about ggml backend
About backend, memory allocation, compute graph, compute process. all ref the simple-backend.cpp.
log
Need register log callback, and use macro like GGML_DEBUG to print log.
device
device means what select decide to run the ggml. device init in the function like ggml_backend_cpu_init. To init devide, user need do like:
cpp
if(backend == cpu){
ggml_backend_cpu_init();
}else if(backend == gpu){
ggml_backend_gpu_init();
}else{
// like this
}In the cpu:
cpp
ggml_backend_t ggml_backend_cpu_init(void) {
// initialize CPU backend now to avoid slowing the first graph computation
// This just allocate FP32 to FP16 Table
ggml_cpu_init();
struct ggml_backend_cpu_context * ctx = new ggml_backend_cpu_context;
if (ctx == NULL) {
return NULL;
}
// use how many thread to computer
ctx->n_threads = GGML_DEFAULT_N_THREADS;
// the threadpool
ctx->threadpool = NULL;
// this is a point to buffer, this buffer save the intermedicate
// variable in the computer process.
ctx->work_data = NULL;
// the buffer size
ctx->work_size = 0;
// the callback of compute finish
ctx->abort_callback = NULL;
ctx->abort_callback_data = NULL;
ggml_backend_t cpu_backend = new ggml_backend {
/* .guid = */ ggml_backend_cpu_guid(),
// this point to the cpu interface
/* .interface = */ ggml_backend_cpu_i,
// this point to the cpu device
/* .device = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
/* .context = */ ctx,
};
if (cpu_backend == NULL) {
delete ctx;
return NULL;
}
return cpu_backend;
}ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0) this point to the device. device contains all the message about cpu.
cpp
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,
};
// this function just return the registered device.
ggml_backend_reg_t ggml_backend_cpu_reg(void) {
// init CPU feature detection
ggml_cpu_init();
static struct ggml_backend_reg ggml_backend_cpu_reg = {
/* .api_version = */ GGML_BACKEND_API_VERSION,
/* .iface = */ ggml_backend_cpu_reg_i,
/* .context = */ NULL,
};
return &ggml_backend_cpu_reg;
}What important in cpu device is ggml_backend_cpu_get_proc_address, this contains the many extra information about cpu, this function can return the function pointer, user can call this function pointer to get some information.
cpp
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) {
ggml_backend_set_n_threads_t fct = ggml_backend_cpu_set_n_threads;
return (void *)fct;
}
if (strcmp(name, "ggml_backend_dev_get_extra_bufts") == 0) {
// this function in cpu is important, the extra matrix extension should use this
ggml_backend_dev_get_extra_bufts_t fct = ggml_backend_cpu_device_get_extra_buffers_type;
return (void *)fct;
}
if (strcmp(name, "ggml_backend_get_features") == 0) {
return (void *)ggml_backend_cpu_get_features;
}
if (strcmp(name, "ggml_backend_set_abort_callback") == 0) {
return (void *)ggml_backend_cpu_set_abort_callback;
}
if (strcmp(name, "ggml_backend_cpu_numa_init") == 0) {
return (void *)ggml_numa_init;
}
if (strcmp(name, "ggml_backend_cpu_is_numa") == 0) {
return (void *)ggml_is_numa;
}
// threadpool - TODO: move to ggml-base
if (strcmp(name, "ggml_threadpool_new") == 0) {
return (void *)ggml_threadpool_new;
}
if (strcmp(name, "ggml_threadpool_free") == 0) {
return (void *)ggml_threadpool_free;
}
if (strcmp(name, "ggml_backend_cpu_set_threadpool") == 0) {
return (void *)ggml_backend_cpu_set_threadpool;
}
return NULL;
GGML_UNUSED(reg);
}all of that is about devide init.
global ggml context
after the device registered, we should create global ggml context.
cpp
struct ggml_init_params params {
/*.mem_size =*/ ggml_tensor_overhead() * num_tensors,
/*.mem_buffer =*/ NULL,
/*.no_alloc =*/ true,
};
// create context
model.ctx = ggml_init(params);ggml context don't have specific meaning in ggml, it just use to allocate some memory, so the ggml init will be call in some memory allocation case. The mem_size means the size the memory allocate want, no_alloc true means backend devide allocate the real buffer, false means ggml allocate real buffer.
new tensor object
cpp
// create tensors
model.a = ggml_new_tensor_2d(model.ctx, GGML_TYPE_F32, cols_A, rows_A);
model.b = ggml_new_tensor_2d(model.ctx, GGML_TYPE_F32, cols_B, rows_B);this just struct to save some config about tensor, no memory allocate.
