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

cuBLAS memory management routines

Browse source
This commit is contained in:
John 2023-04-24 04:26:40 +02:00
parent 7fcfba2e28
commit 6e292d70a5
1 changed files with 36 additions and 13 deletions
Download patch file Download diff file Expand all files Collapse all files

49
ggml-cuda.cu
View file

@ -6,7 +6,7 @@
typedef uint16_t ggml_fp16_t;
static_assert(sizeof(__half) == sizeof(ggml_fp16_t), "wrong fp16 size");
// #define CUDA_MEM_DEBUG 1
#define CUDA_MEM_DEBUG 0
#define QK4_0 32
typedef struct {
@ -181,6 +181,19 @@ static int g_cuda_free_buffer_indices[MAX_CUDA_BUFFERS]; // sorted list of free
static int g_cuda_free_buffer_count = 0;
static size_t g_cuda_pool_total_allocated = 0;
void cuda_pool_dump() {
printf("========================================\n");
printf("| Current CUDA Buffer Pool |\n");
printf("========================================\n");
printf("| %-6s | %-12s | %-10s |\n", "Index", "Buffer Index", "Size (bytes)");
printf("----------------------------------------\n");
for (int i = 0; i < g_cuda_free_buffer_count; ++i) {
int buffer_index = g_cuda_free_buffer_indices[i];
printf("| %-6d | %-12d | %-10zu |\n",
i, buffer_index, g_cuda_buffer_pool[buffer_index].size);
}
printf("========================================\n");
}
void cuda_pool_insert_idx(int index) {
int left = 0;
@ -188,7 +201,7 @@ void cuda_pool_insert_idx(int index) {
while (left <= right) {
int mid = (left + right) / 2;
if (g_cuda_buffer_pool[g_cuda_free_buffer_indices[mid]].size <= g_cuda_buffer_pool[index].size) {
if (g_cuda_buffer_pool[g_cuda_free_buffer_indices[mid]].size < g_cuda_buffer_pool[index].size) {
left = mid + 1;
} else {
right = mid - 1;
@ -202,8 +215,9 @@ void cuda_pool_insert_idx(int index) {
g_cuda_free_buffer_indices[left] = index;
g_cuda_free_buffer_count++;
#ifdef CUDA_MEM_DEBUG
printf("INFO:: Inserted buffer at index %d with size %d bytes\n", index, g_cuda_buffer_pool[index].size);
#if CUDA_MEM_DEBUG
printf("INFO:: Inserted buffer at index %d with size %lld bytes\n", index, g_cuda_buffer_pool[index].size);
cuda_pool_dump();
#endif
}
// find the best fitting block in the pool and remove it from the indexed list
@ -216,7 +230,11 @@ int cuda_pool_get_block(size_t size, size_t * actual_size) {
int mid = (left + right) / 2;
if (g_cuda_buffer_pool[g_cuda_free_buffer_indices[mid]].size >= size) {
index = mid;
right = mid - 1;
if (mid > 0 && g_cuda_buffer_pool[g_cuda_free_buffer_indices[mid - 1]].size >= size) {
right = mid - 1; // continue searching to the left for smaller fitting buffers
} else {
break; // found the smallest fitting buffer
}
} else {
left = mid + 1;
}
@ -231,17 +249,20 @@ int cuda_pool_get_block(size_t size, size_t * actual_size) {
g_cuda_free_buffer_indices[i] = g_cuda_free_buffer_indices[i + 1];
}
g_cuda_free_buffer_count--;
#ifdef CUDA_MEM_DEBUG
printf("INFO:: Found buffer of size %d bytes at index %d\n", *actual_size, buffer_index);
#if CUDA_MEM_DEBUG
printf("INFO:: Found buffer of size %lld bytes at index %d\n", *actual_size, buffer_index);
cuda_pool_dump();
#endif
return buffer_index;
}
#ifdef CUDA_MEM_DEBUG
printf("INFO:: No buffer found for size %d bytes\n", size);
#if CUDA_MEM_DEBUG
printf("INFO:: No buffer found for size %lld bytes\n", size);
cuda_pool_dump();
#endif
return -1;
}
// mark all buffers as free
void ggml_cuda_pool_initialize() {
for (int i = 0; i < MAX_CUDA_BUFFERS; ++i) {
@ -262,8 +283,9 @@ void * ggml_cuda_pool_malloc(size_t size, size_t * actual_size) {
g_cuda_buffer_pool[buffer_index].ptr = nullptr;
g_cuda_buffer_pool[buffer_index].size = 0;
g_cuda_pool_total_allocated -= *actual_size;
#ifdef CUDA_MEM_DEBUG
printf("INFO:: Allocated %d bytes from buffer at index %d (total allocated: %d bytes)\n", *actual_size, buffer_index, g_cuda_pool_total_allocated);
#if CUDA_MEM_DEBUG
printf("INFO:: Allocated %lld bytes from buffer at index %d (total allocated: %lld bytes)\n", *actual_size, buffer_index, g_cuda_pool_total_allocated);
cuda_pool_dump();
#endif
return ptr;
}
@ -274,8 +296,9 @@ void * ggml_cuda_pool_malloc(size_t size, size_t * actual_size) {
CUDA_CHECK(cudaMalloc((void **) &ptr, size));
*actual_size = size;
g_cuda_pool_total_allocated += size;
#ifdef CUDA_MEM_DEBUG
printf("INFO:: Allocated %d bytes from cudaMalloc (total allocated: %d bytes)\n", *actual_size, g_cuda_pool_total_allocated);
#if CUDA_MEM_DEBUG
printf("INFO:: Allocated %lld bytes from cudaMalloc (total allocated: %lld bytes)\n", *actual_size, g_cuda_pool_total_allocated);
cuda_pool_dump();
#endif
return ptr;
}