Configuration entry reference

The following configuration entries are available in .dace.conf, as part of the API, and as environment variables. See Configuring DaCe for more information on how to use the interface.

cache

Compiled cache entry naming policy

• Type: str

• Description: Determine the name of the generated .dacecache folder:

  • name uses the name of the SDFG directly, causing it to be overridden by other programs using the same SDFG name.

  • hash uses a mangled name based on the hash of the SDFG, such that any change to the SDFG will generate a different cache folder.

  • unique uses a name based on the currently running Python process at code generation time, such that no caching or clashes can happen between different processes or subsequent invocations of Python.

  • single uses a single cache folder for all SDFGs, saving space and potentially build time, but disallows executing SDFGs in parallel and caching of more than one simultaneous SDFG.

• Default value: name

call_hooks

Hooks before/after every DaCe program call

• Type: str

• Description: A comma-separated list of functions (or Context Manager classes) that will be called before every DaCe program (SDFG) is compiled and run. Used for functionality such as automatic tuning or instrumentation.

• Default value: (Empty)

compiled_sdfg_call_hooks

Hooks before/after every compiled SDFG call

• Type: str

• Description: A comma-separated list of functions (or Context Manager classes) that will be called before every compiled SDFG’s generated code is invoked. Used for functionality such as low-level profiling.

• Default value: (Empty)

debugprint

Debug printing

• Type: bool

• Description: Enable verbose printouts.

• Default value: False

default_build_folder

Default SDFG build folder

• Type: str

• Description: Default folder in which compiled DaCe programs and SDFGs are stored. Can either be a relative path (by default) or absolute.

• Default value: .dacecache

profiling

Profiling

• Type: bool

• Description: Enable profiling support.

• Default value: False

profiling_status

Status bar for profiling

• Type: bool

• Description: Enable tqdm status bar while profiling. If tqdm is not installed a warning will appear. To disable this feature (and the warning) set this option to false.

• Default value: True

progress

Progress reports

• Type: bool

• Description: Enable progress report printouts.

• Default value: True

store_history

Store SDFG transformation history

• Type: bool

• Description: Store the history of transformations on the SDFG file.

• Default value: True

treps

Profiling Repetitions

• Type: int

• Description: Number of times to run program for profiling.

• Default value: 100

compiler

Preferences of the compiler

compiler.allow_shadowing

Allow variable shadowing

• Type: bool

• Description: Allowing shadowing of variables in the code (reduces exceptions to warnings when shadowing is encountered).

• Default value: True

compiler.allow_view_arguments

Allow numpy views as arguments

• Type: bool

• Description: If true, allows users to call DaCe programs with NumPy views (for example, “A[:, 1]” or “w.T”). As this can create pointer aliasing issues with two arrays pointing to the same memory, or analyzability issue with strides and alignment, this option is disabled by default.

• Default value: False

compiler.build_type

Build configuration

• Type: str

• Description: Configuration type for CMake build (can be Debug, Release, RelWithDebInfo, or MinSizeRel).

• Default value: RelWithDebInfo

compiler.codegen_lineinfo

Annotate code generator lines

• Type: bool

• Description: Keep a source mapping between generated code and the file/line of the code generator that generated it. Used for debugging code generation.

• Default value: False

compiler.codegen_state_struct_suffix

Suffix used by the code generator to mangle the state struct.

• Type: str

• Description: For every SDFG the code generator is is processing a state struct is generated. The typename of this struct is derived by appending this value to the SDFG’s name. Note that the suffix may only contains letters, digits and underscores.

• Default value: _state_t

compiler.default_data_types

Default data types

• Type: str

• Description: Specify the default data types to use in generating code. If “Python”, Python’s semantics will be followed (i.e., float and int are represented using 64 bits). If the property is set to “C”, C’s semantcs will be used (float and int are represented using 32bits).

• Default value: Python

compiler.extra_cmake_args

Additional CMake configuration arguments

• Type: str

• Description: If set, specifies additional arguments to the initial invocation of cmake.

• Default value: (Empty)

compiler.indentation_spaces

Indentation width

• Type: int

• Description: Number of spaces used when indenting generated code.

• Default value: 4

compiler.inline_sdfgs

Inline all nested SDFGs

• Type: bool

• Description: If set to true, inlines all nested SDFGs upon code generation by default.