tensor memory allocate
cpp
// create a backend buffer (backend memory) and alloc the tensors from the context
model.buffer = ggml_backend_alloc_ctx_tensors(model.ctx, model.backend);this will call:
cpp
ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors(struct ggml_context * ctx, ggml_backend_t backend) {
return ggml_backend_alloc_ctx_tensors_from_buft(ctx, ggml_backend_get_default_buffer_type(backend));
}the ggml_backend_buffer_t meneage the backend device memory. Three have ggml_backend_get_default_buffer_type to determine with backend buffer type want to use. The most important is, if you want to use intel-amx and other cpu specific matrix extensions, you must let ggml_backend_get_default_buffer_type to be the extension specific buffer.
ggml_backend_get_default_buffer_type this use device interface to return the buffer type:
cpp
struct ggml_backend_buffer_type_i {
const char * (*get_name) (ggml_backend_buffer_type_t buft);
// allocate a buffer of this type
ggml_backend_buffer_t (*alloc_buffer) (ggml_backend_buffer_type_t buft, size_t size);
// tensor alignment
size_t (*get_alignment) (ggml_backend_buffer_type_t buft);
// (optional) max buffer size that can be allocated (defaults to SIZE_MAX)
size_t (*get_max_size) (ggml_backend_buffer_type_t buft);
// (optional) data size needed to allocate the tensor, including padding (defaults to ggml_nbytes)
size_t (*get_alloc_size)(ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor);
// (optional) check if tensor data is in host memory and uses standard ggml tensor layout (defaults to false)
bool (*is_host) (ggml_backend_buffer_type_t buft);
};
struct ggml_backend_buffer_type {
struct ggml_backend_buffer_type_i iface;
ggml_backend_dev_t device;
void * context;
};the ggml_backend_buffer_type_i seems like a backend buffer allocator, the memory allocate by ggml_backend_buffer_type_i, and operate in the ggml buffer type. The cpu extension implementer must care about this.
cpp
ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, ggml_backend_buffer_type_t buft) {
GGML_ASSERT(ggml_get_no_alloc(ctx) == true);
size_t alignment = ggml_backend_buft_get_alignment(buft);
// see this, the max_size of the backend buffer defined by backend
size_t max_size = ggml_backend_buft_get_max_size(buft);
ggml_backend_buffer_t * buffers = NULL;
size_t n_buffers = 0;
size_t cur_buf_size = 0;
struct ggml_tensor * first = ggml_get_first_tensor(ctx);
// access all the tensors in ggml context
// it seems like the same ggml context can only do that once
for (struct ggml_tensor * t = first; t != NULL; t = ggml_get_next_tensor(ctx, t)) {
size_t this_size = 0;
// calculate this tensor need how many size
if (t->data == NULL && t->view_src == NULL) {
this_size = GGML_PAD(ggml_backend_buft_get_alloc_size(buft, t), alignment);
}
if (cur_buf_size > 0 && (cur_buf_size + this_size) > max_size) {
// per buffer size need < max buffer size
// allocate tensors in the current buffer
// the alloc_tensor_range allocate the memory
if (!alloc_tensor_range(ctx, first, t, buft, cur_buf_size, &buffers, &n_buffers)) {
return NULL;
}
first = t;
cur_buf_size = this_size;
} else {
cur_buf_size += this_size;
}
}
// allocate remaining tensors
if (cur_buf_size > 0) {
if (!alloc_tensor_range(ctx, first, NULL, buft, cur_buf_size, &buffers, &n_buffers)) {
return NULL;
}
}
if (n_buffers == 0) {
#ifndef NDEBUG
GGML_LOG_DEBUG("%s: all tensors in the context are already allocated\n", __func__);
#endif
return NULL;
}
ggml_backend_buffer_t buffer;
if (n_buffers == 1) {
buffer = buffers[0];
} else {
// n_buffers copy to ctx
// this also copy
buffer = ggml_backend_multi_buffer_alloc_buffer(buffers, n_buffers);
}
free(buffers);
return buffer;
}alloc_tensor_range
cpp
static bool alloc_tensor_range(struct ggml_context * ctx,
struct ggml_tensor * first, struct ggml_tensor * last,
ggml_backend_buffer_type_t buft, size_t size,
ggml_backend_buffer_t ** buffers, size_t * n_buffers) {
// the allocate backend buffer in there
// use the buffer type method to allocate buffer
ggml_backend_buffer_t buffer = ggml_backend_buft_alloc_buffer(buft, size);
if (buffer == NULL) {
GGML_LOG_ERROR("%s: failed to allocate %s buffer of size %zu\n", __func__, ggml_backend_buft_name(buft), size);
free_buffers(buffers, n_buffers);
return false;
}
// add buffer to buffers array
*buffers = realloc(*buffers, sizeof(ggml_backend_buffer_t) * (*n_buffers + 1));
(*buffers)[(*n_buffers)++] = buffer;
// simplely thinking, tallocr use ggml_backend_buffer_t to allocate memory for tensor
struct ggml_tallocr tallocr = ggml_tallocr_new(buffer);
for (struct ggml_tensor * t = first; t != last; t = ggml_get_next_tensor(ctx, t)) {
enum ggml_status status = GGML_STATUS_SUCCESS;
if (t->data == NULL) {
if (t->view_src == NULL) {
// no src, true allocate
status = ggml_tallocr_alloc(&tallocr, t);
} else if (t->buffer == NULL) {
// have src, just a view
status = ggml_backend_view_init(t);
}
} else {
if (t->view_src != NULL && t->buffer == NULL) {
// view of a pre-allocated tensor
status = ggml_backend_view_init(t);
}
}
if (status != GGML_STATUS_SUCCESS) {
GGML_LOG_ERROR("%s: failed to initialize tensor %s\n", __func__, t->name);
free_buffers(buffers, n_buffers);
return false;
}
}
return true;
}also should care about that, both new and view will all call the init interface in ggml_backend_buffer_t.
