vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

ggml-kompute.h : remove anything that doesn't need to be public

Browse source 
The remaining functions are either used by llama.cpp or GPT4All.
This commit is contained in:
Jared Van Bortel 2024-01-26 14:57:58 -05:00
parent 6af02b19d1
commit 2ff2d16131
2 changed files with 13 additions and 25 deletions
Download patch file Download diff file Expand all files Collapse all files

17
ggml-kompute.cpp
View file

@ -85,6 +85,15 @@ public:
static kompute_manager komputeManager; static kompute_manager komputeManager;
struct ggml_vk_memory {
void *data = nullptr;
size_t size = 0;
vk::DeviceMemory *primaryMemory = nullptr;
vk::Buffer *primaryBuffer = nullptr;
vk::DeviceMemory *stagingMemory = nullptr;
vk::Buffer *stagingBuffer = nullptr;
};
#ifdef __linux__ #ifdef __linux__
__attribute__((constructor)) __attribute__((constructor))
static void enable_sam() { static void enable_sam() {
@ -302,13 +311,13 @@ ggml_vk_device ggml_vk_current_device() {
return devices.front(); return devices.front();
} }
ggml_kompute_context *ggml_vk_init() { static ggml_kompute_context * ggml_vk_init() {
GGML_ASSERT(s_kompute_context == nullptr); GGML_ASSERT(s_kompute_context == nullptr);
s_kompute_context = new ggml_kompute_context; s_kompute_context = new ggml_kompute_context;
return s_kompute_context; return s_kompute_context;
} }
void ggml_vk_free(struct ggml_kompute_context * ctx) { static void ggml_vk_free(struct ggml_kompute_context * ctx) {
assert(ctx == s_kompute_context); assert(ctx == s_kompute_context);
s_kompute_context = nullptr; s_kompute_context = nullptr;
if (ctx != nullptr) { if (ctx != nullptr) {
@ -457,7 +466,7 @@ static ggml_vk_memory ggml_vk_allocate(size_t size) {
return memory; return memory;
} }
void ggml_vk_free_memory(ggml_vk_memory &memory) static void ggml_vk_free_memory(ggml_vk_memory &memory)
{ {
komputeManager()->device()->destroy( komputeManager()->device()->destroy(
*memory.primaryBuffer, *memory.primaryBuffer,
@ -1376,7 +1385,7 @@ static bool ggml_vk_supports_op(const struct ggml_tensor * op) {
return false; return false;
} }
void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml_cgraph * gf) { static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml_cgraph * gf) {
const int n_seq = 8; const int n_seq = 8;
// FIXME: Figure out if we can somehow optimize the size of the pool... right now we're setting // FIXME: Figure out if we can somehow optimize the size of the pool... right now we're setting

21
ggml-kompute.h
View file

@ -6,22 +6,6 @@
#include <vector> #include <vector>
#include <string> #include <string>
struct ggml_kompute_context;
namespace vk {
class DeviceMemory;
class Buffer;
};
struct ggml_vk_memory {
void *data = nullptr;
size_t size = 0;
vk::DeviceMemory *primaryMemory = nullptr;
vk::Buffer *primaryBuffer = nullptr;
vk::DeviceMemory *stagingMemory = nullptr;
vk::Buffer *stagingBuffer = nullptr;
};
struct ggml_vk_device { struct ggml_vk_device {
int index = 0; int index = 0;
int type = 0; // same as VkPhysicalDeviceType int type = 0; // same as VkPhysicalDeviceType
@ -40,11 +24,6 @@ bool ggml_vk_has_vulkan();
bool ggml_vk_has_device(); bool ggml_vk_has_device();
bool ggml_vk_using_vulkan(); bool ggml_vk_using_vulkan();
ggml_vk_device ggml_vk_current_device(); ggml_vk_device ggml_vk_current_device();
struct ggml_kompute_context * ggml_vk_init(void);
void ggml_vk_free(struct ggml_kompute_context * ctx);
void ggml_vk_free_memory(ggml_vk_memory &memory);
void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml_cgraph * gf);
// //
// backend API // backend API