Introduce a shared array in DAG runInfo that stores DAG intermediate tensors (instead of the DagLocalContext dict). Every op holds the indices of its input and output tensors, and uses this array to load and store them in turn.

Author: alonre24

src/DAG/dag.c

@@ -88,6 +88,24 @@ void RedisAI_DagOpBatchInfo(RedisAI_RunInfo *rinfo, RAI_DagOp *op, size_t *batch
 void RedisAI_DagOpBatchingMatch(RedisAI_RunInfo *rinfo1, RAI_DagOp *op1, RedisAI_RunInfo *rinfo2,
                                 RAI_DagOp *op2, int *batched, size_t *inbatchsize);
 
+/**
+ * @brief Get a tensor from the dag local context in a given index
+ * (this access to a shared array, require read lock)
+ * @param rinfo The DAG runInfo.
+ * @param index The index of the tensor in the Dag shared array to return
+ * @return The tensor of the given index (NULL is returned if this tensor hasn't been realized yet)
+ */
+RAI_Tensor *Dag_GetInternalTensor(RedisAI_RunInfo *rinfo, size_t index);
+
+/**
+ * @brief Shallow copy and set a tensor in the dag local context in a given index.
+ * (this access to a shared array, require write lock)
+ * @param rinfo The DAG runInfo.
+ * @param index The index to put in the given tensor in the Dag shared array.
+ * @param t The tensor to shallow copy and store in the given index.
+ */
+void Dag_SetInternalTensor(RedisAI_RunInfo *rinfo, size_t index, RAI_Tensor *t);
+
 /**
  * Run the first unrealized DAG operation in rinfo for the given device.
  * @param rinfo context in which RedisAI blocking commands operate.

src/DAG/dag.h

@@ -29,13 +29,13 @@ int RAI_DAGLoadTensor(RAI_DAGRunCtx *run_info, const char *t_name, RAI_Tensor *t
 
     RedisAI_RunInfo *rinfo = (RedisAI_RunInfo *)run_info;
     RedisModuleString *key_name = RedisModule_CreateString(NULL, t_name, strlen(t_name));
-    // Add the tensor under its "mangled" key name to the DAG local context dict.
-    char buf[16];
-    sprintf(buf, "%04d", 1);
-    RedisModule_StringAppendBuffer(NULL, key_name, buf, strlen(buf));
-    AI_dictAdd(rinfo->dagTensorsContext, (void *)key_name,
-               (void *)RAI_TensorGetShallowCopy(tensor));
-    RedisModule_FreeString(NULL, key_name);
+
+    // Add the tensor to the DAG shared tensors and map its name to the relevant index.
+    int *instance = RedisModule_Alloc(sizeof(int));
+    *instance = array_len(rinfo->dagSharedTensors);
+    AI_dictAdd(rinfo->tensorsNamesToIndices, (void *)key_name, (void *)instance);
+    RAI_TensorGetShallowCopy(tensor);
+    rinfo->dagSharedTensors = array_append(rinfo->dagSharedTensors, (void *)tensor);
 
     return REDISMODULE_OK;
 }

src/DAG/dag_builder.c

@@ -3,155 +3,59 @@
 #include "background_workers.h"
 #include "util/string_utils.h"
 
-void _DAG_SetTensorsInLocalContext(RedisAI_RunInfo *rinfo) {
-    for (size_t i = 0; i < rinfo->dagOpCount; i++) {
-        RAI_DagOp *op = rinfo->dagOps[i];
-        if (op->commandType == REDISAI_DAG_CMD_TENSORSET) {
-            // Insert the tensor with its mangled (unique) name.
-            void *t = (void *)RAI_TensorGetShallowCopy(op->outTensor);
-            AI_dictReplace(rinfo->dagTensorsContext, (void *)op->outkeys[0], t);
-        }
-    }
-}
-
-int MangleTensorsNames(RedisAI_RunInfo *rinfo) {
-
-    int res = REDISMODULE_ERR;
-    AI_dict *mangled_tensors = AI_dictCreate(&AI_dictTypeHeapRStrings, NULL);
+int ValidatePersistKeys(RedisAI_RunInfo *rinfo, AI_dict *tensorsNamesToInd,
+                        AI_dict *persistTensorsNames) {
 