cpp
ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer_t * buffers, size_t n_buffers) {
ggml_backend_multi_buffer_context * ctx = (ggml_backend_multi_buffer_context *) malloc(sizeof(struct ggml_backend_multi_buffer_context));
ctx->n_buffers = n_buffers;
ctx->buffers = (ggml_backend_buffer_t *) malloc(n_buffers * sizeof(ggml_backend_buffer_t));
GGML_ASSERT(ctx->buffers != NULL);
size_t total_size = 0;
for (size_t i = 0; i < n_buffers; i++) {
ctx->buffers[i] = buffers[i];
total_size += ggml_backend_buffer_get_size(buffers[i]);
}
return ggml_backend_buffer_init(buffers[0]->buft, ggml_backend_multi_buffer_i, ctx, total_size);
}if more than one backend buffer use, thay will be save in ggml_backend_multi_buffer_context, return the ggml_backend_buffer_t, this ggml_backend_buffer_t merge this message.
copy tensor to backend
cpp
// load data from cpu memory to backend buffer
ggml_backend_tensor_set(model.a, a, 0, ggml_nbytes(model.a));
ggml_backend_tensor_set(model.b, b, 0, ggml_nbytes(model.b));this copy the data from host to backend buffer, this also will call the ggml_backend_buffer_t set tensor function.
create graph
cpp
// calculate the temporaly memory required to compute
// ggml_gallocr_t is a data struct contains some allocator and graph message
// Gragh allocator
ggml_gallocr_t allocr = NULL;
{
// use glloc to manage buffer
allocr = ggml_gallocr_new(ggml_backend_get_default_buffer_type(model.backend));
// create the worst case graph for memory usage estimation
struct ggml_cgraph * gf = build_graph(model);
// allocate memory
ggml_gallocr_reserve(allocr, gf);
size_t mem_size = ggml_gallocr_get_buffer_size(allocr, 0);
fprintf(stderr, "%s: compute buffer size: %.4f KB\n", __func__, mem_size/1024.0);
}build graph process
cpp
struct ggml_cgraph * build_graph(const simple_model& model) {
static size_t buf_size = ggml_tensor_overhead()*GGML_DEFAULT_GRAPH_SIZE + ggml_graph_overhead();
static std::vector<uint8_t> buf(buf_size);
struct ggml_init_params params0 = {
/*.mem_size =*/ buf_size,
/*.mem_buffer =*/ buf.data(),
/*.no_alloc =*/ true, // the tensors will be allocated later by ggml_allocr_alloc_graph()
};
// create a temporally context to build the graph
struct ggml_context * ctx0 = ggml_init(params0);
struct ggml_cgraph * gf = ggml_new_graph(ctx0);
// result = a*b^T
struct ggml_tensor * result = ggml_mul_mat(ctx0, model.a, model.b);
// build operations nodes
ggml_build_forward_expand(gf, result);
// delete the temporally context used to build the graph
ggml_free(ctx0);
return gf;
}we can see that ggml_context was create again, but it just a temp variable. ggml_context does not have specical meaning in ggml, if you want allocate some memory for data struct, you can create ggml context.
build_graph use ggml_new_graph, first looking for the ggml_cgraph struct:
cpp
struct ggml_cgraph {
int size; // maximum number of nodes/leafs/grads/grad_accs
// this is not memory size, just the max count of node
int n_nodes; // number of nodes currently in use
int n_leafs; // number of leafs currently in use
struct ggml_tensor ** nodes; // tensors with data that can change if the graph is evaluated
struct ggml_tensor ** grads; // the outputs of these tensors are the gradients of the nodes, use for train?
struct ggml_tensor ** grad_accs; // accumulators for node gradients, use for train?
