DarthAffe/stable-diffusion.cpp
1
0 Fork 0
mirror of https://github.com/leejet/stable-diffusion.cpp.git synced 2026-11-11 03:24:30 +00:00
Code Issues Packages Projects Releases Wiki Activity

perf: optimize Vulkan kernels, graph fusion, and workspace reuse (#2251)

Browse Source
This commit is contained in:
leejet authored and GitHub committed 2026-10-31 14:55:58 +08:00
1 parent e239e4b0fd
commit 0a067120b3
6 files changed
+89 -17
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

src/core/compute_workspace.cpp
+54 -1
View File
@@ -92,7 +92,8 @@ namespace sd {
ComputeWorkspace::Measurement ComputeWorkspace::measure(
ggml_cgraph* graph,
const std::function<ggml_backend_t(const ggml_tensor*)>& external_backend,
const AssignNodes& assign_nodes) {
const AssignNodes& assign_nodes,
bool cache_measurement) {
std::vector<const ggml_tensor*> tensors;
std::unordered_set<const ggml_tensor*> seen;
auto visit = [&](const ggml_tensor* tensor) {
@@ -112,6 +113,52 @@ namespace sd {
visit(source);
}
}
std::vector<uint8_t> key;
if (cache_measurement) {
key.reserve(tensors.size() * sizeof(ggml_tensor));
auto append = [&](const auto& value) {
const auto* bytes = reinterpret_cast<const uint8_t*>(&value);
key.insert(key.end(), bytes, bytes + sizeof(value));
};
append(graph->n_nodes);
append(graph->n_leafs);
append(graph->order);
for (auto tensor : tensors) {
append(tensor);
append(tensor->type);
append(tensor->ne);
append(tensor->nb);
append(tensor->op);
append(tensor->op_params);
append(tensor->flags);
append(tensor->src);
append(tensor->view_src);
append(tensor->view_offs);
append(tensor->name);
const auto external = external_backend(tensor);
append(external);
if (external == nullptr || tensor->view_src != nullptr) {
append(tensor->buffer);
append(tensor->data);
append(tensor->extra);
if (tensor->buffer != nullptr) {
const auto usage = ggml_backend_buffer_get_usage(tensor->buffer);
const auto type = ggml_backend_buffer_get_type(tensor->buffer);
append(usage);
append(type);
}
}
const size_t index = ggml_hash_find(&graph->visited_hash_set, tensor);
const int32_t uses = index != GGML_HASHSET_FULL && ggml_bitset_get(graph->visited_hash_set.used, index)
? graph->use_counts[index]
: 0;
append(uses);
}
// Match the full allocation metadata, including topology and external placement.
if (key == measurement_key_) {
return {measured_buffers_};
}
}
const size_t graph_size = tensors.size() + 8;
auto context = ggml_init({tensors.size() * ggml_tensor_overhead() + ggml_graph_overhead_custom(graph_size, false), nullptr, true});
if (context == nullptr) {
@@ -178,6 +225,10 @@ namespace sd {
ggml_backend_buffer_free(entry.second);
}
ggml_free(context);
if (cache_measurement && !result.buffers.empty()) {
measurement_key_ = std::move(key);
measured_buffers_ = result.buffers;
}
return result;
}
@@ -212,6 +263,8 @@ namespace sd {
if (active_) {
return false;
}
measurement_key_.clear();
measured_buffers_.clear();
ggml_backend_sched_free(scheduler_);
scheduler_ = nullptr;
scheduler_capacity_ = 0;