Bfloat16 Inference

Bfloat16 Inference Usage (C++)



Inference Engine with the bfloat16 inference implemented on CPU must support the native avx512_bf16 instruction and therefore the bfloat16 data format. It is possible to use bfloat16 inference in simulation mode on platforms with Intel® Advanced Vector Extensions 512 (Intel® AVX-512), but it leads to significant performance degradation in comparison with FP32 or native avx512_bf16 instruction usage.


Bfloat16 computations (referred to as BF16) is the Brain Floating-Point format with 16 bits. This is a truncated 16-bit version of the 32-bit IEEE 754 single-precision floating-point format FP32. BF16 preserves 8 exponent bits as FP32 but reduces precision of the sign and mantissa from 24 bits to 8 bits.


Preserving the exponent bits keeps BF16 to the same range as the FP32 (~1e-38 to ~3e38). This simplifies conversion between two data types: you just need to skip or flush to zero 16 low bits. Truncated mantissa leads to occasionally less precision, but according to investigations, neural networks are more sensitive to the size of the exponent than the mantissa size. Also, in lots of models, precision is needed close to zero but not so much at the maximum range. Another useful feature of BF16 is possibility to encode INT8 in BF16 without loss of accuracy, because INT8 range completely fits in BF16 mantissa field. It reduces data flow in conversion from INT8 input image data to BF16 directly without intermediate representation in FP32, or in combination of INT8 inference and BF16 layers.

See the `”BFLOAT16 – Hardware Numerics Definition” white paper”</a> for more bfloat16 format details.

There are two ways to check if CPU device can support bfloat16 computations for models: 1. Query the instruction set via system <tt>lscpu | grep avx512_bf16</tt> or <tt>cat /proc/cpuinfo | grep avx512_bf16</tt>. 2. Use @ref openvino_docs_IE_DG_InferenceEngine_QueryAPI “Query API” with <tt>METRIC_KEY(OPTIMIZATION_CAPABILITIES)</tt>, which should return <tt>BF16</tt> in the list of CPU optimization options:

@snippet snippets/Bfloat16Inference0.cpp part0

Current Inference Engine solution for bfloat16 inference uses Intel® Math Kernel Library for Deep Neural Networks (Intel® MKL-DNN) and supports inference of the significant number of layers in BF16 computation mode.

<h1>Lowering Inference Precision</h1>

Lowering precision to increase performance is <a href=”” >widely used</a> for optimization of inference. The bfloat16 data type usage on CPU for the first time opens the possibility of default optimization approach. The embodiment of this approach is to use the optimization capabilities of the current platform to achieve maximum performance while maintaining the accuracy of calculations within the acceptable range.

Bfloat16 data usage provides the following benefits that increase performance: 1. Faster multiplication of two BF16 numbers because of shorter mantissa of bfloat16 data. 2. No need to support denormals and handling exceptions as this is a performance optimization. 3. Fast conversion of float32 to bfloat16 and vice versa. 4. Reduced size of data in memory, as a result, larger models fit in the same memory bounds. 5. Reduced amount of data that must be transferred, as a result, reduced data transition time.

For default optimization on CPU, source model is converted from FP32 or FP16 to BF16 and executed internally on platforms with native BF16 support. This is done by setting <tt>KEY_ENFORCE_BF16</tt> <tt>YES</tt> in the <tt>PluginConfigParams</tt> for <tt>GetConfig()</tt>. The code below demonstrates how to check if the key is set:

@snippet snippets/Bfloat16Inference1.cpp part1

To disable BF16 internal transformations, set the <tt>KEY_ENFORCE_BF16</tt> to <tt>NO</tt>. In this case, the model infers as is without modifications with precisions that were set on each layer edge.

@snippet snippets/Bfloat16Inference2.cpp part2 To disable BF16 in C API:

@code ie_config_t config = { “ENFORCE_BF16”, “NO”, NULL}; ie_core_load_network(core, network, device_name, &config, &exe_network); @endcode

An exception with message <tt>Platform doesn’t support BF16 format</tt> is formed in case of setting <tt>KEY_ENFORCE_BF16</tt> to <tt>YES</tt> on CPU without native BF16 support or BF16 simulation mode.

Low-Precision 8-bit integer models cannot be converted to BF16, even if bfloat16 optimization is set by default. <br>

<h1>Bfloat16 Simulation Mode</h1>

Bfloat16 simulation mode is available on CPU and Intel® AVX-512 platforms that do not support the native <tt>avx512_bf16</tt> instruction. The simulator does not guarantee an adequate performance. To enable Bfloat16 simulator: * In @ref openvino_inference_engine_samples_benchmark_app_README “Benchmark App”, add the <tt>-enforcebf16=true</tt> option * In C++ API, set <tt>KEY_ENFORCE_BF16</tt> to <tt>YES</tt> * In C API: @code ie_config_t config = { “ENFORCE_BF16”, “YES”, NULL}; ie_core_load_network(core, network, device_name, &config, &exe_network); @endcode Performance Counters Information about layer precision is stored in the performance counters that are available from the Inference Engine API. The layers have the following marks:marks:Suffix for layers that had bfloat16 data type input and were computed in BF16 precisionprecisionSuffix for layers computed in 32-bit precision For example, the performance counters table for the Inception model can look as follows: @code pool5 EXECUTED layerType: Pooling realTime: 143 cpu: 143 execType: jit_avx512_BF16 fc6 EXECUTED layerType: FullyConnected realTime: 47723 cpu: 47723 execType: jit_gemm_BF16 relu6 NOT_RUN layerType: ReLU realTime: 0 cpu: 0 execType: undef fc7 EXECUTED layerType: FullyConnected realTime: 7558 cpu: 7558 execType: jit_gemm_BF16 relu7 NOT_RUN layerType: ReLU realTime: 0 cpu: 0 execType: undef fc8 EXECUTED layerType: FullyConnected realTime: 2193 cpu: 2193 execType: jit_gemm_BF16 prob EXECUTED layerType: SoftMax realTime: 68 cpu: 68 execType: jit_avx512_FP32 @endcode The column of the table includes inference primitives with specific suffixes. <>`__