CANN: refactor ACL graph cache matching

docs/backend/CANN.md

@@ -314,3 +314,10 @@ Controls automatic cleanup of the memory pool. This option is only effective whe
 
 Converting the matmul weight format from ND to NZ can significantly improve performance on the 310I DUO NPU.
 
+### GGML_CANN_ACL_GRAPH
+
+Operators are executed using ACL graph execution, rather than in op-by-op (eager) mode. Enabled by default.
+
+### GGML_CANN_GRAPH_CACHE_CAPACITY
+
+Maximum number of compiled CANN graphs kept in the LRU cache, default is 12. When the number of cached graphs exceeds this capacity, the least recently used graph will be evicted.

ggml/src/ggml-cann/common.h

@@ -38,6 +38,7 @@
 #include <unistd.h>
 #include <functional>
 #include <optional>
+#include <list>
 
 #include "../include/ggml-cann.h"
 #include "../include/ggml.h"
@@ -106,6 +107,7 @@ int32_t ggml_cann_get_device();
 
 std::optional<std::string> get_env(const std::string& name);
 bool parse_bool(const std::string& value);
+int parse_integer(const std::string& value);
 
 /**
  * @brief Abstract base class for memory pools used by CANN.
@@ -339,24 +341,227 @@ class cann_task_queue {
 
 #ifdef USE_ACL_GRAPH
 struct ggml_graph_node_properties {
+    // dst tensor
     void * node_address;
-    ggml_op node_op;
     int64_t ne[GGML_MAX_DIMS];
     size_t nb[GGML_MAX_DIMS];
+
+    // src tensor
     void * src_address[GGML_MAX_SRC];
+    int64_t src_ne[GGML_MAX_SRC][GGML_MAX_DIMS];
+    size_t  src_nb[GGML_MAX_SRC][GGML_MAX_DIMS];
+
+    // op
+    ggml_op node_op;
     int32_t op_params[GGML_MAX_OP_PARAMS / sizeof(int32_t)];
+
+    /**
+     * @brief Check if a ggml tensor node matches this property set.
+     *
+     * This function compares all relevant fields (address, op type, shape, source inputs, op params)
+     * to determine whether the current node matches these previously recorded properties.
+     *
+     * @param node The current ggml tensor node.
+     * @return true if all fields match (excluding GGML_OP_VIEW); false otherwise.
+     */
+    bool has_matching_properties(ggml_tensor * node) {
+        if (node->data != this->node_address && node->op != GGML_OP_VIEW) {
+            return false;
+        }
+
+        if (node->op != this->node_op) {
+            return false;
+        }
+
+        for (int i = 0; i < GGML_MAX_DIMS; i++) {
+            if (node->ne[i] != this->ne[i]) {
+                return false;
+            }
+            if (node->nb[i] != this->nb[i]) {
+                return false;
+            }
+        }
+
+        for (int i = 0; i < GGML_MAX_SRC; i++) {
+            if (node->src[i]) {
+                if (node->src[i]->data != this->src_address[i] && node->op != GGML_OP_VIEW) {
+                    return false;
+                }
+
+                for (int d = 0; d < GGML_MAX_DIMS; d++) {
+                    if (node->src[i]->ne[d] != this->src_ne[i][d]) {
+                        return false;
+                    }
+                    if (node->src[i]->nb[d] != this->src_nb[i][d]) {
+                        return false;
+                    }
+                }
+            } else {
+                if (this->src_address[i] != nullptr) {
+                    return false;
+                }
+            }
+        }
+
+        if (node->op == GGML_OP_SCALE || node->op == GGML_OP_UNARY || node->op == GGML_OP_GLU) {
+            return memcmp(this->op_params, node->op_params, GGML_MAX_OP_PARAMS) == 0;
+        }
+        return true;
+    }
 };
 
 struct ggml_cann_graph {
     ~ggml_cann_graph() {
         if (graph != nullptr) {
-            aclmdlRIDestroy(graph);
+            ACL_CHECK(aclmdlRIDestroy(graph));
         }
     }
 
     aclmdlRI graph = nullptr;
 
