Open topic with navigation

Overview

Source-level experiments collect information per line of SASS assembly, and for kernels compiled with line info, the information can be correlated to the original source and PTX as well. Since kernel-level experiments collect information summed or averaged over the lifetime of the kernel, they are helpful for determining if a kernel is making inefficient use of the GPU, but not for isolating which parts of the kernel are causing the inefficiency. Due to the difficulty of collecting fine-grained timing information without altering timing, source-level experiments only collect behavioral data that is invariant to execution speed. Use source-level experiments to search for inefficient behavioral patterns.

Integration with CUDA Source View

Source-level experiments add columns to the Source View. Use the Column Chooser to display the columns of interest.

CUDA Launches View Correlation

Source-level experiments add correlation data to launches in the CUDA Launches View, just as kernel-level experiments do. Source-level experiments present a single data table, where each row represents a line of SASS assembly for which the experiment collected data. Use the Column Chooser to display the columns of interest. Click the arrow in the Line # columns or right-click a row to jump to that line in the CUDA Source View or the Visual Studio editor.

Results of the Memory Transactions source-level experiment in the correlation pane for a kernel launch.

List of CUDA source-level experiments

Instruction Count

Divergent Branch

Memory Transactions

Open topic with navigation