• Default value: False

compiler.library_extension

Library extension

• Type: str

• Description: File extension of shared libraries.

• Default value: so

• Default value (on Linux): so

• Default value (on Windows): dll

• Default value (on Darwin): dylib

compiler.library_prefix

Library prefix

• Type: str

• Description: Filename prefix for shared libraries.

• Default value: (Empty)

• Default value (on Linux): lib

• Default value (on Darwin): lib

compiler.max_stack_array_size

Max stack-allocated array size (bytes)

• Type: int

• Description: All stack allocated arrays (i.e. StorageType.Register) with size larger than this will be allocated on the heap.

• Default value: 65536

compiler.unique_functions

Generate unique functions

• Type: str

• Description: Determine if and how to generate the code for equivalent NestedSDFGs: “hash”: hashing is used to determine if multiple NestedSDFGs with equivalent contents exist. If this is the case, the code is generated only once. “unique_name”: the unique_name property of SDFG is used to determine if two NestedSDFGs are equal, generating the code only once. This gives more control to the programmer, that can explicitly decide what NestedSDFG code can be replicated and what not. “none”: a separate function is code generated for each NestedSDFG

• Default value: hash

compiler.use_cache

Use cache

• Type: bool

• Description: If enabled, does not recompile code generated from SDFGs if shared library (.so/.dll) file is present.

• Default value: False

compiler.cpu

CPU compiler preferences

compiler.cpu.args

Arguments

• Type: str

• Description: Compiler argument flags

• Default value: -std=c++14 -fPIC -Wall -Wextra -O3 -march=native -ffast-math -Wno-unused-parameter -Wno-unused-label

• Default value (on Windows): /O2 /fp:fast /arch:AVX2 /D_USRDLL /D_WINDLL /D__restrict__=__restrict

compiler.cpu.executable

Compiler executable override

• Type: str

• Description: File path or name of compiler executable

• Default value: (Empty)

compiler.cpu.libs

Additional libraries

• Type: str

• Description: Additional linked libraries required by target

• Default value: (Empty)

compiler.cpu.openmp_sections

Use OpenMP sections

• Type: bool

• Description: If set to true, multiple connected components will generate “#pragma omp parallel sections” code around them.

• Default value: False

compiler.cuda

GPU (CUDA/HIP) compiler preferences

compiler.cuda.args

nvcc Arguments

• Type: str

• Description: Compiler argument flags for CUDA

• Default value: -Xcompiler -march=native --use_fast_math -Xcompiler -Wno-unused-parameter

• Default value (on Windows): -O3 --use_fast_math

compiler.cuda.backend

Compilation backend

• Type: str

• Description: Backend to compile for (‘auto’ for automatic detection, ‘cuda’ for NVIDIA, or ‘hip’ for AMD).

• Default value: auto

compiler.cuda.block_size_lastdim_limit

Maximum last dimension thread-block size in code generation

• Type: int

• Description: Threshold for the GPU code generator to fail in generating a kernel with a specified larger block size in the third dimension. Default value is derived from hardware limits on common GPUs.

• Default value: 64

compiler.cuda.block_size_limit

Maximum thread-block size in code generation

• Type: int

• Description: Threshold for the GPU code generator to fail in generating a kernel with a specified overall larger block size. Default value is derived from hardware limits on common GPUs.

• Default value: 1024

compiler.cuda.cuda_arch

Additional CUDA architectures

• Type: str

• Description: Additional CUDA architectures (separated by commas) to compile GPU code for, excluding the current architecture on the compiling machine.

• Default value: 60

compiler.cuda.default_block_size

Default thread-block size

• Type: str

• Description: Default thread-block size for GPU kernels when explicit GPU block maps are not defined. Can be set to ‘max’ to maximize occupancy.

• Default value: 32,1,1

compiler.cuda.dynamic_map_block_size

Thread-Block size for GPU_ThreadBlock_Dynamic

• Type: str

• Description: Thread-Block size for maps using GPU_ThreadBlock_Dynamic scheduler. Can be set to ‘max’ to maximize occupancy.

