Contents

Persistent Cache

Creating oneDNN abstractions can be costly for various reasons. Usually, oneDNN mitigates that overhead by caching such objects but cache has no effect when the objects are created for the first time. For some applications, reducing that overhead is critical.

oneDNN provides an API that can be used to create a persistent cache for oneDNN abstractions. Through the API, users can obtain a cache blob ID and a cache blob to use as a key and value respectively.

Note

Content and size of the cache blob ID and cache blob objects are not specified.

Note

The oneDNN version and git commit hash (dnnl_version_t::hash) affect the equality of the cache blob IDs. That is, the queried cache blob ID will differ for different oneDNN versions and git commit hashes.

Warning

The git commit hash may not be available if the git package is not found during a CMake call. In this case, the cache blob ID will be the same for different hashes. This may result in fetching a wrong cache blob from persistent cache.

Primitive

Relation to Primitive Cache

When a primitive is created from a cache blob and the identical primitive is present in the primitive cache, the one from primitive cache will be returned to the user, and the given cache blob will not be used. Otherwise, the cache blob will be used to speed up the primitive creation. The information about how the primitive was created (cache_miss, cache_hit or persistent_cache_hit) is part of the verbose output for verbose level 2 (Verbose Mode).

API Usage Example

The following pseudo-code demonstrates a simple example of persistent cache implementation for primitives using the oneDNN API:

using namespace dnnl;

{
    convolution_forward::primitive_desc conv_pd(desc, attr, engine);
    convolution_forward conv(conv_pd);

    std::vector<uint8_t> key = conv_pd.get_cache_blob_id();
    std::vector<uint8_t> value = conv.get_cache_blob();
    store_cache_blob_on_disk(key, value);
}

{
    convolution_forward::primitive_desc conv_pd(desc, attr, engine);
    std::vector<uint8_t> key = conv_pd.get_cache_blob_id();
    std::vector<uint8_t> value = load_cache_blob_from_disk(key);
    convolution_forward conv_from_cache_blob(conv_pd, value);
}

Engine

API Usage Example

The following pseudo-code demonstrates a simple example of persistent cache implementation for OpenCL engines using the oneDNN API:

using namespace dnnl;

{
    cl_device_id device = ...;
    cl_context context = ...;

    engine ocl_engine = ocl_interop::make_engine(device, context);
    std::vector<uint8_t> key = get_engine_cache_blob_id(ocl_interop::get_device(ocl_engine));
    std::vector<uint8_t> value = get_engine_cache_blob(ocl_engine);

    store_cache_blob_on_disk(key, value);
}

{
    cl_device_id device = ...;
    cl_context context = ...;

    std::vector<uint8_t> key = get_engine_cache_blob_id(device);
    std::vector<uint8_t> value = load_cache_blob_from_disk(key);
    engine ocl_engine = ocl_interop::make_engine(device, context, value);
}

Memory descriptor

When serializing primitives, a binary blob can be obtained from a memory descriptor using dnnl::memory::desc::get_blob. Any binary blob obtained from dnnl::memory::desc::get_blob can be used to create a memory descriptor dnnl::memory::desc.

Note

When deserializing a constant tensor, the user must verify that the deserialized memory descriptor matches the memory descriptor expected by the primitive that will use that memory. The only circumstance where both are guaranteed to match is when serialization/deserialization happens on the same system and in the same environment.

Limitations

  • The primitive and engine APIs are implemented for OpenCL runtime only. For CPU engine and other runtimes, the library will return dnnl_unimplemented (in the case of the C API) or throw a corresponding dnnl::error exception (in the case of the C++ API).

  • Currently, the library cannot differentiate cache blobs created for devices that have different stepping; therefore, the cache blob can be safely used only on the system where it is created.

Primitive Cache Using oneDNN with Threadpool-Based Threading