Hi,How can we get the matrix multiply counter (total Instructions & cycles) from the streamline tool?
I want to approximate the improvement gained using MMUL unit over FMA instructions
You can measure GPU activity and arithmetic unit utilization in the shader core, and compare before and after.
There is no specific counter for MMUL instructions.
Thanks for reply. Is the MMUL seperate unit or part of CVT/FMA
For shader-based work, e.g. using VK_KHR_cooperative_matrix, it will use the FMA/CVT/SFU pipes in the shader core.For data graph based work, using VK_ARM_data_graph, there is also dedicated neural accelerator which has its own set of counters. Whether a subgraph ends up on shader core or the neural accelerator depends on data type and operator. Neural accelerator counters are documented here:arm-software.github.io/.../mali_g2_counter_reference.html