• Default value: 128,1,1

compiler.cuda.dynamic_map_fine_grained

Enable fine grained load balancing for GPU_ThreadBlock_Dynamic

• Type: bool

• Description: If true the scheduler will dynamically redistribute the combined work of all threads in the warp equally across the warp (fine grained). Otherwise, each warp works sequentially only on its tasks (potential load imbalance).

• Default value: True

compiler.cuda.hip_arch

Additional HIP architectures

• Type: str

• Description: Additional HIP architectures (separated by commas) to compile GPU code for, excluding the current architecture on the compiling machine.

• Default value: gfx906

compiler.cuda.hip_args

hipcc Arguments

• Type: str

• Description: Compiler argument flags for HIP

• Default value: -std=c++17 -fPIC -O3 -ffast-math -Wno-unused-parameter

compiler.cuda.libs

Additional libraries

• Type: str

• Description: Additional linked libraries required by target

• Default value: (Empty)

compiler.cuda.max_concurrent_streams

Concurrent execution streams

• Type: int

• Description: Maximum number of concurrent CUDA/HIP streams to generate. Special values: -1 only uses the default stream, 0 uses infinite concurrent streams.

• Default value: 0

compiler.cuda.mempool_release_threshold

Memory pool memory release threshold

• Type: int

• Description: A value that determines how large a memory allocation has to be before it is automatically released from the memory pool to the system. The default is -1, which indicates “never release”. Other values may be 0 (always release), or any byte value. For more information, see cudaMemPoolAttrReleaseThreshold in the CUDA toolkit documentation.

• Default value: -1

compiler.cuda.path

CUDA/HIP path override

• Type: str

• Description: Path to CUDA toolkit or ROCm/HIP root directory

• Default value: (Empty)

compiler.cuda.persistent_map_SM_fraction

Fraction of SMs to use for persistent GPU map

• Type: float

• Description: Sets the fraction of the number of SMs of the Device that the GPU_Persistent map can use. Together with persistent_map_occupancy this specifies the grid size of the kernel being launched. 0.0 < persistent_map_SM_fraction <= 1.0 The fraction will be rounded up to the next integer number of SMs. The max value of SMs that can/will be used is equal to cudaDevAttrMultiProcessorCount.

• Default value: 1.0

compiler.cuda.persistent_map_occupancy

Number of blocks to launch per SM used

• Type: int

• Description: Sets the number of thread block to be launched per SM being used. Essentially this is a simple multiplier to persistent_map_SM_fraction. It is up to the user to check if the resulting number of thread blocks can run efficiently on the GPU.

• Default value: 2

compiler.cuda.syncdebug

Synchronous Debugging

• Type: bool

• Description: Enables Synchronous Debugging mode, where each library call is followed by full-device synchronization and error checking.

• Default value: False

compiler.cuda.thread_id_type

Thread/block index data type

• Type: str

• Description: Defines the data type for a thread and block index in the generated code. The type is based on the type-classes in dace.dtypes. For example, uint64 is equivalent to dace.uint64. Change this setting when large index types are needed to address memory offsets that are beyond the 32-bit range, or to reduce memory usage.

• Default value: int32

compiler.fpga

Common preferences for FPGA compilation.

compiler.fpga.autobuild_bitstreams

Automatically build bitstreams

• Type: bool

• Description: If set to true, CMake will automatically build missing bitstreams when running an FPGA program. This can take a very long time, and users might want to do this manually. If set to false, the program will optimistically assume that the bitstream is present in the build directory, and will crash if this is not the case.

• Default value: True

compiler.fpga.concurrent_kernel_detection

Detect parts of an SDFG that can run in parallel

• Type: bool

• Description: If set to false, DaCe will place each weakly connected component found in an SDFG state in a different Kernel/Processing Element. If true, a heuristic will further inspect each independent component for other parallelism opportunities (e.g., branches of the SDFG that can be executed in parallel), creating the corresponding kernels.

• Default value: False

compiler.fpga.minimum_fifo_depth

Minimum depth of FIFOs

• Type: int

• Description: Sets the minimum depth of any generated FIFO.

• Default value: (Empty)

compiler.fpga.vendor

FPGA vendor

• Type: str

• Description: Target Xilinx (“xilinx”) or Intel (“intel_fpga”) FPGAs when generating code.

