Deep Neural Network Library (DNNL)  1.2.0
Performance library for Deep Learning
 All Classes Namespaces Files Functions Variables Typedefs Enumerations Enumerator Groups Pages
Version 1.0 Transition Guide

NOTE

Starting with version 1.1 Intel(R) MKL-DNN is renamed to DNNL. For consistency, only this guide uses Intel MKL-DNN nomenclature.

Introduction

This article describes user-visible and some important internal changes to Intel MKL-DNN that occurred between v0.20 and v1.0.

The v0.x branch (mnt-v0) is deprecated and users are strongly encouraged to migrate to v1.x.

See Also
Discussion on the API changes occurred in PR #384: RFC: API changes for the upcoming v1.0.

Summary of Changes

We tried to keep changes minimal to make migration as simple as possible. In particular, the Intel MKL-DNN programming model stays the same. Nevertheless, the new version brings a lot of incompatible changes requiring developers to revisit significant portions of the integrated code.

All changes can be split into the following groups:

  1. Minor API changes
  2. Improving the library robustness
  3. Simplified execution model
  4. Changes in memory description
  5. Changes in the build system

These groups are discussed in detail below.

1. Minor API Changes

1.1. Remove deprecated functionality

Deprecated functionality Replacement
ReLU primitive Eltwise with algorithm kind ReLU
ConvolutionReLU (single primitive) Convolution with ReLU as a post operation
Double precision scales Single precision scales
RNN backward pd w/o forward pd hint RNN backward pd w/ forward pd hint
mkldnn_omit_stats batch norm. flag mkldnn_use_global_stats
mkldnn_eltwise_desc_t.negative_slope mkldnn_eltwise_desc_t.alpha
mkldnn_rnn_cell_flags_t Not available anymore – RNN primitives are separated into RNN, LSTM, and GRU
mkldnn_padding_kind_t Not used anymore

The complete list of the removed C functions:

mkldnn_relu_forward_desc_init(...);
mkldnn_relu_backward_desc_init(...);
mkldnn_convolution_relu_desc_init(...);
mkldnn_rnn_cell_desc_init(...);
mkldnn_rnn_cell_get_gates_count(...);
mkldnn_rnn_cell_get_states_count(...);
mkldnn_rnn_forward_desc_init(...);
mkldnn_rnn_backward_desc_init(...);

The complete list of the removed C++ classes and functions:

struct mkldnn::convolution_relu_forward {}
struct mkldnn::relu_forward {}
struct mkldnn::relu_backward {}
struct mkldnn::rnn_cell {}
struct mkldnn::rnn_forward {}
struct mkldnn::rnn_backward {}
mkldnn::sum::primitive_desc(const memory::desc &output, std::vector<double> scale, std::vector<memory::primitive_desc> inputs);
mkldnn::sum::primitive_desc(std::vector<double> scale, std::vector<memory::primitive_desc> inputs);
mkldnn::eltwise_forward::desc(prop_kind aprop_kind, const memory::desc &src_desc, T negative_slope);
mkldnn::eltwise_backward::desc(const memory::desc &diff_data_desc, const memory::desc &data_desc, T negative_slope);

1.2. Rename foo_v2() to foo() and remove old foo() (C API only)

The functions like:

mkldnn_primitive_desc_create_v2(...);

were renamed to:

mkldnn_primitive_desc_create(...);

In v0.x, the foo_v2() functions typically were used to pass attributes, and foo() assumed empty attributes. In v1.0, the attributes parameter is mandatory. A user can still pass NULL to indicate that the default (empty) attributes should be used.

The list of functions that had the _v2 suffix:

mkldnn_primitive_desc_iterator_create_v2(...);
mkldnn_primitive_desc_create_v2(...);
mkldnn_reorder_primitive_desc_create_v2(...);

1.3. Remove s16 (int16_t) data type support

The experimental s16 data type is not supported any more and has been dropped.

1.4. Disallow setting the rounding mode

Rounding mode that was a part of attributes has been dropped. All computations respect the MXCSR register when performing rounding. Unless the rounding mode is set explicitly, rounding to the nearest even integer (RNE) is used.

