These environment variables can be used to control OpenCL behavior and provide visibility for debugging.
When OpenCL C kernels are compiled for DSPs, the result is a binary .out file in the /tmp sub-directory. They are then subsequently available for download to the DSPs for running. The process of compiling generates several intermediate files for each source file. OpenCL typically removes these temporary files. However, it can sometimes be useful to inspect these files. This environment variable can be set to instruct the runtime to leave the temporary files in /tmp. Inspecting the assembly file associated with the out file, can be useful to see how well your code was optimized.
Setting this environment variable modifies the execution of OpenCL applications to enable debug of the OpenCL C kernels. If your application uses on-line compilation (i.e. it compiles from a string rather than a binary), then that on-line compilation asserts the debug flag to the compiler. (If the application is using off-line compilation, i.e. creating a program from binary, then you would need to pass -g as an option to the off-line compiler clocl. The OpenCL runtime pauses your application before dispatch of all kernels. While paused the runtime indicates to the user that a kernel dispatch is pending.
Currently two debugging methods are supported by setting TI_OCL_DEBUG to either “gdb” or “ccs”. If set to “gdb”, the runtime provides the gdbc6x commands to connect to the DSP and setup appropriate breakpoints. If set to “ccs”, the runtime provides the Code Composer Studio (CCS) instructions to connect to the DSP and setup appropriate breakpoints. With either method, the runtime forces all kernels and work-groups within kernels to execute only on DSP core 0. Details can be found in Debug.
On-line compilation of kernels is a useful feature for portable OpenCL programs. All the detail required to compile kernels for devices is encapsulated in the OpenCL API calls. However, on-line compilation for the DSPs can be time-consuming. Setting this environment variable causes the OpenCL runtime to perform one on-line compilation of your kernels and cache the results in a database in /tmp. Running the application again without modification of the kernel source or the options used to compile it, results in the compilation step being bypassed, and the use of the cached kernel binary.
If OpenCL C kernels call standard C code, modifications to the standard C code are not seen by the OpenCL runtime and a cached result may be used when it is not appropriate. If calling standard C code, either disable this environment variable or clean the cache anytime the standard C code is modified.
Using this environment variable causes persistent data to accumulate in /tmp, and /tmp may grow to capacity causing run-time errors. If this occurs, remove cached objects in /tmp or increase the size of the /tmp partition. To explicitly remove the cache, execute the command: rm -f /tmp/opencl*.
The OpenCL compilation cache is automatically removed during a Linux reboot
Specify the compute units available to the OpenCL runtime as a comma separated list of compute unit indices starting at 0. The specified compute unit list must be consecutive, e.g.
"1,2,3,4"on K2H. If the environment variable is not specified, the runtime defaults to using all the compute units available on the device (or starting from OpenCL product version 1.1.13, all available compute units specified in
Example usage on AM572x:
-> TI_OCL_COMPUTE_UNIT_LIST="0" ./vecadd
-> TI_OCL_COMPUTE_UNIT_LIST="1" ./vecadd
-> TI_OCL_COMPUTE_UNIT_LIST="0, 1" ./vecadd
Prior to OpenCL product version 1.1.13, this environment variable is available only on AM572x.
By default, OpenCL kernel related code and global data is allocated out of DDR memory. If this environment variable is set, kernel related code and global data is allocated out of MSMC memory.
Rarely used K2x only feature, will be deprecated starting with OpenCL version 188.8.131.52.
Currently, OpenCL ARM CPU devices only support native kernels (see the OpenCL 1.1 spec for a description of native kernels). As a result, the ARM CPU is not, by default, treated as a COMPUTE DEVICE when doing an OpenCL platform query. If your application only uses the ARM CPU for native kernels, then this environment variable can be used to enable it as a COMPUTE DEVICE for OpenCL. Enqueueing NDRangeKernels or Tasks to the CPU is not supported, even when this environment variable is set.
Removed in OpenCL 1.1.17, no longer required.
The OpenCL runtime starts a new CPU thread for every OpenCL command queue defined in your application. These threads manage the OpenCL command queues and the communication between the CPU and the device to which the command queue is associated. If there are any OpenCL kernels actively running on the device, the thread assigned to monitor the communication with the device on behalf of those kernels consumes CPU resources, checking the status of those kernels. This environment variable can be used to provide a level of control on how much CPU resource is consumed. When TI_OCL_WORKER_SLEEP is unset, the OpenCL runtime uses more CPU capacity to ensure the fastest turnaround latency on kernel execution. When the TI_OCL_WORKER_SLEEP environment variable is set to a number of microseconds, it degrades the turnaround latency for a kernel execution to reduce the CPU capacity needed to monitor the kernel. If an application is not performance limited by CPU cycles or if the application enqueues many fine-grained kernels, then having the TI_OCL_WORKER_SLEEP environment variable unset is appropriate. In the opposite cases, when CPU cycles are limiting the performance of an application or if fewer, but longer running kernels are enqueued, then setting TI_OCL_WORKER_SLEEP to some number of microseconds is appropriate. The correct number of microseconds to use depends on the execution platform and the particular application. However, using a microseconds value in the range from 80 to 150 is a reasonable starting point.
The C66x DSP is double precision floating point capable and all the optional features in the OpenCL specification for double precision floating point are supported in this OpenCL implementation, except for the requirement that double FP support include subnormal behavior or graceful underflow. The 64-bit floating point hardware on the C66x DSP does not support subnormal behavior. It supports flush to zero behavior. To support subnormal behavior for doubles would require software emulation that would entail a significant performance penalty versus the hardware capabilities of the C66x DSP. Therefore, by default the platform and devices supported in the TI OpenCL implementation do not report support for double floating point. That is, if the platform or device is queried for extensions, cl_khr_fp64 is not listed by default. Additionally the OpenCL C predefined macro cl_khr_fp64 is not be defined by default. When the TI_OCL_ENABLE_FP64 environment variable is set, the TI OpenCL implementation reports support for double floating point, i.e. cl_khr_fp64 is listed as an extension for the platform and the DSP device and cl_khr_fp64 is defined when compiling OpenCL C kernels. This environment variable controls whether the OpenCL implementation reports support for double. However, double, all double vector types and all built-in functions using doubles are supported and available without regard to the setting of this environment variable.
The OpenCL specification provides a well-defined mechanism for returning error codes from API functions. However, It is often the case that a generic error code is returned for differing reasons. When this environment variable is set, the OpenCL runtime may print more description error messages in addition to the defined return code error mechanism.
OpenCL provides a query to a device for the maximum number of work-items allowed in a work-group. The DSP device in TI’s implementation allows a large number of work-items per work-group. Other OpenCL implementations have much smaller max work-group size limit. When running code designed and optimized for other OpenCL implementations, this environment variable can be used to limit the max work-group size reported.
Specifies the hardware event type to profile. The two basic divisions, stall cycle events and memory events, are described in Profiling. If 1 is specified, OpenCL runtime will profile a stall cycle event. If 2 is specified, OpenCL runtime will profile one or two memory event(s). Otherwise, profiling is disabled.
Specifies the event number to profile. The exact value of this variable represents the offset from either AET_GEM_STALL_EVT_START or AET_GEM_MEM_EVT_START, depending on the event type. For a full event list, see Profiling.
Specifies the second memory event number to profile. Can be skipped.
Specified the threshold of stall cycles to count. Only stall events with stall cycles higher than this threshold are captured in the counter. Default value in OpenCL runtime is 1, i.e. all stall events are captured.