• Default value: xilinx

compiler.intel_fpga

Intel FPGA compiler preferences.

compiler.intel_fpga.board

Target FPGA board

• Type: str

• Description: FPGA board to compile for, obtain list by running aoc --list-boards.

• Default value: a10gx

compiler.intel_fpga.enable_debugging

Enable debugging for hardware kernels

• Type: bool

• Description: Injects debugging cores where available.

• Default value: False

compiler.intel_fpga.host_flags

Host arguments

• Type: str

• Description: Extra host compiler argument flags

• Default value: -Wno-unknown-pragmas

compiler.intel_fpga.kernel_flags

Kernel flags

• Type: str

• Description: High-level synthesis C++ flags

• Default value: -fp-relaxed -cl-no-signed-zeros -cl-fast-relaxed-math -cl-single-precision-constant -no-interleaving=default

compiler.intel_fpga.mode

Compilation mode

• Type: str

• Description: Target of FPGA kernel build (emulator/simulator/hardware).

• Default value: emulator

compiler.intel_fpga.path

Intel FPGA OpenCL SDK installation override

• Type: str

• Description: Path to specific Intel FPGA OpenCL SDK installation to use instead of just searching PATH and environment variables.

• Default value: (Empty)

compiler.linker

Linker preferences

compiler.linker.args

Arguments

• Type: str

• Description: Linker argument flags

• Default value: -Wl,--disable-new-dtags

• Default value (on Darwin): (Empty)

• Default value (on Windows): (Empty)

compiler.linker.executable

Linker executable override

• Type: str

• Description: File path or name of linker executable

• Default value: (Empty)

compiler.mpi

MPI compiler preferences

compiler.mpi.executable

Compiler executable override

• Type: str

• Description: File path or name of compiler executable

• Default value: (Empty)

compiler.rtl

RTL (SystemVerilog) compiler preferences

compiler.rtl.verbose

Verbose Build & Execution Output

• Type: bool

• Description: Output full build and execution (incl internal state) log.

• Default value: False

compiler.rtl.verilator_enable_debug

Verilator Enable Debug

• Type: bool

• Description: Enable/disable verbose internal state debug output.

• Default value: False

compiler.rtl.verilator_flags

Additional Verilator Arguments

• Type: str

• Description: Additional arguments feed to verilator.

• Default value: (Empty)

compiler.rtl.verilator_lint_warnings

Verilator Lint Warnings

• Type: bool

• Description: Enable/Disable detailed SV lint checker output.

• Default value: True

compiler.xilinx

FPGA (Xilinx) compiler preferences

compiler.xilinx.build_flags

Arguments

• Type: str

• Description: Kernel build C++ flags

• Default value: (Empty)

compiler.xilinx.decouple_array_interfaces

Decouple array memory interfaces

• Type: bool

• Description: If an array is both read and written, this option decouples its accesses, by creatin a memory interface for reading and one for writing. Note that this may hide potential Read-After-Write or Write-After-Read dependencies.

• Default value: False

compiler.xilinx.enable_debugging

Enable debugging for hardware kernels

• Type: bool

• Description: Injects debugging cores on the interfaces of the kernel, allowing fine-grained debugging of hardware runs at the cost of additional resources. This is always enabled for emulation runs.

• Default value: False

compiler.xilinx.frequency

Target frequency for Xilinx kernels

• Type: str

• Description: Target frequency, in MHz, when compiling kernels for Xilinx. Will not necessarily be achieved in practice. To enable multiple clocks, enter values in the format “clock_id:frequency”, with frequency being specified in MHz separated by an escaped bar, all enclosed in quotes. E.g. “0:250|1:500”.

• Default value: (Empty)

compiler.xilinx.host_flags

Host arguments

• Type: str

• Description: Extra host compiler argument flags

• Default value: -Wno-unknown-pragmas -Wno-unused-label

compiler.xilinx.mode

Compilation mode

• Type: str

• Description: Target of FPGA kernel build (simulation/software_emulation/hardware_emulation/hardware)

• Default value: simulation

compiler.xilinx.path

Vitis installation override

• Type: str

• Description: Path to specific Vitis/SDx/SDAccel installation to use instead of just searching PATH and environment variables.

• Default value: (Empty)

compiler.xilinx.platform

Target platform for Xilinx

• Type: str

• Description: Platform name of Vitis/SDx/SDAccel target.

