 +201,7 @@ void ggml_backend_tensor_set(struct ggml_tensor * tensor, const void * data, siz
tensor->buffer->iface.set_tensor(buf, tensor, data, offset, size);
}
void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
GGML_CALL void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
@ -318,9 +318,9 @@ struct ggml_backend_reg {
static struct ggml_backend_reg ggml_backend_registry[GGML_MAX_BACKENDS_REG];
static size_t ggml_backend_registry_count = 0;
static ggml_backend_t ggml_backend_reg_cpu_init(const char * params, void * user_data);
GGML_CALL static ggml_backend_t ggml_backend_reg_cpu_init(const char * params, void * user_data);
static void ggml_backend_registry_init(void) {
GGML_CALL static void ggml_backend_registry_init(void) {
static bool initialized = false;
if (initialized) {
@ -333,18 +333,18 @@ static void ggml_backend_registry_init(void) {
// add forward decls here to avoid including the backend headers
#ifdef GGML_USE_CUBLAS
extern void ggml_backend_cuda_reg_devices(void);
extern GGML_CALL void ggml_backend_cuda_reg_devices(void);
ggml_backend_cuda_reg_devices();
#endif
#ifdef GGML_USE_METAL
extern ggml_backend_t ggml_backend_reg_metal_init(const char * params, void * user_data);
extern ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
extern GGML_CALL ggml_backend_t ggml_backend_reg_metal_init(const char * params, void * user_data);
extern GGML_CALL ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
ggml_backend_register("Metal", ggml_backend_reg_metal_init, ggml_backend_metal_buffer_type(), NULL);
#endif
}
void ggml_backend_register(const char * name, ggml_backend_init_fn init_fn, ggml_backend_buffer_type_t default_buffer_type, void * user_data) {
GGML_CALL void ggml_backend_register(const char * name, ggml_backend_init_fn init_fn, ggml_backend_buffer_type_t default_buffer_type, void * user_data) {
GGML_ASSERT(ggml_backend_registry_count < GGML_MAX_BACKENDS_REG);
size_t id = ggml_backend_registry_count;
@ -439,33 +439,33 @@ ggml_backend_buffer_t ggml_backend_reg_alloc_buffer(size_t i, size_t size) {
// backend CPU
static const char * ggml_backend_cpu_buffer_name(ggml_backend_buffer_t buffer) {
GGML_CALL static const char * ggml_backend_cpu_buffer_name(ggml_backend_buffer_t buffer) {
return "CPU";
GGML_UNUSED(buffer);
}
static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) {
GGML_CALL static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) {
return (void *)buffer->context;
}
static void ggml_backend_cpu_buffer_free_buffer(ggml_backend_buffer_t buffer) {
GGML_CALL static void ggml_backend_cpu_buffer_free_buffer(ggml_backend_buffer_t buffer) {
free(buffer->context);
}
static void ggml_backend_cpu_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
GGML_CALL static void ggml_backend_cpu_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
memcpy((char *)tensor->data + offset, data, size);
GGML_UNUSED(buffer);
}
static void ggml_backend_cpu_buffer_get_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
GGML_CALL static void ggml_backend_cpu_buffer_get_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
memcpy(data, (const char *)tensor->data + offset, size);
GGML_UNUSED(buffer);
}
static bool ggml_backend_cpu_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst) {
GGML_CALL static bool ggml_backend_cpu_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)) {
memcpy(dst->data, src->data, ggml_nbytes(src));
return true;
@ -475,7 +475,7 @@ static bool ggml_backend_cpu_buffer_cpy_tensor(ggml_backend_buffer_t buffer, con
GGML_UNUSED(buffer);
}
static void ggml_backend_cpu_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
GGML_CALL static void ggml_backend_cpu_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
memset(buffer->context, value, buffer->size);
}
@ -506,13 +506,13 @@ static struct ggml_backend_buffer_i cpu_backend_buffer_i_from_ptr = {
static const size_t TENSOR_ALIGNMENT = 64; // should be enough for AVX 512
static const char * ggml_backend_cpu_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
GGML_CALL static const char * ggml_backend_cpu_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
return "CPU";
GGML_UNUSED(buft);
}
static ggml_backend_buffer_t ggml_backend_cpu_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
GGML_CALL static ggml_backend_buffer_t ggml_backend_cpu_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
size += TENSOR_ALIGNMENT; // malloc may return an address that is not aligned
void * data = malloc(size); // TODO: maybe use GGML_ALIGNED_MALLOC?
