Using GUI - Using GUI - 57368

If data is collected using CLI, then use Import Session to import the session into GUI to analyze data in GUI. The following are the supported views to analyze GPU Profile:

Legend:

1. Target Application and Profile summary.
2. Hotspots Summary shows a glimpse of GPU Kernel summary about Kernel Launch count, total execution time(s) etc.
3. System Speed of Light : System Speed-of-Light summarizes some of the key metrics from various sections of profiling report.

Legend:

1. Profile Details
2. Target Details
3. System Details
4. GPU Device Details

Legend:

1. All supported views are segregated into three categories as follows:
   • Overview
     • System Speed of Light: System Speed-of-Light summarizes some of the key metrics from various sections of profiling report.
   • Compute
     • Command Processor: Command Processor handles interacting with AMDGPU kernel driver- Linux Kernel on CPU and user space HSA clients when commands are submitted to HSA queues.
     • Shader Processor Input: Shader Processor passes dispatches from command processor to workgroup manager onto the CU.
     • Wavefront Launch: The wavefront launch stats panel gives general information about the kernel launch.
     • Compute Units Instruction Mix: The instruction mix panel shows a breakdown of the various types of instructions executed by the user’s kernel, and which pipelines on the CU they were executed on.
     • Compute Units Compute Pipeline: This section reports the number of floating-point and integer operations executed on the VALU and MFMA units in various precisions.
   • Cache
     • Memory Chart: Metrices related to memory including caches and type of Read and Write.
     • Local Data Share: The LDS statistics panel gives a more detailed view of the H/W
     • Instruction Cache: Detail on the hit/miss statistics of the L1 Instruction (L1I) cache.
     • Scalar L1D Cache: The Scalar L1 Data cache (sL1D) can cache data accessed from scalar load instructions.
     • Texture Address and Texture Data: The Texture Address shares in-depth understanding of memory instructions, write and atomic data from CU and passes this info to data processing unit.
     • Vector L1D Cache: The vector L1 data (vL1D) cache is local to each CU on the accelerator and handles vector memory operations issued by a wavefront.
     • L2 Cache: This panel shares performance, accesses, misses and delays in accessing L2 cache.
     • L2 Cache Per Channel: This panel shares aggregated L2 cache performance.
2. The filters pane lets you filter the profile data by providing the following options.
   • Kernel Filter: Evaluate GPU Metrics per view for selected kernel(s)
   • GPU ID Filter: Filter out kernels launched on a specific GPU device
   • Normalization Filter: Evaluate GPU Metrics in all views according to selected normalization. Below is a list of supported normalization:
     • Per Wave: The total value of the measured counter or metric that occurred per kernel invocation divided by the total number of wavefronts launched in the kernel.
     • Per Cycle: The total value of the measured counter or metric that occurred per kernel invocation divided by the kernel cycles, that is, the total number of cycles the kernel executed as measured by the command processor.
     • Per Second: The total value of the measured counter or metric that occurred per kernel invocation divided by the kernel time, that is, the total runtime of the kernel in seconds, as measured by the command processor.
     • Per Kernel: The total value of the measured counter or metric that occurred per kernel invocation.
       Note: Per Wave is default normalization.
   • All Kernel Info: Resets all filters.
3. This section lists all launched GPU Kernels in descending order of total execution time with total launch count, Min, Max and Avg time taken by each kernel. This section also supports sorting data on all columns.
4. Any selected kernel(s) will be displayed in this Label.
5. Select Appropriate view which needs to be analyzed from drop down.
6. Subsequent Metrics for selected views are listed in this section.

Use these views to analyze how efficiently GPUs are used by the application. In other words, how much time specific GPU kernel took for executions with subsequent H/W counters evaluation for that kernel.