• Default value: xilinx_u250_xdma_201830_2

compiler.xilinx.synthesis_flags

Synthesis arguments

• Type: str

• Description: High-level synthesis C++ flags

• Default value: -std=c++14

experimental

Experimental features

experimental.validate_undefs

Undefined Symbol Check

• Type: bool

• Description: Check for undefined symbols in memlets during SDFG validation.

• Default value: False

frontend

Python frontend preferences

frontend.avoid_wcr

Avoid using WCR for augmented assignments when possible

• Type: bool

• Description: Perform a map-symbol-dependency check on the write-subsets of augmented assignments that appear inside Maps to avoid using WCR when possible. This feature works correctly only when there is a single augmented assignment for each data dimension inside a Map.

• Default value: False

frontend.cache_size

Program cache size

• Type: int

• Description: The number of compiled programs to cache (based on argument types, closure constants, and closure array types) to avoid reparsing/compiling when calling a @dace.program or method.

• Default value: 32

frontend.check_args

Check arguments on SDFG call

• Type: bool

• Description: Perform an early type check on arguments passed to an SDFG when called directly (from SDFG.__call__). Another type check is performed when calling compiled SDFGs.

• Default value: False

frontend.dont_fuse_callbacks

Do not fuse callbacks

• Type: bool

• Description: Stricter mode of operation where callbacks into Python don’t participate in state fusion transformations.

• Default value: False

frontend.implicit_recursion_depth

Auto-parsing recursion depth

• Type: int

• Description: The maximum call-stack depth allowed when automatically parsing called dace functions or methods.

• Default value: 64

frontend.preprocessing_passes

Number of preprocessing passes on Python code

• Type: int

• Description: Number of times to run the Python preprocessing passes (e.g., constant folding) on the input code. Set to zero to disable preprocessing optimizations, set to -1 to run until the code has not changed.

• Default value: 5

frontend.raise_nested_parsing_errors

Raise nested parsing errors

• Type: bool

• Description: Raise all errors out of nested function parsing contexts instead of trying to create a callback implicitly.

• Default value: False

frontend.typed_callbacks_only

Only allow typed callbacks

• Type: bool

• Description: Stricter mode of operation where callbacks into Python must have explicit return value types in order to compile.

• Default value: False

frontend.unroll_threshold

Automatic unroll loop size threshold

• Type: int

• Description: Threshold for automatic loop unrolling of any generator (e.g., including range) with a compile-time size. A value of -1 (default) means not to unroll any loop automatically, a value of 0 means unrolling every loop, and a value above zero sets a size threshold beyond which a constant-sized loop will not be automatically unrolled.

• Default value: -1

frontend.verbose_errors

Show preprocessed AST on parsing errors

• Type: bool

• Description: Prints out the preprocessed unparsed AST in case of a parsing error.

• Default value: False

instrumentation

Instrumentation preferences

instrumentation.print_fpga_runtime

Print FPGA runtime

• Type: bool

• Description: Prints the runtime of instrumented FPGA kernel states to standard output.

• Default value: False

instrumentation.report_each_invocation

Save report for each invocation

• Type: bool

• Description: Save an instrumentation report file for each invocation of the SDFG, rather than one report that spans from SDFG initialization to finalization.

• Default value: True

instrumentation.papi

PAPI configuration

instrumentation.papi.default_counters

Default PAPI counters

• Type: str

• Description: Sets the default PAPI counter list, formatted as a Python list of strings.

• Default value: ['PAPI_TOT_INS', 'PAPI_TOT_CYC', 'PAPI_L2_TCM', 'PAPI_L3_TCM']

instrumentation.papi.overhead_compensation

Compensate Overhead

• Type: bool

• Description: Subtracts the minimum measured overhead from every measurement.

• Default value: True

instrumentation.papi.vectorization_analysis

Enable vectorization check

• Type: bool

• Description: Enables analysis of gcc vectorization information. Only gcc/g++ is supported.

• Default value: False

library

Settings for handling the use of DaCe libraries.

library.blas

Built-in BLAS DaCe library.

library.blas.default_implementation

Default implementation

• Type: str

• Description: Default implementation for BLAS library nodes.