struct ggml_tensor ** leafs; // tensors with constant data
struct ggml_hash_set visited_hash_set;
enum ggml_cgraph_eval_order order; // define the computer order
};cpp
struct ggml_cgraph * ggml_new_graph(struct ggml_context * ctx) {
return ggml_new_graph_custom(ctx, GGML_DEFAULT_GRAPH_SIZE, false);
}
struct ggml_cgraph * ggml_new_graph_custom(struct ggml_context * ctx, size_t size, bool grads) {
const size_t obj_size = ggml_graph_nbytes(size, grads);
// this allocate the memory for the compute graph from ggml_context
struct ggml_object * obj = ggml_new_object(ctx, GGML_OBJECT_TYPE_GRAPH, obj_size);
struct ggml_cgraph * cgraph = (struct ggml_cgraph *) ((char *) ctx->mem_buffer + obj->offs);
// the size of the hash table is doubled since it needs to hold both nodes and leafs
size_t hash_size = ggml_hash_size(size * 2);
// move the p to the memory begin
void * p = cgraph + 1;
// set these pointer, p will move in the process
struct ggml_tensor ** nodes_ptr = incr_ptr_aligned(&p, size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *));
struct ggml_tensor ** leafs_ptr = incr_ptr_aligned(&p, size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *));
struct ggml_tensor ** hash_keys_ptr = incr_ptr_aligned(&p, hash_size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *));
struct ggml_tensor ** grads_ptr = grads ? incr_ptr_aligned(&p, hash_size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)) : NULL;
struct ggml_tensor ** grad_accs_ptr = grads ? incr_ptr_aligned(&p, hash_size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)) : NULL;
ggml_bitset_t * hash_used = incr_ptr_aligned(&p, ggml_bitset_size(hash_size) * sizeof(ggml_bitset_t), sizeof(ggml_bitset_t));
// check that we allocated the correct amount of memory
assert(obj_size == (size_t)((char *)p - (char *)cgraph));
*cgraph = (struct ggml_cgraph) {
/*.size =*/ size,
/*.n_nodes =*/ 0,
/*.n_leafs =*/ 0,
/*.nodes =*/ nodes_ptr,
/*.grads =*/ grads_ptr,
/*.grad_accs =*/ grad_accs_ptr,
/*.leafs =*/ leafs_ptr,
/*.hash_table =*/ { hash_size, hash_used, hash_keys_ptr },
/*.order =*/ GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT,
};
// just memset the hash_table
ggml_hash_set_reset(&cgraph->visited_hash_set);
if (grads) {
memset(cgraph->grads, 0, hash_size*sizeof(struct ggml_tensor *));
memset(cgraph->grad_accs, 0, hash_size*sizeof(struct ggml_tensor *));
}
return cgraph;
}after create, add the compute to the graph:
cpp
// result = a*b^T
struct ggml_tensor * result = ggml_mul_mat(ctx0, model.a, model.b);this just create the tensor, set tensors src then return the tensor.
After that, add tensor to the graph:
cpp
// build operations nodes
ggml_build_forward_expand(gf, result);
void ggml_build_forward_expand(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor) {
ggml_build_forward_impl(cgraph, tensor, true);
}
static void ggml_build_forward_impl(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor, bool expand) {
if (!expand) {
// TODO: this branch isn't accessible anymore, maybe move this to ggml_build_forward_expand
ggml_graph_clear(cgraph);
}
const int n0 = cgraph->n_nodes;
// visit graph, and add
ggml_visit_parents(cgraph, tensor);
const int n_new = cgraph->n_nodes - n0;
GGML_PRINT_DEBUG("%s: visited %d new nodes\n", __func__, n_new);
if (n_new > 0) {
// the last added node should always be starting point
GGML_ASSERT(cgraph->nodes[cgraph->n_nodes - 1] == tensor);
}
}
static void ggml_visit_parents(struct ggml_cgraph * cgraph, struct ggml_tensor * node) {
// check if already visited
// already visited will be set in hash_table
if (ggml_hash_insert(&cgraph->visited_hash_set, node) == GGML_HASHSET_ALREADY_EXISTS) {
return;
}
for (int i = 0; i < GGML_MAX_SRC; ++i) {
const int k =
(cgraph->order == GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT) ? i :