     std::vector<ggml_graph_node_properties> ggml_graph_properties;
+
+    /**
+     * @brief Create a new CANN graph from a ggml computation graph.
+     *
+     * This function creates a new ggml_cann_graph object and fills its node properties
+     * (operation type, dimensions, strides, input sources, and operation parameters)
+     * based on the current ggml computation graph.
+     *
+     * Each node in the ggml graph is mapped to a property entry in the new CANN graph:
+     * - node address
+     * - operation type
+     * - shape (ne) and strides (nb)
+     * - source tensor addresses
+     * - operation parameters
+     *
+     * @param cgraph The current ggml computation graph.
+     * @return Pointer to the newly created ggml_cann_graph object.
+     */
+    static ggml_cann_graph * create_from_cgraph(ggml_cgraph * cgraph) {
+        ggml_cann_graph * new_graph = new ggml_cann_graph();
+        new_graph->ggml_graph_properties.resize(cgraph->n_nodes);
+
+        for (int node_idx = 0; node_idx < cgraph->n_nodes; ++node_idx) {
+            ggml_tensor * node = cgraph->nodes[node_idx];
+            auto &        prop = new_graph->ggml_graph_properties[node_idx];
+
+            prop.node_address = node->data;
+            prop.node_op      = node->op;
+
+            std::copy_n(node->ne, GGML_MAX_DIMS, prop.ne);
+            std::copy_n(node->nb, GGML_MAX_DIMS, prop.nb);
+
+            for (int src = 0; src < GGML_MAX_SRC; ++src) {
+                if (node->src[src]) {
+                    prop.src_address[src] = node->src[src]->data;
+                    std::copy_n(node->src[src]->ne, GGML_MAX_DIMS, prop.src_ne[src]);
+                    std::copy_n(node->src[src]->nb, GGML_MAX_DIMS, prop.src_nb[src]);
+                } else {
+                    prop.src_address[src] = nullptr;
+                    std::fill_n(prop.src_ne[src], GGML_MAX_DIMS, 0);
+                    std::fill_n(prop.src_nb[src], GGML_MAX_DIMS, 0);
+                }
+            }
+
+            memcpy(prop.op_params, node->op_params, GGML_MAX_OP_PARAMS);
+        }
+
+        return new_graph;
+    }
+
+    /**
+     * @brief Check whether this CANN graph matches the given ggml computation graph.
+     *
+     * This function compares the number of nodes and each node's properties
+     * (operation type, dimensions, strides, inputs, and operation parameters)
+     * to determine whether this CANN graph matches the given ggml graph.
+     *
+     * @param cgraph The current ggml computation graph.
+     * @return true if this CANN graph matches the ggml graph; false otherwise.
+     */
+    bool matches_cgraph(ggml_cgraph * cgraph) {
+        if (this->ggml_graph_properties.size() != static_cast<size_t>(cgraph->n_nodes)) {
+            return false;
+        }
+
+        for (int i = 0; i < cgraph->n_nodes; ++i) {
+            if (!this->ggml_graph_properties[i].has_matching_properties(cgraph->nodes[i])) {
+                return false;
+            }
+        }
+
+        return true;
+    }
+};
+
+/**
+ * @brief LRU cache for managing ggml_cann_graph objects.
+ *
+ * This class maintains a list of shared_ptr to ggml_cann_graph objects
+ * and enforces a maximum capacity. It provides methods to push new graphs,
+ * move existing graphs to the front (most recently used), and clear the cache.
+ */
+struct ggml_cann_graph_lru_cache {
+    size_t capacity;  /**< Maximum number of graphs in the cache. */
+
+    std::list<ggml_cann_graph*> cache_list; /**< List storing cached graphs as raw pointers. */
+
+    ggml_cann_graph_lru_cache() {
+        capacity = parse_integer(get_env("GGML_CANN_GRAPH_CACHE_CAPACITY").value_or("12"));
+    }
+
+    /**
+     * @brief Push a new graph to the front of the cache.
+     * If the cache exceeds capacity, the least recently used graph is deleted.
+     * @param new_node Pointer to the new ggml_cann_graph to cache.
+     *        Ownership is transferred to the cache (cache will delete it).
+     */
+    void push(ggml_cann_graph* new_node) {
+        if (cache_list.size() >= capacity) {
+            ggml_cann_graph* old = cache_list.back();
+            cache_list.pop_back();
+            delete old; // free the old graph
+        }
+        cache_list.push_front(new_node);
+    }
+
+    /**
+     * @brief Clear all graphs from the cache (also frees memory).
+     */
+    void clear() {
+        for (auto ptr : cache_list) {
+            delete ptr;
+        }
+        cache_list.clear();
+    }
+
+    /**
+     * @brief Destructor that clears the cache and frees all cached graphs.
+     */
+    ~ggml_cann_graph_lru_cache() { clear(); }
+
+    /**
+     * @brief Find a cached CANN graph that matches the given ggml graph and move it to front.
+     *
+     * This function iterates through the cached CANN graphs stored in the LRU cache and
+     * compares them against the given ggml computation graph. If a matching graph is found,
+     * it is promoted to the front of the LRU cache and returned. Otherwise, the function
+     * returns nullptr.
+     *
+     * @param cgraph The current ggml computation graph.
+     * @return true if found; false otherwise.
+     */
+    bool find_and_move_to_front(ggml_cgraph * cgraph) {
+        for (auto & graph_ptr : this->cache_list) {
+            if (graph_ptr->matches_cgraph(cgraph)) {
+                cache_list.remove(graph_ptr);
+                cache_list.push_front(graph_ptr);
+                return true;
+            }
+        }
+        return false;
+    }
 };
 #endif  // USE_ACL_GRAPH
 
@@ -370,7 +575,8 @@ struct ggml_backend_cann_context {
     aclrtEvent copy_event = nullptr; /**< Event for managing copy operations. */
 #ifdef USE_ACL_GRAPH
     /// Cached CANN ACL graph used for executing the current ggml computation graph.
-    std::unique_ptr<ggml_cann_graph> cann_graph;
+    ggml_cann_graph_lru_cache graph_lru_cache;
+    bool acl_graph_mode = true;
 #endif
     cann_task_queue task_queue;
     bool async_mode;