     {
-        AI_dictIterator *iter = AI_dictGetSafeIterator(rinfo->dagTensorsContext);
-        AI_dictEntry *entry = AI_dictNext(iter);
-        while (entry) {
-            RedisModuleString *key = (RedisModuleString *)AI_dictGetKey(entry);
-            size_t key_len;
-            const char *key_str = RedisModule_StringPtrLen(key, &key_len);
-            RedisModuleString *demangled_key = RedisModule_CreateString(NULL, key_str, key_len - 4);
-            int *instance = RedisModule_Alloc(sizeof(int));
-            *instance = 1;
-            AI_dictAdd(mangled_tensors, (void *)demangled_key, (void *)instance);
-            RedisModule_FreeString(NULL, demangled_key);
-            entry = AI_dictNext(iter);
+        AI_dictIterator *iter = AI_dictGetSafeIterator(persistTensorsNames);
+        AI_dictEntry *persist_entry;
+        while ((persist_entry = AI_dictNext(iter))) {
+            RedisModuleString *persist_key = (RedisModuleString *)AI_dictGetKey(persist_entry);
+            AI_dictEntry *entry = AI_dictFind(tensorsNamesToInd, persist_key);
+            if (!entry) {
+                RAI_SetError(rinfo->err, RAI_EDAGRUN, "ERR PERSIST key cannot be found in DAG");
+                return REDISMODULE_ERR;
+            }
         }
         AI_dictReleaseIterator(iter);
     }
+    return REDISMODULE_OK;
+}
+
+int MapTensorsKeysToIndices(RedisAI_RunInfo *rinfo, AI_dict *tensorsNamesToInd) {
 
     for (long long i = 0; i < array_len(rinfo->dagOps); i++) {
         RAI_DagOp *currentOp = rinfo->dagOps[i];
 
-        RedisModuleString **mangled_inkeys =
-            array_new(RedisModuleString *, array_len(currentOp->inkeys));
         for (long long j = 0; j < array_len(currentOp->inkeys); j++) {
             RedisModuleString *key = currentOp->inkeys[j];
-            AI_dictEntry *entry = AI_dictFind(mangled_tensors, key);
+            AI_dictEntry *entry = AI_dictFind(tensorsNamesToInd, key);
             if (!entry) {
-                array_free(mangled_inkeys);
                 RAI_SetError(rinfo->err, RAI_EDAGRUN, "ERR INPUT key cannot be found in DAG");
-                goto cleanup;
+                return REDISMODULE_ERR;
             }
-            int *instance = AI_dictGetVal(entry);
-            char buf[16];
-            sprintf(buf, "%04d", *instance);
-            RedisModuleString *mangled_key = RedisModule_CreateStringFromString(NULL, key);
-            RedisModule_StringAppendBuffer(NULL, mangled_key, buf, strlen(buf));
-            mangled_inkeys = array_append(mangled_inkeys, mangled_key);
+            int *ind = AI_dictGetVal(entry);
+            currentOp->inkeys_indices = array_append(currentOp->inkeys_indices, *ind);
         }
 