(cgraph->order == GGML_CGRAPH_EVAL_ORDER_RIGHT_TO_LEFT) ? (GGML_MAX_SRC-1-i) :
/* unknown order, just fall back to using i*/ i;
if (node->src[k]) {
// recur visit parent to genertee all parent in the graph
ggml_visit_parents(cgraph, node->src[k]);
}
}
// add as leaf
if (node->op == GGML_OP_NONE && !(node->flags & GGML_TENSOR_FLAG_PARAM)) {
// reached a leaf node, not part of the gradient graph (e.g. a constant)
GGML_ASSERT(cgraph->n_leafs < cgraph->size);
if (strlen(node->name) == 0) {
ggml_format_name(node, "leaf_%d", cgraph->n_leafs);
}
cgraph->leafs[cgraph->n_leafs] = node;
cgraph->n_leafs++;
// add as node
} else {
GGML_ASSERT(cgraph->n_nodes < cgraph->size);
if (strlen(node->name) == 0) {
ggml_format_name(node, "node_%d", cgraph->n_nodes);
}
cgraph->nodes[cgraph->n_nodes] = node;
cgraph->n_nodes++;
}
}then ggml_gallocr_reserve:
cpp
bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, const int * node_buffer_ids, const int * leaf_buffer_ids) {
// this seems like create hash_table for quick find node
// the hash_table in cgraph use to check if the node in the cgraph
size_t min_hash_size = graph->n_nodes + graph->n_leafs;
// add 25% margin to avoid hash collisions
min_hash_size += min_hash_size / 4;
// initialize hash table
if (galloc->hash_set.size < min_hash_size) {
ggml_hash_set_free(&galloc->hash_set);
galloc->hash_set = ggml_hash_set_new(min_hash_size);
GGML_ASSERT(galloc->hash_set.keys != NULL);
free(galloc->hash_values);
galloc->hash_values = malloc(sizeof(struct hash_node) * galloc->hash_set.size);
GGML_ASSERT(galloc->hash_values != NULL);
}
// reset ggml_dyn_tallocr
for (int i = 0; i < galloc->n_buffers; i++) {
ggml_dyn_tallocr_reset(galloc->buf_tallocs[i]);
}
// allocate in hash table
ggml_gallocr_alloc_graph_impl(galloc, graph, node_buffer_ids, leaf_buffer_ids);
// set the node_allocs from the hash table
if (galloc->n_nodes < graph->n_nodes) {
free(galloc->node_allocs);
galloc->node_allocs = calloc(graph->n_nodes, sizeof(struct node_alloc));
GGML_ASSERT(galloc->node_allocs != NULL);
}
galloc->n_nodes = graph->n_nodes;
for (int i = 0; i < graph->n_nodes; i++) {
struct ggml_tensor * node = graph->nodes[i];
struct node_alloc * node_alloc = &galloc->node_allocs[i];
if (node->view_src || node->data) {
node_alloc->dst.buffer_id = -1;
node_alloc->dst.offset = SIZE_MAX;
node_alloc->dst.size_max = 0;
} else {
struct hash_node * hn = ggml_gallocr_hash_get(galloc, node);
node_alloc->dst.buffer_id = hn->buffer_id;
node_alloc->dst.offset = hn->offset;
node_alloc->dst.size_max = ggml_backend_buft_get_alloc_size(galloc->bufts[hn->buffer_id], node);
}
for (int j = 0; j < GGML_MAX_SRC; j++) {
struct ggml_tensor * src = node->src[j];
if (!src || src->view_src || src->data) {
node_alloc->src[j].buffer_id = -1;
node_alloc->src[j].offset = SIZE_MAX;
node_alloc->src[j].size_max = 0;
} else {
struct hash_node * hn = ggml_gallocr_hash_get(galloc, src);
node_alloc->src[j].buffer_id = hn->buffer_id;
node_alloc->src[j].offset = hn->offset;
node_alloc->src[j].size_max = ggml_backend_buft_get_alloc_size(galloc->bufts[hn->buffer_id], src);
}
}
}
if (galloc->n_leafs < graph->n_leafs) {
free(galloc->leaf_allocs);
galloc->leaf_allocs = calloc(graph->n_leafs, sizeof(galloc->leaf_allocs[0]));
GGML_ASSERT(galloc->leaf_allocs != NULL);
}
galloc->n_leafs = graph->n_leafs;
for (int i = 0; i < graph->n_leafs; i++) {
struct ggml_tensor * leaf = graph->leafs[i];
struct hash_node * hn = ggml_gallocr_hash_get(galloc, leaf);
if (leaf->view_src || leaf->data) {
galloc->leaf_allocs[i].leaf.buffer_id = -1;
galloc->leaf_allocs[i].leaf.offset = SIZE_MAX;
galloc->leaf_allocs[i].leaf.size_max = 0;
} else {
galloc->leaf_allocs[i].leaf.buffer_id = hn->buffer_id;
galloc->leaf_allocs[i].leaf.offset = hn->offset;
galloc->leaf_allocs[i].leaf.size_max = ggml_backend_buft_get_alloc_size(galloc->bufts[hn->buffer_id], leaf);
