Cog
Related Doc:
package parameters
Enable merging of the two kernels involved in the sequence blockReduceSum(convolve(...BackProjectFrame...))
Enable virtual field registers to share OpenCl buffers
The InOrderSharedLatchAllocator before Cog 4.3.6 mishandled some models.
The InOrderSharedLatchAllocator before Cog 4.3.6 mishandled some models. Alert the user if they were effected.
The Reshape operator before Cog 4.3 had a different behavior for some uses.
The Reshape operator before Cog 4.3 had a different behavior for some uses. Alert the user if they were effected.
Enable warning about user CPU Operators having apply() invoked multiple times- bad if they have mutable state.
Enables the cpu kernel Akka actors to share threads, rather than use a custom dispatcher that gives each actor its own thread.
Enables the cpu kernel Akka actors to share threads, rather than use a custom dispatcher that gives each actor its own thread. The system has not been extensively tested with this set to "true" and so may deadlock.
Release version, is not automatically discernible from the release jar manifest
Set this true for additional compiler information.
Set this true for additional compiler information. This deletes any cached programs, forcing a recompilation that will provide the information. This should always be set false except when debugging.
Enable multi-plane/workgroup kernel optimization for frame projection followed by block reduce
Enable experimental merger optimizations (e.g.
Enable experimental merger optimizations (e.g. forward-merging of kernels that use local memory)
Maximum number of probed fields that the graphical layout pane will attempt to display.
Allow the kernels to execute in a different order from which they were submitted to the GPU command queues.
Allow the kernels to execute in a different order from which they were submitted to the GPU command queues. This will result in bigger model GPU memory usage with sadly no performance boost on current OpenCL platforms.
Filter string for pruning down list of available OpenCL platforms.
Enable printing of OpenCL "binary" code.
Enable printing of OpenCL "binary" code. For Nvidia, the ptx assembly.
Print out optimized circuit for ComputeGraph.
Enable printing of OpenCL source code during compilation
Enable profiling.
Number of steps between writing profile statistics to the console.
Enable merging of the two kernels involved in the sequence blockReduceSum(convolve(...ProjectFrame...))
return the release version as a String.
return the release version as a String. When this is invoked out of the libcog jar, the getImplementationVersion will return the value from the file ./META-INF/MANIFEST.MF as put there by sbt, for example. When invoked from the development environment, this approach returns null, so we supply a value whose major and minor release number are correct.
The release version as a string
Looks up and sets a boolean parameter based on a JVM arg, e.g.
Looks up and sets a boolean parameter based on a JVM arg, e.g. -Dcog.verboseOptimizer
Looks up and sets an integer parameter based on a JVM arg, e.g.
Looks up and sets an integer parameter based on a JVM arg, e.g. -Dcog.profileSize=100
Looks up and sets a string parameter based on a JVM arg, e.g.
Looks up and sets a string parameter based on a JVM arg, e.g. -Dcog.platform="amd"
Enables the core platform Akka actors to share threads, rather than use a custom dispatcher that gives each actor its own thread.
Enables the core platform Akka actors to share threads, rather than use a custom dispatcher that gives each actor its own thread. The system has not been extensively tested with this set to "true" and so may deadlock.
Make the code generator chatty.
Display debugger steps.
Print out details of merging GPU kernels.
Print out details of OpenCL device allocation.
Print out details of OpenCL platform.
Print OpenCL resources allocated (device kernels, CPU kernels, buffers).
Display optimizer results.
Chatty version of User GPU Expressions to ease debugging.
Chatty version of User GPU Operators to ease debugging.
Controls compilation process, allowing various levels of verbosity for insight into the process. No module should print out information without consulting this. The user may modify the values of the fields here within their program. Alternatively, the user can set a JVM command line arg to modify the behavior of an already compiled program.