1.5. Rename a few types, enumerations, and functions

1.5.1. Types

API v0.x v1.0
C mkldnn_batch_normalization_flag_t mkldnn_normalization_flags_t
C mkldnn_format_t mkldnn_format_tag_t
C++ mkldnn::batch_normalization_flag mkldnn::normalization_flags
C++ mkldnn::memory::format mkldnn::memory::format_tag

1.5.2. Enumerations

API v0.x v1.0
C mkldnn_fuse_bn_relu mkldnn_fuse_norm_relu
C++ mkldnn::fuse_bn_relu mkldnn::normalization_flags::fuse_norm_relu
C++ mkldnn::query::eengine mkldnn::query::engine

1.5.3. Functions

API v0.x v1.0
C mkldnn_memory_desc_init() mkldnn_memory_desc_init_by_tag()

1.6. Unscoped enumerations become scoped (C++ API only)

All enum became enum class. This requires the following changes:

Type Value in v0.x Value in v1.0
mkldnn::prop_kind mkldnn::forward_inference mkldnn::prop_kind::forward_inference
mkldnn::algorithm mkldnn::eltwise_tanh mkldnn::algorithm::eltwise_tanh
mkldnn::normalization_flags mkldnn::fuse_bn_norm_relu mkldnn::normalization_flags::fuse_norm_relu
mkldnn::query mkldnn::eengine mkldnn::query::engine
mkldnn::memory::data_type mkldnn::memory::f32 mkldnn::memory::data_type::f32
mkldnn::memory::format_tag mkldnn::memory::nchw mkldnn::memory::format_tag::nchw

1.7. Remove view primitive

Version 0.x had an implementation of view that was simply an alias for memory. In Intel MKL-DNN v1.0, we removed view as a type and replaced it with a memory descriptor directly. In order to initialize sub-memory, use dnnl::memory::desc::submemory_desc( "mkldnn::memory::desc::submemory_desc()").

See Also
For more detail, refer to section 4. View rework of the RFC for v1.0.

1.8. RNN-specific changes

Each type of RNN (Vanilla RNN, LSTM, and two types of GRU) is now initialized by a separate function/operation descriptor constructor.

For instance, instead of using mkldnn::rnn_forward with specified RNN types a user is expected to use:

Also, the hidden and cell states in LSTM are now separated. This means that instead of one src_iter tensor of shape (layers, directions, states, batch, channels) a user passes src_iter tensor of shape (layers, directions, batch, channels) for hidden states and src_iter_c tensor of shape (layers, directions, batch, channels) for cell states. The same applies to dst_iter; the hidden state and the cell state are split into dst_iter and dst_iter_c respectively.

1.9. GEMM API changes

Intel MKL-DNN provides three GEMM-like functions:

With version 1.0 we switched from a Fortran-style to a C-style API, meaning that the parameters are passed by value rather than by address, and matrices are assumed to be in row-major format rather than column-major format.

Moreover, to broaden the applicability of integer matrix-matrix multiply functions we changed the formula from:

\[ C_{s32} = \alpha \cdot (op(A_{i8}) + o_A) \cdot (op(B_{s8}) + o_B) + \beta \cdot C_{s32} + o_C \]

to

\[ C_{s32} = \alpha \cdot (op(A_{i8}) - o_A) \cdot (op(B_{s8}) - o_B) + \beta \cdot C_{s32} + o_C \]

where for both mkldnn_gemm_u8s8s32() and mkldnn_gemm_s8s8s32() the types of offsets for matrices A and B correspond to the type of the matrices themselves; that is:

1.10. Primitive descriptor queries for memory descriptors

In version 0.x when querying the primitive descriptor for a memory descriptor that is not used, the C API returned NULL and the C++ API threw an exception. In version 1.0, both the C and C++ APIs return a zero memory descriptor.

Zero memory descriptor means that the number of dimensions equals 0 and all the fields are set to zero. A memory object created with such a memory descriptor does not require any buffer allocations.

These changes enable simplifying the code that handles workspace or scratchpad:

// The code works fine even if scratchpad is not required.
    // In this case the memory would be just zero memory.
    auto scratchpad_md = pd.scratchpad_desc();
    auto scratchpad = memory(scratchpad_md, pd.get_engine());
    primitive.execute(stream, {
...,
{MKLDNN_SCRATCHPAD, scratchpad}};

1.11. Default constructors for C++ classes (C++ API only)

In Intel MKL-DNN v1.0, all C++ objects (primitives, memory objects, engines, and streams) now have default empty constructors. This enables defining the object, and then initializing it later on. An attempt to use any methods of an uninitialized object will result in the throwing of an exception.

This improvement can be especially useful when Intel MKL-DNN objects are members of the user's classes. For example:

class RELU_layer {
public:
RELU_layer() {} // no need to initialize eltwise here
void init() {
...
// deferred initialization
eltwise = eltwise_forward(...);
}
private:
eltwise_forward eltwise;
};

2. Improving the Library Robustness

2.1. Memory allocation in the C API

In Intel MKL-DNN v1.0, constructing a memory object using special value MKLDNN_MEMORY_ALLOCATE for a handle results in the buffer being allocated by the library. This makes the behavior of the C API memory object constructor aligned with its C++ API mkldnn::memory counterpart. Note that the C++ API memory object class still has an extra constructor that doesn't take a handle at all, and asks the library to allocate the buffer (that is, the same behavior as calling with the handle equal to MKLDNN_MEMORY_ALLOCATE).