}
}
// reallocate buffers if needed
// after before are calculate size, in this are allocate
for (int i = 0; i < galloc->n_buffers; i++) {
// if the buffer type is used multiple times, we reuse the same buffer
for (int j = 0; j < i; j++) {
if (galloc->buf_tallocs[j] == galloc->buf_tallocs[i]) {
galloc->buffers[i] = galloc->buffers[j];
break;
}
}
size_t cur_size = galloc->buffers[i] ? ggml_backend_buffer_get_size(galloc->buffers[i]) : 0;
size_t new_size = ggml_dyn_tallocr_max_size(galloc->buf_tallocs[i]);
// even if there are no tensors allocated in this buffer, we still need to allocate it to initialize views
if (new_size > cur_size || galloc->buffers[i] == NULL) {
#ifndef NDEBUG
GGML_LOG_DEBUG("%s: reallocating %s buffer from size %.02f MiB to %.02f MiB\n", __func__, ggml_backend_buft_name(galloc->bufts[i]), cur_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
#endif
ggml_backend_buffer_free(galloc->buffers[i]);
galloc->buffers[i] = ggml_backend_buft_alloc_buffer(galloc->bufts[i], new_size);
if (galloc->buffers[i] == NULL) {
GGML_LOG_ERROR("%s: failed to allocate %s buffer of size %zu\n", __func__, ggml_backend_buft_name(galloc->bufts[i]), new_size);
return false;
}
ggml_backend_buffer_set_usage(galloc->buffers[i], GGML_BACKEND_BUFFER_USAGE_COMPUTE);
}
}
return true;
}cpp
static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgraph * graph, const int * node_buffer_ids, const int * leaf_buffer_ids) {
// clear hash tables
ggml_hash_set_reset(&galloc->hash_set);
memset(galloc->hash_values, 0, sizeof(struct hash_node) * galloc->hash_set.size);
// allocate leafs
// these may be tensors that the application is not using in the graph, but may still want to allocate for other purposes
for (int i = 0; i < graph->n_leafs; i++) {
struct ggml_tensor * leaf = graph->leafs[i];
// I cant see any allocate from ggml_backend_buffer_type, this means before build the graph, leaf node should allocate memory itself
// this just count some message about leaf in galloc and ggml_dyn_tallocr
ggml_gallocr_allocate_node(galloc, leaf, get_node_buffer_id(leaf_buffer_ids, i));
}
// count number of children and views
// allocate other graph inputs and leafs first to avoid overwriting them
// count the child and other message, make efficient
for (int i = 0; i < graph->n_nodes; i++) {
struct ggml_tensor * node = graph->nodes[i];
// TODO: better way to add external dependencies
// GGML_OP_NONE does not appear normally in the graph nodes, but is used by ggml-backend to add dependencies to
// control when some tensors are allocated and freed. in this case, the dependencies are in `src`, but the node
// itself is never used and should not be considered a dependency
if (ggml_is_view(node) && node->op != GGML_OP_NONE) {
struct ggml_tensor * view_src = node->view_src;
ggml_gallocr_hash_get(galloc, view_src)->n_views += 1;
}
if (node->flags & GGML_TENSOR_FLAG_INPUT) {
ggml_gallocr_allocate_node(galloc, graph->nodes[i], get_node_buffer_id(node_buffer_ids, i));
}
for (int j = 0; j < GGML_MAX_SRC; j++) {
struct ggml_tensor * src = node->src[j];
if (src == NULL) {
continue;
}
ggml_gallocr_hash_get(galloc, src)->n_children += 1;
// allocate explicit inputs
if (src->flags & GGML_TENSOR_FLAG_INPUT) {
ggml_gallocr_allocate_node(galloc, src, get_node_buffer_id(node_buffer_ids, i));
}
}
}
// allocate tensors
for (int i = 0; i < graph->n_nodes; i++) {
struct ggml_tensor * node = graph->nodes[i];
int buffer_id = get_node_buffer_id(node_buffer_ids, i);
// allocate parents (only leafs need to be allocated at this point)
for (int j = 0; j < GGML_MAX_SRC; j++) {
struct ggml_tensor * parent = node->src[j];
if (parent == NULL) {
continue;
}
ggml_gallocr_allocate_node(galloc, parent, buffer_id);
}
// allocate node
// memory allocate by buffer_id?
