Skip to content

Environment Variables

Braden Stefanuk edited this page Sep 30, 2024 · 6 revisions

This file is deprecated as of 2024-09-30 and will be removed for ROCm 6.4. New documentation is under active development.

The Tensile client recognizes the following environment variables that enable test or debugging features:

Env Variable Usage
TENSILE_DB Enables debugging features based on the value set.
TENSILE_DB is a bit field, so options can be set individually or combined. To enable all debug output, set TENSILE_DB=0xFFFF. Individual options are listed below:

0x2 or 0x4 - Prints extra information about the solution selection process. Indicates if a kernel was an exact match, or the sequence of kernels considered for a closest match.
0x8 - Prints extra information about the hardware selection process.
0x10 - Prints debug-level information about predicate evaluations.
0x20 - Prints a list of loaded or missing code object libraries.
0x40 - Prints kernel launch arguments, including the kernel name, work group size and count, and all arguments passed.
0x80 - Prints size of allocated tensors.
0x100 - Prints debug information about convolution reference calculations.
0x200 - Prints more detailed information about convolution reference calculations.
0x400 - Currently unused.
0x800 - Currently unused.
0x1000 - Prints information about the loading of embedded, YAML, or MessagePack libraries.
0x2000 - Currently unused.
0x4000 - Prints solution lookup efficiency.
0x8000 - Prints the name of selected kernels.
0x80000 - Prints the name of selected kernels and number of the common kernel parameters (e.g. Matrix Instruction, MacroTile, ThreadTile, DepthU, etc) .
TENSILE_DB2 Enables extended debugging features based on the value set.
If TENSILE_DB2 is set to 1, Tensile will skip launching kernels for debug purposes. All other steps such as kernel selected, data allocation, and initialization are still performed.
TENSILE_NAIVE_SEARCH Force a naive search for matching kernels instead of the standard optimized search.
TENSILE_TAM_SELECTION_ENABLE Enables tile aware solution selection.
TENSILE_SOLUTION_INDEX Prints the index of the selected solution.
TENSILE_METRIC Override the default distance metrix for solution selection. Can be set to "Euclidean", "JSD", "Manhattan", "Ratio", or "Random".
TENSILE_EXPERIMENTAL_SELECTION Experimental kernel selection for GEMM.

Options:
0 or unset - default kernel selection is used.
1 - grid experimental kernel selection is used.
2 - decision trees experimental kernel selection is used.

See Experimental Kernel Selection page for details.
Clone this wiki locally