2.2. Explicit scratchpad management

Intel MKL-DNN primitives may require temporary scratchpad memory for storing intermediate computational results. For instance, convolution backward by weights typically requires extra space to perform a reduction of the diff_weights computed by different threads (the work is divided across images). Starting with version 1.0, the library supports two modes:

  1. Implicit scratchpad, managed by the library (default). See mkldnn::scratchpad_mode::library.
  2. Explicit scratchpad, provided by the user. See mkldnn::scratchpad_mode::user.

The former mode matches the behavior of Intel MKL-DNN v0.x. It is kept for user convenience and cases in which memory is not a concern.

In the explicit scratchpad mode, a new mkldnn_query_scratchpad_md query will return the amount of scratchpad memory needed for a primitive, and the user will be responsible for allocating and providing the scratchpad memory to a primitive at runtime. The explicit scratchpad mode should be explicitly enabled by passing an attribute with mkldnn::scratchpad_mode::user to primitive descriptors.

Warning
Scratchpad memory is not the same as workspace.

With explicit scratchpad it is possible to make Intel MKL-DNN primitives stateless and hence thread safe: the same primitive can be executed in multiple independent threads as long as different threads use different scratchpads.

However, if a user chooses implicit scratchpad mode, there is no thread-safety guarantee.

3. Simplified Execution Model

This is the most notable change in the library. The main idea was to change the execution API so that memory arguments are specified at primitive execution time and not at primitive creation time. This leads to the following changes.

3.1. Memory is not a primitive anymore

In version 0.x, memory had a type of primitive. With the new API, memory becomes a distinct data type. Moreover, a memory primitive descriptor becomes redundant and has been dropped. The functions that use memory primitive descriptors now take memory descriptor and (optionally) engine, if the latter cannot be inferred.

These changes bring new data types and functions, such as:

#define MKLDNN_NATIVE_HANDLE_ALLOCATE ((void *)-1)
#define MKLDNN_NATIVE_HANDLE_NONE ((void *)0)
struct mkldnn_memory_t; // memory type, no more equal to mkldnn_primitive_t
// create a memory
// native_handle can:
// - point to the user allocated memory, i.e. valid handle. In this case the
// library doesn't own allocated memory.
// - be MKLDNN_NATIVE_HANDLE_ALLOCATE to ask the library to allocate and
// attach memory. In this case the library owns allocated memory.
// - be MKLDNN_NATIVE_HANDLE_NONE to create mkldnn_memory w/o attached memory.
mkldnn_status_t mkldnn_memory_create(mkldnn_memory_t *mem,
const mkldnn_memory_desc_t *md, mkldnn_engine_t engine,
void *handle);

3.2. Operation primitives cannot be used as inputs (use memory instead)

Version 0.x allowed passing an operation primitive as an input to another primitive. For instance, a convolution primitive could be passed as an input to a consequent ReLU. During the execution the ReLU primitive queried the convolution for its output memory and used it as an input.

In version 1.0, users are allowed to pass only memory type as inputs and outputs for primitives.

3.3. Remove the mkldnn_primitive_at_t type

Another consequence is that mkldnn_primitive_at_t, which is logically equivalent to {primitive, output_index}, becomes redundant. Previously the type was used to specify the exact memory to use (if a primitive had several outputs).

3.4. Passing stream and input/output memories at primitive execution

Finally, users are now able to directly run primitives by calling an execute function instead of putting primitives into a stream and running the latter. This change affects how primitives interact with streams and input/output memory objects: with the new API they become arguments to be passed to the primitive execution function.

The change significantly simplifies primitive creation, which now requires a primitive descriptor only:

mkldnn_status_t mkldnn_primitive_create(mkldnn_primitive_t *primitive,
const_mkldnn_primitive_desc_t *pd);

To remove the ambiguity in which order input and output memories need to be passed, we introduced a map-like argument in which each memory argument is paired with a tag indicating what kind of argument it is: destination, source, weights, and so on.

