Add support for offline planning in case of multiple subgraphs (#1195)

In #947 support for multiple subgraphs was added to the memory planner. However, in case memory planning information was provided offline this would not work since it would only read the planning data for the first subgraph. This PR fixes that.

The fix itself is simple and local to the function, assuming all offline planned tensors are just concatenated, e.g.: `[s0t0, s0t1, s1t0, s2t1, s1t2, s3t0, s3t1, s3t2, s3t3]`. The documentation is updated to reflect this.

The metadata format also had a subgraph ID according to the documentation, but that was never read. I tried to use that but that wouldn't work either, because it would only hold one subgraph ID and we can only have one meta-data with a given name. So I'm not sure how this was intended to be used originally but I added a note to the documentation and the code to clarify that it is unused at the moment.

BUG=see PR description

Author: CNugteren

tensorflow/lite/micro/docs/memory_management.md

@@ -87,13 +87,18 @@ integers of the following format:
 | Offset | Value |
 |-|-|
 | 0 | Offline allocation format version |
-| 1 | Subgraph index to which this allocation applies |
+| 1 | Number of subgraphs |
 | 2 | Number offsets following: n |
 | 3 | Byte offset of tensor #0 or -1 to allocate at runtime |
 | 4 | Byte offset of tensor #1 or -1 to allocate at runtime |
 | ... | ... |
 | 3+(n-1) | Byte offset of tensor #(n-1) or -1 to allocate at runtime |
 
+Note that offsets 0 (the version) and 1 (the number of subgraphs) are currently
+ignored by the micro memory allocator. In case of multiple subgraphs, it assumes
+all tensors for all subgraphs are concatenated: all tensors for the first
+subgraph are first, followed by those of the second subgraph, etc.
+
 The `tflite::GreedyMemoryPlanner` treats the provided offline tensor allocation
 plan as constant fixed offset to the start of the head section and will attempt
 to fit any other tensors (such as scratch tensors added a runtime using the

tensorflow/lite/micro/micro_allocation_info.cc

@@ -179,6 +179,7 @@ TfLiteStatus AllocationInfoBuilder::InitializeAllocationInfo(
     const int32_t* offline_offsets, SubgraphAllocations* allocations) {
   AllocationInfo* allocation_info = info_.allocation_info;
   // Initialize allocation info for every tensor in every subgraph.
+  int offline_index = 0;
   for (size_t subgraph_idx = 0; subgraph_idx < model_->subgraphs()->size();
        subgraph_idx++) {
     const SubGraph* subgraph = model_->subgraphs()->Get(subgraph_idx);
@@ -203,15 +204,14 @@ TfLiteStatus AllocationInfoBuilder::InitializeAllocationInfo(
           (!subgraph->tensors()->Get(i)->is_variable()) &&
           (current->bytes != 0);
       if (offline_offsets) {
-        current->offline_offset = offline_offsets[i];
+        current->offline_offset = offline_offsets[offline_index++];
 
         // Mark offline planned variable tensors so they can get an offline
         // offset and be handled offline.
         if (subgraph->tensors()->Get(i)->is_variable() &&
             current->offline_offset != kOnlinePlannedBuffer) {
           current->needs_allocating = true;
         }
-
       } else {
         current->offline_offset = kOnlinePlannedBuffer;
       }