Compute Modes

oneMKL - Data Parallel C++ Developer Reference

Download PDF

ID 772045

Date 3/31/2021

Version

Public

A newer version of this document is available. Customers should click here to go to the newest version.

Visible to Intel only — GUID: GUID-47B7D585-E87B-4FDA-A8E7-41818762B07E

View Details

Compute Modes

BLAS level-3 routines and extensions support alternate compute modes, which can provide increased performance in exchange for different numerical properties or reduced accuracy.

A list of one or more allowed modes can be specified either at compile time, on a per-call or per-source-file basis, or at runtime, using the MKL_BLAS_COMPUTE_MODE environment variable. oneMKL will automatically select an appropriate implementation from this list, taking into account routine parameters and hardware characteristics. In case none of the allowed alternate modes are supported by the given routine on the selected device, or if none of the allowed alternate modes are expected to improve performance, oneMKL will automatically fall back to a standard implementation.

NOTE:

Whether a particular mode provides additional performance will depend on a number of factors, including routine type, matrix sizes, transpose parameters, and hardware configuration. In the current oneMKL release, alternate implementations are available for gemm, gemmt, syrk, and syr2k on selected hardware.

By default, oneMKL does not enable any alternate compute modes. The MKL_BLAS_COMPUTE_MODE environment variable is intended for quickly evaluating whether alternate compute modes provide performance benefits and acceptable accuracy for an application. After initial testing, alternate mode settings can be permanently applied within the application using the per-call or per-source-file APIs.

When running on GPU, oneMKL’s verbose output indicates which mode is used for each call, whether the standard mode or one of the alternate modes. See Checking Which Mode Is Used for more details.

Mode Settings

Available alternate modes are described in the table below. Multiple modes can be combined, allowing oneMKL to choose any of the allowed modes that is expected to provide best performance.

For per-call or per-source-file mode, alternate compute modes are selected by OR’ing together one or more oneapi::mkl::blas::compute_mode values from the following table, e.g.:

using oneapi::mkl::blas;
auto mode_settings = compute_mode::float_to_bf16x2 | compute_mode::float_to_tf32; /* allow either of these two modes */

When making global changes with the MKL_BLAS_COMPUTE_MODE environment variable, multiple modes are combined with commas:

set MKL_BLAS_COMPUTE_MODE=FLOAT_TO_BF16X2,FLOAT_TO_TF32

`compute_mode` enum value	Environment variable setting	Description
`compute_mode::float_to_bf16`	`FLOAT_TO_BF16`	Convert single-precision inputs to bfloat16 format internally; output is accumulated in single precision. Accuracy will be reduced, with possibly much higher performance.
`compute_mode::float_to_bf16x2`	`FLOAT_TO_BF16X2`	Convert each single-precision input value to a sum of two bfloat16 values internally; output is accumulated in single precision. Expected accuracy is between that of standard single precision arithmetic and bfloat16 arithmetic. Note: infinite inputs may produce unexpected NaNs in the output.
`compute_mode::float_to_bf16x3`	`FLOAT_TO_BF16X3`	Convert each single-precision input value to a sum of three bfloat16 values internally; output is accumulated in single precision. Expected accuracy is comparable to standard single precision arithmetic in most cases. Note: infinite inputs may produce unexpected NaNs in the output.
`compute_mode::float_to_tf32`	`FLOAT_TO_TF32`	Convert each single-precision input value to tf32 format internally; output is accumulated in single precision. Expected accuracy is between `float_to_bf16` and `float_to_bf16x2`.
`compute_mode::complex_3m`	`COMPLEX_3M`	Reduce the four real multiplications in a standard complex multiplication to three real multiplications. Expected accuracy is comparable to standard arithmetic in most cases.
`compute_mode::any`	`ANY`	Allow any alternate compute mode.
`compute_mode::standard`	`STANDARD`	Do not allow any alternate compute modes.
`compute_mode::prefer_alternate`	`PREFER_ALTERNATE`	Used in conjunction with one or more alternate modes. Prefer an alternate compute mode over the standard implementation whenever available, even if performance may be reduced.
`compute_mode::force_alternate`	`FORCE_ALTERNATE`	Used in conjunction with one or more alternate modes. Never use the standard implementation; if no allowed alternate implementation is available, an exception will be thrown.