• Default value: pure

library.blas.override

Force configured implementation

• Type: bool

• Description: Force the default implementation, even if an implementation has been explicitly set on a node.

• Default value: False

library.blas.fpga

FPGA-specific BLAS options.

library.blas.fpga.default_stream_depth

Default FPGA stream depth

• Type: int

• Description: Default FPGA stream depth used in the BLAS library nodes and the corresponding streaming transformations

• Default value: 32

library.lapack

Built-in LAPACK DaCe library.

library.lapack.default_implementation

Default implementation

• Type: str

• Description: Default implementation for LAPACK library nodes.

• Default value: OpenBLAS

library.lapack.override

Force configured implementation

• Type: bool

• Description: Force the default implementation, even if an implementation has been explicitly set on a node.

• Default value: False

library.linalg

Built-in NumPy linalg DaCe library.

library.linalg.default_implementation

Default implementation

• Type: str

• Description: Default implementation for linalg library nodes.

• Default value: OpenBLAS

library.linalg.override

Force configured implementation

• Type: bool

• Description: Force the default implementation, even if an implementation has been explicitly set on a node.

• Default value: False

library.pblas

Built-in PBLAS DaCe library.

library.pblas.default_implementation

Default implementation

• Type: str

• Description: Default implementation PBLAS library nodes.

• Default value: MKLMPICH

library.pblas.override

Force configured implementation

• Type: bool

• Description: Force the default implementation, even if an implementation has been explicitly set on a node.

• Default value: False

optimizer

Preferences of the SDFG Optimizer

optimizer.automatic_simplification

Automatic SDFG simplification

• Type: bool

• Description: Automatically performs SDFG simplification on programs.

• Default value: True

optimizer.autooptimize

Run auto-optimization heuristics

• Type: bool

• Description: Automatically runs the set of optimizing transformation heuristics on any program called via the Python frontend.

• Default value: False

optimizer.autospecialize

Auto-specialize symbols

• Type: bool

• Description: Automatically specialize every SDFG to the symbol values at call-time. Requires all symbols to be set.

• Default value: False

optimizer.autotile_partial_parallelism

Prefer partial parallelism over write-conflict tiling

• Type: bool

• Description: If true, sets the auto-optimizer to prefer extracting map parallel dimensions over tiling for atomic write-conflict resolution edges. This may be slower in case of small parallel dimensions vs. conflicted dimensions. This preference only applies to symbolic ranges or ranges over the autotile_size parameter.

• Default value: True

optimizer.autotile_size

Default tile size in auto-optimization

• Type: int

• Description: Sets the default tile size for the optimization heuristics.

• Default value: 128

optimizer.detect_control_flow

Detect control flow from state transitions

• Type: bool

• Description: Attempts to infer control flow constructs “if”, “for” and “while” from state transitions, allowing code generators to generate appropriate code.

• Default value: True

optimizer.match_exception

Treat exceptions in “can_be_applied” as errors

• Type: bool

• Description: When an exception is raised in a transformation “can_be_applied” function, if True the exception is raised further. Otherwise the exception is printed as a warning.

• Default value: False

optimizer.save_intermediate

Save intermediate SDFGs

• Type: bool

• Description: Save SDFG files after every transformation.

• Default value: False

optimizer.symbolic_positive

Treat all symbolic expressions as positive

• Type: bool

• Description: Every expression in which a symbolic value appears is treated as strictly positive. This is necessary for certain Range evaluations using Subgraph Fusion.

• Default value: True

optimizer.visualize_sdfv

Visualize SDFG

• Type: bool

• Description: Open an SDFG in browser every transformation.

• Default value: False

testing

Unit testing settings

testing.deserialize_exception

Treat exceptions in deserialization as errors

• Type: bool

• Description: When an exception is raised in a deserialization process (e.g., due to missing library node), by default a warning is issued. If this setting is True, the exception will be raised as-is.

• Default value: False

testing.serialization

Test Serialization on validation

• Type: bool

• Description: Before generating code, verify that a serialization/deserialization loop generates the same SDFG.

• Default value: False

testing.serialize_all_fields

Serialize all unmodified fields in SDFG files

• Type: bool

• Description: If False (default), saving an SDFG keeps only the modified non-default properties. If True, saves all fields.

• Default value: False