ggml_gallocr_allocate_node(galloc, node, buffer_id);
AT_PRINTF("exec: %s (%s) <= ", ggml_op_desc(node), node->name);
for (int j = 0; j < GGML_MAX_SRC; j++) {
struct ggml_tensor * parent = node->src[j];
if (parent == NULL) {
continue;
}
AT_PRINTF("%s", parent->name);
if (j < GGML_MAX_SRC - 1 && node->src[j + 1] != NULL) {
AT_PRINTF(", ");
}
}
AT_PRINTF("\n");
// update parents
for (int j = 0; j < GGML_MAX_SRC; j++) {
struct ggml_tensor * parent = node->src[j];
if (parent == NULL) {
continue;
}
struct hash_node * p_hn = ggml_gallocr_hash_get(galloc, parent);
p_hn->n_children -= 1;
AT_PRINTF("parent %s: %d children, %d views, allocated: %d\n",
parent->name, p_hn->n_children, p_hn->n_views, p_hn->allocated);
if (p_hn->n_children == 0 && p_hn->n_views == 0) {
if (ggml_is_view(parent)) {
struct ggml_tensor * view_src = parent->view_src;
struct hash_node * view_src_hn = ggml_gallocr_hash_get(galloc, view_src);
view_src_hn->n_views -= 1;
AT_PRINTF("view_src %s: %d children, %d views\n",
view_src->name, view_src_hn->n_children, view_src_hn->n_views);
if (view_src_hn->n_views == 0 && view_src_hn->n_children == 0 && view_src_hn->allocated) {
ggml_gallocr_free_node(galloc, view_src);
}
}
else if (p_hn->allocated) {
ggml_gallocr_free_node(galloc, parent);
}
}
AT_PRINTF("\n");
}
}
}compute
cpp
struct ggml_tensor * compute(const simple_model & model, ggml_gallocr_t allocr) {
// reset the allocator to free all the memory allocated during the previous inference
struct ggml_cgraph * gf = build_graph(model);
// allocate tensors
// connect buffer with tensor
ggml_gallocr_alloc_graph(allocr, gf);
int n_threads = 1; // number of threads to perform some operations with multi-threading
if (ggml_backend_is_cpu(model.backend)) {
// this is user set n_threads to computer
// set the how many threads use in cpu
// just set cpucontext n_threads
ggml_backend_cpu_set_n_threads(model.backend, n_threads);
}
ggml_backend_graph_compute(model.backend, gf);
// in this case, the output tensor is the last one in the graph
// return the result
return ggml_graph_node(gf, -1);
}computer the graph:
cpp
enum ggml_status ggml_backend_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
// do computer
enum ggml_status err = ggml_backend_graph_compute_async(backend, cgraph);
// if hava sync callback, sync
ggml_backend_synchronize(backend);
return err;
}in the cpu, the graph compute is:
cpp
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;
// create compute plan, how many intermedicate memory should be use?
struct ggml_cplan cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads, cpu_ctx->threadpool);
// if the cpu_ctx work_size < plan work_size, realloc
if (cpu_ctx->work_size < cplan.work_size) {
delete[] cpu_ctx->work_data;
// realloc memory
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;
// set the abort callback
cplan.abort_callback = cpu_ctx->abort_callback;
cplan.abort_callback_data = cpu_ctx->abort_callback_data;
return ggml_graph_compute(cgraph, &cplan);
}cpp
struct ggml_cplan ggml_graph_plan(
const struct ggml_cgraph * cgraph,
int n_threads,
struct ggml_threadpool * threadpool) {
if (threadpool == NULL) {
//GGML_PRINT_DEBUG("Threadpool is not specified. Will create a disposable threadpool : n_threads %d\n", n_threads);
}
if (n_threads <= 0) {
n_threads = threadpool ? threadpool->n_threads_max : GGML_DEFAULT_N_THREADS;
}
size_t work_size = 0;
// plan is a buffer use for data
struct ggml_cplan cplan;
memset(&cplan, 0, sizeof(struct ggml_cplan));
int max_tasks = 1;
// thread scheduling for the different operations + work buffer size estimation
// iter node to calculate the intermedicate buffer size and task size
for (int i = 0; i < cgraph->n_nodes; i++) {
struct ggml_tensor * node = cgraph->nodes[i];
// return this node op need how many thread
const int n_tasks = ggml_get_n_tasks(node, n_threads);
max_tasks = MAX(max_tasks, n_tasks);
size_t cur = 0;
// may need more woek size?
// is have extra, this function will change cur
// return the size caculate should use
if (!ggml_cpu_extra_work_size(n_threads, node, &cur)) {
switch (node->op) {
// ...
case GGML_OP_ADD:
case GGML_OP_ADD1:
{
if (ggml_is_quantized(node->src[0]->type)) {
cur = ggml_type_size(GGML_TYPE_F32) * node->src[0]->ne[0] * n_tasks;
}
} break;
case GGML_OP_MUL_MAT:
{
const enum ggml_type vec_dot_type = type_traits_cpu[node->src[0]->type].vec_dot_type;
if (node->src[1]->type != vec_dot_type) {
cur = ggml_row_size(vec_dot_type, ggml_nelements(node->src[1]));
}
} break;
// ...