Per-Call Mode Settings

BLAS level-3 routines and extensions support an optional oneapi::mkl::blas::compute_mode argument to specify a desired mode setting, at the end of the parameter list. For USM APIs, the compute_mode argument goes before the list of input dependencies, if any; either argument may be omitted. For example:

using oneapi::mkl::blas;

sycl::buffer<float, 1> a_buffer, c_buffer;
float *a_ptr, *c_ptr;
/* ... */

// Buffer API
syrk(my_queue, n, k, uplo, trans, alpha, a_buffer, lda, beta, c_buffer, ldc, compute_mode::float_to_bf16);

// Buffer API, forcing float_to_bf16 mode
syrk(my_queue, n, k, uplo, trans, alpha, a_buffer, lda, beta, c_buffer, ldc, compute_mode::float_to_bf16 | compute_mode::force_alternate);

// USM API, without dependencies
syrk(my_queue, n, k, uplo, trans, alpha, a_ptr, lda, beta, c_ptr, ldc, compute_mode::float_to_bf16);

// USM API, with dependencies
syrk(my_queue, n, k, uplo, trans, alpha, a_ptr, lda, beta, c_ptr, ldc, compute_mode::float_to_bf16, {event1, event2});

// USM API, dependencies but no special compute_mode settings
syrk(my_queue, n, k, uplo, trans, alpha, a_ptr, lda, beta, c_ptr, ldc, {event1, event2});

Per-Source-File Mode Settings

You can provide default mode settings for all calls within a source file by defining the MKL_BLAS_COMPUTE_MODE macro before including any oneMKL header files. This macro must be set to an expression of type oneapi::mkl::blas::compute_mode.

#define MKL_BLAS_COMPUTE_MODE oneapi::mkl::blas::compute_mode::complex_3m
#include <oneapi/mkl.hpp>

void my_function() {
    /* ... */

    // 3M mode will be allowed by default:
    gemm(my_queue, m, n, k, trans_a, trans_b, alpha, a, lda, b, ldb, beta, c, ldc);
}

compute_mode parameters passed to a oneMKL routine take precedence over the default setting:

#define MKL_BLAS_COMPUTE_MODE oneapi::mkl::blas::compute_mode::complex_3m
#include <oneapi/mkl.hpp>

void my_function() {
    /* ... */

    // Provided compute_mode overrides the default 3M mode.
    gemm(my_queue, m, n, k, trans_a, trans_b, alpha, a, lda, b, ldb, beta, c, ldc, compute_mode::standard);
}

Runtime Mode Settings

The MKL_BLAS_COMPUTE_MODE environment variable allows you to set default application-wide alternate compute mode settings, as a quick method for evaluating alternate compute modes at runtime. For example, with bash or a similar shell:

// my_application.cpp
#include <oneapi/mkl.hpp>

int main() {
    /* ... */

    // Call to gemm without a compute_mode argument:
    oneapi::mkl::blas::gemm(my_queue, /* ... */);
}

export MKL_BLAS_COMPUTE_MODE=FLOAT_TO_BF16X3
my_application      # gemm may use bf16x3 arithmetic.

Any per-call or per-source-file mode settings take precedence over the MKL_BLAS_COMPUTE_MODE environment variable. One result of this is that compiling an application with -DMKL_BLAS_COMPUTE_MODE=oneapi::mkl::blas::compute_mode::standard effectively disables the environment variable.

Checking Which Mode Is Used

On GPU, when oneMKL’s verbose mode is enabled, information on the compute mode(s) enabled and used for each call is provided in the verbose log. For example, in the following call, float_to_bf16 and float_to_tf32 were both enabled, and float_to_bf16 was selected (some output omitted for clarity):

MKL_VERBOSE oneapi::mkl::blas::column_major::gemm[float](0x7ffd39046350,...,float_to_bf16|float_to_tf32) mode:float_to_bf16 host:nan device:nan GPU0

Verbose mode can be enabled by setting the MKL_VERBOSE environment variable to 1, or via the mkl_verbose API. For more information, see “Using oneMKL Verbose Mode” in the Intel(R) oneAPI Math Kernel Library Developer Guide, available in the Intel Software Documentation Library.

Select Your Language

Using Intel.com Search

Quick Links

Recent Searches

Advanced Search

Only search in

oneMKL - Data Parallel C++ Developer Reference

Compute Modes

Mode Settings

Per-Call Mode Settings

Per-Source-File Mode Settings

Runtime Mode Settings

Checking Which Mode Is Used