// types
#define MKLDNN_ARG_SRC_0 1
#define MKLDNN_ARG_SRC MKLDNN_ARG_SRC_0
#define MKLDNN_ARG_FROM MKLDNN_ARG_SRC_0
// ...
// C API
typedef struct {
int arg; // MKLDNN_ARG_SRC, ...
mkldnn_memory_t memory;
} mkldnn_exec_arg_t;
mkldnn_status_t mkldnn_primitive_execute(mkldnn_primitive_t prim,
mkldnn_stream_t stream, int nargs, const mkldnn_exec_arg_t *args);
// C++ API
convolution_forward::execute(mkldnn::stream &stream,
const std::map<int, mkldnn::memory> &exec_args);
// ... other primitives ...
// example C, convolution forward w/ bias
mkldnn_exec_arg_t conv_exec_args[] = {
{MKLDNN_ARG_SRC, src_mem},
{MKLDNN_ARG_WEIGHTS, weights_mem},
{MKLDNN_ARG_BIAS, bias_mem},
{MKLDNN_ARG_DST, dst_mem},
};
mkldnn_primitive_execute(conv_fwd, stream, 4, conv_exec_args);
// example C++, in-place eltwise
eltwise.execute(stream, {{MKLDNN_ARG_SRC, mem}, {MKLDNN_ARG_DST, mem}});

3.5 Short summary

The example below shows conceptual code transformations between versions. The C++ API is used for brevity.

Version 0.x:

// create a convolution, specify all inputs and outputs
auto conv = convolution(conv_pd,
{src_mem, 0}, {wei_mem, 0}, dst_conv_mem);
// create a relu (note that one of inputs is the convolution)
auto relu = relu(relu_pd,
{conv, 0}, dst_relu_mem);
// create a stream, submit convolution and relu, and wait for the result
stream().submit({conv, relu}).wait();

Version 1.0:

// create convolution and relu. no inputs/outputs
auto conv = convolution(conv_pd);
auto relu = relu(relu_pd);
// create stream (based on engine)
stream s(engine, 0);
// execute the convolution with given inputs, outputs
conv.execute(s, {
{MKLDNN_ARG_SRC, src_mem},
{MKLDNN_ARG_WEIGHTS, wei_mem},
{MKLDNN_ARG_DST, dst_conv_mem}});
// execute the relu. cannot pass convolution as an input, only memory is allowed
relu.execute(s, {
{{MKLDNN_ARG_SRC, dst_conv_mem},
{MKLDNN_ARG_DST, dst_relu_mem}});
s.wait(); // wait for async streams

4. Changes in Memory Description

The way of describing memory format in version 0.x had multiple issues. From the user's perspective, the main issues were:

There were more substantial issues from the library development perspective: code bloat to support special cases, etc.

We addressed the issues above by reworking memory descriptors. From the user's perspective, the main changes are:

  1. Memory descriptors support arbitrary strides for plain layouts. For example, initializing a memory descriptor with strides={h*w, o*h*w, w, 1} should be a valid way to define iohw format even if Intel MKL-DNN does not support it explicitly. Functions to use:
  2. Dimensions are of type int64_t instead of int, and the maximum number of tensor dimensions is decreased from 16 to 12. The mkldnn_strides_t is removed; use mkldnn_dims_t instead.
  3. The memory_desc_t.format field is replaced with memory_desc_t.format_kind, which also has different semantics.

While the first two items are self-explanatory, the last one requires some elaboration.

In version 0.x, most memory formats could be described directly by using appropriate format names (for example, nchw) that fully describe how data is laid out in memory. However, Intel MKL-DNN also had the blocked memory format and the corresponding memory_desc_t.layout_desc.blocking_desc structure, which could describe a memory format in a unified fashion by specifying block sizes and strides. The original idea was to use format tags like nchw during memory descriptor initialization only, and always use the blocked format internally. Unfortunately, that was never implemented.

With the new design, Intel MKL-DNN starts distinguishing between the actual memory format and convenience memory format tags that can be used to describe memory format concisely.

Users are still able to initialize memory descriptors with format tags like nchw using mkldnn::memory::desc::desc(dims, data_type, format_tag) or mkldnn_memory_desc_init_by_tag(), but the memory_desc_t.format_kind is set to a canonicalized kind like blocked, and the format name is not recorded in the memory descriptor structure. Initialization with strides will always result in blocked format. The API also uses different types for memory format tags and kinds to aid correctness.

For more details, refer to the Memory descriptor article of the RFC for v1.0.

5. Changes in the Build System

The build options were slightly changed in the new version of Intel MKL-DNN. That was done mainly to avoid name collisions with other projects that include Intel MKL-DNN as a subproject and to accommodate future extensions to the library. The change are:

Old option New option Notes
WITH_EXAMPLE MKLDNN_BUILD_EXAMPLES
WITH_TEST MKLDNN_BUILD_TESTS
MKLDNN_THREADING MKLDNN_CPU_RUNTIME
MKLDNN_USE_MKL N/A Intel MKL-DNN does not use Intel MKL anymore
VTUNEROOT N/A Not required, as Intel MKL-DNN contains all the necessary code internally

By default, the -Werror flag is disabled. MKLDNN_WERROR controls the behavior.

For more information about build options, refer to Build Options.