@ -521,25 +521,25 @@ static ggml_backend_buffer_t ggml_backend_cpu_buffer_type_alloc_buffer(ggml_back
return ggml_backend_buffer_init(buft, cpu_backend_buffer_i, data, size);
}
static size_t ggml_backend_cpu_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
GGML_CALL static size_t ggml_backend_cpu_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
return TENSOR_ALIGNMENT;
GGML_UNUSED(buft);
}
static bool ggml_backend_cpu_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) {
GGML_CALL static bool ggml_backend_cpu_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) {
return ggml_backend_is_cpu(backend);
GGML_UNUSED(buft);
}
static bool ggml_backend_cpu_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
GGML_CALL static bool ggml_backend_cpu_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
return true;
GGML_UNUSED(buft);
}
ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void) {
GGML_CALL ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void) {
static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type = {
/* .iface = */ {
/* .get_name = */ ggml_backend_cpu_buffer_type_get_name,
@ -561,23 +561,23 @@ ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void) {
#include <hbwmalloc.h>
static const char * ggml_backend_cpu_hbm_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
GGML_CALL static const char * ggml_backend_cpu_hbm_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
return "CPU_HBM";
GGML_UNUSED(buft);
}
static const char * ggml_backend_cpu_hbm_buffer_get_name(ggml_backend_buffer_t buf) {
GGML_CALL static const char * ggml_backend_cpu_hbm_buffer_get_name(ggml_backend_buffer_t buf) {
return "CPU_HBM";
GGML_UNUSED(buf);
}
static void ggml_backend_cpu_hbm_buffer_free_buffer(ggml_backend_buffer_t buffer) {
GGML_CALL 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) {
GGML_CALL static ggml_backend_buffer_t ggml_backend_cpu_hbm_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
//void * ptr = hbw_malloc(size);
void * ptr;
int result = hbw_posix_memalign(&ptr, ggml_backend_cpu_buffer_type_get_alignment(buft), size);
@ -617,20 +617,20 @@ struct ggml_backend_cpu_context {
size_t work_size;
};
static const char * ggml_backend_cpu_name(ggml_backend_t backend) {
GGML_CALL static const char * ggml_backend_cpu_name(ggml_backend_t backend) {
return "CPU";
GGML_UNUSED(backend);
}
static void ggml_backend_cpu_free(ggml_backend_t backend) {
GGML_CALL static void ggml_backend_cpu_free(ggml_backend_t backend) {
struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
free(cpu_ctx->work_data);
free(cpu_ctx);
free(backend);
}
static ggml_backend_buffer_type_t ggml_backend_cpu_get_default_buffer_type(ggml_backend_t backend) {
GGML_CALL static ggml_backend_buffer_type_t ggml_backend_cpu_get_default_buffer_type(ggml_backend_t backend) {
return ggml_backend_cpu_buffer_type();
GGML_UNUSED(backend);
@ -641,7 +641,7 @@ struct ggml_backend_plan_cpu {
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) {
GGML_CALL 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 = malloc(sizeof(struct ggml_backend_plan_cpu));
@ -656,7 +656,7 @@ static ggml_backend_graph_plan_t ggml_backend_cpu_graph_plan_create(ggml_backend
return cpu_plan;
}
static void ggml_backend_cpu_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
GGML_CALL 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;
free(cpu_plan->cplan.work_data);
@ -665,7 +665,7 @@ static void ggml_backend_cpu_graph_plan_free(ggml_backend_t backend, ggml_backen
GGML_UNUSED(backend);
}
static void ggml_backend_cpu_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
GGML_CALL static void 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;
ggml_graph_compute(&cpu_plan->cgraph, &cpu_plan->cplan);
@ -673,7 +673,7 @@ static void ggml_backend_cpu_graph_plan_compute(ggml_backend_t backend, ggml_bac
GGML_UNUSED(backend);
}
static bool ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
GGML_CALL static bool 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);
@ -690,7 +690,7 @@ static bool ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_c
return true;
}
static bool ggml_backend_cpu_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
GGML_CALL static bool ggml_backend_cpu_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
switch (op->op) {
case GGML_OP_MUL_MAT:
return op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == ggml_internal_get_type_traits(op->src[0]->type).vec_dot_type;
@ -732,7 +732,7 @@ ggml_backend_t ggml_backend_cpu_init(void) {
return cpu_backend;
}
bool ggml_backend_is_cpu(ggml_backend_t backend) {
GGML_CALL bool ggml_backend_is_cpu(ggml_backend_t backend) {
return backend && backend->iface.get_name == ggml_backend_cpu_name;
}
@ -743,11 +743,11 @@ void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads) {
ctx->n_threads = n_threads;
}
ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size) {
GGML_CALL ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size) {
return ggml_backend_buffer_init(ggml_backend_cpu_buffer_type(), cpu_backend_buffer_i_from_ptr, ptr, size);
}
static ggml_backend_t ggml_backend_reg_cpu_init(const char * params, void * user_data) {
GGML_CALL static ggml_backend_t ggml_backend_reg_cpu_init(const char * params, void * user_data) {
return ggml_backend_cpu_init();
GGML_UNUSED(params);