default:
break;
}
}
// need how many memory
work_size = MAX(work_size, cur);
}
// per thread have cache line
if (work_size > 0) {
work_size += CACHE_LINE_SIZE*(n_threads);
}
cplan.threadpool = threadpool;
cplan.n_threads = MIN(max_tasks, n_threads);
// intermedicate data size, max of all task
cplan.work_size = work_size;
// work_data wait allocate by buffer
cplan.work_data = NULL;
return cplan;
}all is ready, do the graph computer
cpp
enum ggml_status ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan) {
ggml_cpu_init();
GGML_ASSERT(cplan);
GGML_ASSERT(cplan->n_threads > 0);
GGML_ASSERT(cplan->work_size == 0 || cplan->work_data != NULL);
int n_threads = cplan->n_threads;
struct ggml_threadpool * threadpool = cplan->threadpool;
bool disposable_threadpool = false;
// if threadpool is null, allocate a new threadpool
if (threadpool == NULL) {
//GGML_PRINT_DEBUG("Threadpool is not specified. Will create a disposable threadpool : n_threads %d\n", n_threads);
disposable_threadpool = true;
struct ggml_threadpool_params ttp = ggml_threadpool_params_default(n_threads);
threadpool = ggml_threadpool_new_impl(&ttp, cgraph, cplan);
} else {
// Reset some of the parameters that need resetting
// No worker threads should be accessing the parameters below at this stage
threadpool->cgraph = cgraph;
threadpool->cplan = cplan;
threadpool->current_chunk = 0;
threadpool->abort = -1;
threadpool->ec = GGML_STATUS_SUCCESS;
}
#ifdef GGML_USE_OPENMP
if (n_threads > 1) {
#pragma omp parallel num_threads(n_threads)
{
#pragma omp single
{
// update the number of threads from the actual number of threads that we got from OpenMP
n_threads = omp_get_num_threads();
atomic_store_explicit(&threadpool->n_threads_cur, n_threads, memory_order_relaxed);
}
// openmp will unroll
// use openmp to unroll
ggml_graph_compute_thread(&threadpool->workers[omp_get_thread_num()]);
}
} else {
atomic_store_explicit(&threadpool->n_threads_cur, 1, memory_order_relaxed);
ggml_graph_compute_thread(&threadpool->workers[0]);
}
#else
if (n_threads > threadpool->n_threads_max) {
GGML_LOG_WARN("cplan requested more threads (%d) than available (%d)\n", n_threads, threadpool->n_threads_max);
n_threads = threadpool->n_threads_max;
}
// Kick all threads to start the new graph
ggml_graph_compute_kickoff(threadpool, n_threads);
// This is a work thread too
ggml_graph_compute_thread(&threadpool->workers[0]);
#endif
// don't leave affinity set on the main thread
clear_numa_thread_affinity();
enum ggml_status ret = threadpool->ec;
if (disposable_threadpool) {
ggml_threadpool_free(threadpool);
}
return ret;
}the important function is ggml_graph_compute_thread
cpp
static thread_ret_t ggml_graph_compute_thread(void * data) {
struct ggml_compute_state * state = (struct ggml_compute_state *) data;
struct ggml_threadpool * tp = state->threadpool;
const struct ggml_cgraph * cgraph = tp->cgraph;
const struct ggml_cplan * cplan = tp->cplan;
set_numa_thread_affinity(state->ith);
struct ggml_compute_params params = {
/*.ith =*/ state->ith, // ith means cur threads id
/*.nth =*/ atomic_load_explicit(&tp->n_threads_cur, memory_order_relaxed),
/*.wsize =*/ cplan->work_size,
/*.wdata =*/ cplan->work_data,
/*.threadpool=*/ tp,
};
// thread parallel in node level
for (int node_n = 0; node_n < cgraph->n_nodes && atomic_load_explicit(&tp->abort, memory_order_relaxed) != node_n; node_n++) {
struct ggml_tensor * node = cgraph->nodes[node_n];
// per thread do own work
ggml_compute_forward(¶ms, node);
if (state->ith == 0 && cplan->abort_callback &&
cplan->abort_callback(cplan->abort_callback_data)) {
atomic_store_explicit(&tp->abort, node_n + 1, memory_order_relaxed);
tp->ec = GGML_STATUS_ABORTED;
}
// wait for other thread
if (node_n + 1 < cgraph->n_nodes) {
ggml_barrier(state->threadpool);
}
// all thread finish goto next node
}
ggml_barrier(state->threadpool);
return 0;
}the ggml_compute_forward is real computer function:
cpp
static void ggml_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor) {
GGML_ASSERT(params);
if (tensor->op == GGML_OP_NONE || ggml_is_empty(tensor)) {
return;
}
// extra_buffer use to cpu extension
if (ggml_cpu_extra_compute_forward(params, tensor)) {
return;
}
// this is the default implement
switch (tensor->op) {
// ...
case GGML_OP_MUL_MAT:
{
ggml_compute_forward_mul_mat(params, tensor);
} break;
case GGML_OP_MUL_MAT_ID:
{
ggml_compute_forward_mul_mat_id(params, tensor);
} break;
// ...
case GGML_OP_COUNT:
{
GGML_ABORT("fatal error");
}
}
}after that, the compute is finish, the result is in the node wait be use.
extra buffer
extra buffer use for cpu extension like intel amx and other.
cpp
bool ggml_cpu_extra_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * op) {
for (auto extra : ggml_backend_cpu_get_extra_buffers_type()) {
if (extra && extra->context) {
auto buf_extra = (ggml::cpu::extra_buffer_type *) extra->context;
auto tensor_traits = buf_extra->get_tensor_traits(op);
if (tensor_traits && tensor_traits->compute_forward(params, op)) {
return true;
}
}
}
return false;
}this check is any cpu extension can accelerate the computer, such as gemm, these can be use DSA to computer.