-        RedisModuleString **mangled_outkeys =
-            array_new(RedisModuleString *, array_len(currentOp->outkeys));
         for (long long j = 0; j < array_len(currentOp->outkeys); j++) {
             RedisModuleString *key = currentOp->outkeys[j];
-            AI_dictEntry *entry = AI_dictFind(mangled_tensors, key);
-            int *instance = NULL;
-            if (entry) {
-                instance = AI_dictGetVal(entry);
-                *instance += 1;
+            int *ind = RedisModule_Alloc(sizeof(int));
+            *ind = array_len(rinfo->dagSharedTensors);
+
+            // Add a new empty place holder in the array for an output tensor.
+            // If this is MODELSET op, the tensor is already realized.
+            if (currentOp->commandType == REDISAI_DAG_CMD_TENSORSET) {
+                RAI_Tensor *t = RAI_TensorGetShallowCopy(currentOp->outTensor);
+                rinfo->dagSharedTensors = array_append(rinfo->dagSharedTensors, t);
             } else {
-                instance = RedisModule_Alloc(sizeof(int));
-                *instance = 1;
-                AI_dictAdd(mangled_tensors, (void *)key, (void *)instance);
-            }
-            char buf[16];
-            sprintf(buf, "%04d", *instance);
-            RedisModuleString *mangled_key = RedisModule_CreateStringFromString(NULL, key);
-            RedisModule_StringAppendBuffer(NULL, mangled_key, buf, strlen(buf));
-            mangled_outkeys = array_append(mangled_outkeys, mangled_key);
-        }
-
-        if (currentOp->inkeys) {
-            for (size_t j = 0; j < array_len(currentOp->inkeys); j++) {
-                RedisModule_FreeString(NULL, currentOp->inkeys[j]);
-            }
-            array_free(currentOp->inkeys);
-        }
-
-        if (currentOp->outkeys) {
-            for (size_t j = 0; j < array_len(currentOp->outkeys); j++) {
-                RedisModule_FreeString(NULL, currentOp->outkeys[j]);
+                rinfo->dagSharedTensors = array_append(rinfo->dagSharedTensors, NULL);
             }
-            array_free(currentOp->outkeys);
+            currentOp->outkeys_indices = array_append(currentOp->outkeys_indices, *ind);
+            AI_dictReplace(tensorsNamesToInd, (void *)key, (void *)ind);
         }
-
-        currentOp->inkeys = mangled_inkeys;
-        currentOp->outkeys = mangled_outkeys;
-    }
-
-    AI_dict *mangled_persisted = AI_dictCreate(&AI_dictTypeHeapRStrings, NULL);
-    {
-        AI_dictIterator *iter = AI_dictGetSafeIterator(rinfo->dagTensorsPersistedContext);
-        AI_dictEntry *entry = AI_dictNext(iter);
-        while (entry) {
-            RedisModuleString *key = (RedisModuleString *)AI_dictGetKey(entry);
-            AI_dictEntry *mangled_entry = AI_dictFind(mangled_tensors, key);
-            if (!mangled_entry) {
-                AI_dictRelease(mangled_persisted);
-                AI_dictReleaseIterator(iter);
-                RAI_SetError(rinfo->err, RAI_EDAGRUN, "ERR PERSIST key cannot be found in DAG");
-                goto cleanup;
-            }
-            if (AI_dictFind(mangled_persisted, key) != NULL) {
-                AI_dictRelease(mangled_persisted);
-                AI_dictReleaseIterator(iter);
-                RAI_SetError(rinfo->err, RAI_EDAGRUN, "ERR PERSIST keys must be unique");
-                goto cleanup;
-            }
-            int *instance = AI_dictGetVal(mangled_entry);
-            char buf[16];
-            sprintf(buf, "%04d", *instance);
-            RedisModuleString *mangled_key = RedisModule_CreateStringFromString(NULL, key);
-            RedisModule_StringAppendBuffer(NULL, mangled_key, buf, strlen(buf));
-            AI_dictAdd(mangled_persisted, (void *)mangled_key, (void *)1);
-            RedisModule_FreeString(NULL, mangled_key);
-            entry = AI_dictNext(iter);
-        }
-        AI_dictReleaseIterator(iter);
     }
-
-    AI_dictRelease(rinfo->dagTensorsPersistedContext);
-    rinfo->dagTensorsPersistedContext = mangled_persisted;
-
-    for (long long i = 0; i < array_len(rinfo->dagOps); i++) {
-        if (rinfo->dagOps[i]->devicestr == NULL) {
-            rinfo->dagOps[i]->devicestr = "CPU";
-        }
-    }
-    // Tensors from TENSORSET ops are ready to be put in DAG local context under their mangled
-    // names.
-    _DAG_SetTensorsInLocalContext(rinfo);
-    res = REDISMODULE_OK;
-
-cleanup : {
-    AI_dictIterator *iter = AI_dictGetSafeIterator(mangled_tensors);
-    AI_dictEntry *entry = AI_dictNext(iter);
-    while (entry) {
-        int *val = (int *)AI_dictGetVal(entry);
-        RedisModule_Free(val);
-        entry = AI_dictNext(iter);
-    }
-    AI_dictReleaseIterator(iter);
-}
-    AI_dictRelease(mangled_tensors);
-    return res;
+    return REDISMODULE_OK;
 }
 
 // Add Shallow copies of the DAG run info to the devices' queues.
@@ -242,7 +146,7 @@ int RAI_DAGRun(RAI_DAGRunCtx *run_info, RAI_OnFinishCB DAGAsyncFinish, void *pri
     }
     // Make the inkeys and outkeys of the DAG ops unique, to ensure that the operations
     // will be execute in the right order.
-    if (MangleTensorsNames(rinfo) != REDISMODULE_OK) {
+    if (MapTensorsKeysToIndices(rinfo, rinfo->tensorsNamesToIndices) != REDISMODULE_OK) {
         RAI_SetError(err, rinfo->err->code, rinfo->err->detail);
         return REDISMODULE_ERR;
     }
@@ -269,16 +173,13 @@ size_t RAI_DAGNumOutputs(RAI_OnFinishCtx *finish_ctx) {
 const RAI_Tensor *RAI_DAGOutputTensor(RAI_OnFinishCtx *finish_ctx, size_t index) {
     size_t tensor_get_op_ind = -1;
     RedisAI_RunInfo *rinfo = (RedisAI_RunInfo *)finish_ctx;
+
     for (size_t i = 0; i < rinfo->dagOpCount; i++) {
         RAI_DagOp *op = rinfo->dagOps[i];
         if (op->commandType == REDISAI_DAG_CMD_TENSORGET) {
             tensor_get_op_ind++;
             if (tensor_get_op_ind == index) {
-                RAI_Tensor *t;
-                int res = RAI_getTensorFromLocalContext(rinfo->dagTensorsContext, op->inkeys[0], &t,
-                                                        op->err);
-                RedisModule_Assert(res == REDISMODULE_OK);
-                return t;
+                return Dag_GetInternalTensor(rinfo, op->inkeys_indices[0]);
             }
         }
     }

src/DAG/dag_execute.c

@@ -4,17 +4,39 @@
 #include "run_info.h"
 
 /**
- * @brief We are given a DAG runInfo of a sequence of operations, each with its own
+ @brief We are given a DAG runInfo of a sequence of operations, each with its own
  input and output keys. The names of the keys will be used to look whether the
  inputs to a DAG operation have all been realized by previous operations (or if
  they are available as part of LOADed keys from keyspace).
  This strategy is fine if keys are not aliased, that is, if a command's output
  overwrites the key of a previous command. This would trick DAG operations into
  thinking that their input is ready when it's not.
- To overcome this, we make key names unique, so that names are not aliased. We
- mangle the names by appending a numerical suffix ":0001". After computing, we
- demangle the keys in order to persist them.*/
-int MangleTensorsNames(RedisAI_RunInfo *rinfo);
+ To overcome this, we map the input and output tensors for every operation to indices,
+ in the following way. For every input of an operation having the key "x", we map the index
+ for which "x" was last mapped to when, it was an output of a previous operation.
+ For every output of an operation "y", we map the next available index in the array.
+ Every entry in the DAG array contains NULL (except for tensors that where loaded
+ before the DAG run starts).
+ @param rinfo The DAG runInfo.
+ @param tensorsNamesToInd A dict mapping every key name of a tensor that appeared
+ in DAG operation, to the maximal index of the DAG shared array for which they were mapped to.
+ @returns REDISMODULE_ERR if there exists an operation for which one of the input
+ tensors didn't appear as an output of a previous operation, REDISMODULE_OK otherwise
+ */
+int MapTensorsKeysToIndices(RedisAI_RunInfo *rinfo, AI_dict *tensorsNamesToInd);
+
+/**
+ * @brief Validates that tensors key names to persist appeared in the DAG operations.
+ * @param rinfo The DAG runInfo.
+ * @param tensorsNamesToInd A dict mapping every key name of a tensor that appeared
+ * in DAG operation, to the maximal index of the DAG shared array for which they were mapped to.
+ * @param persistTensorsNames A hash table the contains the names of the tensors
+ * to persist when the DAG run is finished.
+ * @return REDISMODULE_ERR if there exists a tensor key to persist that didn't
+ * appear in DAG operation, REDISMODULE_OK otherwise
+ */
+int ValidatePersistKeys(RedisAI_RunInfo *rinfo, AI_dict *tensorsNamesToInd,
+                        AI_dict *persistTensorsNames);
 
 /**
  * @brief Run asynchronously a DAG. This will validate that the sequence of DAG ops