Running GPU Analysis from Command Line

You can use the amplxe-cl command tool to profile applications that do computations on an Intel® HD Graphics device.

Analyzing GPU activity helps you identify whether:

your workload is CPU- or GPU-bound.
Intel integrated graphics are used at their full potential and the headroom for improvement, if any.

VTune Amplifier proposes the following workflow for GPU analysis:

Platform analysis: Start with CPU/GPU Concurrency analysis to explore code execution on the various CPU and GPU cores in your system, correlate CPU and GPU activity and identify whether your application is GPU or CPU bound.
GPU analysis: Monitor overall GPU activity (graphics, media, and compute), collect Intel Integrated Graphics hardware metrics, and then view this data correlated with CPU processes and threads.

Platform Analysis Syntax

To perform CPU/GPU Concurrency analysis from command line, use the collect action and specify the cpugpu-concurrency predefined analysis type:

amplxe-cl -collect cpugpu-concurrency -app-working-dir <dir> -- <target_app> [target_app_options]

where:

dir is the directory where the application runs
target_app is the application or process (for Android*) to run
target_app_options are application arguments

GPU Analysis Syntax

To perform GPU analysis from command line, use the collect action and configure the level of GPU analysis with the knob option:

amplxe-cl -collect <analysis_type> -knob <knob_name>=<knob_value> -app-working-dir <dir> -- <target_app> [target_app_options]

where:

analysis_type is the specified type of analysis
knob_name is a name of the configuration option
knob_value is a string value for this configuration option
dir is the directory where the application runs
target_app is the application or process (for Android) to run
target_app_options are application arguments

For remote Linux and Android systems, make sure to use the --target-system option to specify your target platform. For example, to collect GPU usage when running com.example.android.basicglsurfaceview package, enter:

amplxe-cl -collect hotspots -knob enable-gpu-usage=true --target-system=android -- com.example.android.basicglsurfaceview

Configuring GPU Analysis from Command Line

Use This Knob	To Do This
enable-gpu-usage=true \| false	Analyze frame rate and usage of Processor Graphics engines.
gpu-counters-mode=none \|overview \| global-local-accesses \| compute-extended	Analyze performance data from Processor Graphics based on the predefined groups of GPU hardware metrics. `overview` - track general GPU memory accesses such as Memory Read/Write Bandwidth, GPU L3 Misses, Sampler Busy, Sampler Is Bottleneck, and GPU Memory Texture Read Bandwidth. These metrics can be useful for both graphics and compute-intensive applications. `global-local-accesses` - include metrics that distinguish accessing different types of data on a GPU: Untyped Memory Read/Write Bandwidth, Typed Memory Read/Write Transactions, SLM Read/Write Bandwidth, Render/GPGPU Command Streamer Loaded, and GPU EU Array Usage. This metrics are useful for compute-intensive workloads on the GPU. `compute-extended` - analyze GPU activity on the Intel processor code name Broadwell. This metrics set is disabled for other systems. This option is available only for supported platforms with the Intel® graphics driver installed.
gpu-sampling-interval=<value in us>	Set the interval between GPU samples between 10 and 1000 microseconds. Default is 1000us. An interval of less than 100us is not recommended.
enable-gpu-runtimes=true \| false	Capture the execution time of OpenCL™ kernels and Intel® Media SDK programs on a GPU, identify performance-critical GPU computing tasks, and analyze the performance per GPU hardware metrics. Note OpenCL kernels analysis is currently supported for Windows and Linux target systems with Intel HD graphics. Intel Media SDK program analysis is supported for Linux targets only and should be started with root privileges.

Examples:

Start amplxe-cl as root and launch a Basic Hotspots analysis for an Intel Media SDK application:

$ amplxe-cl  --collect hotspots --knob enable-gpu-usage=true --knob enable-gpu-runtimes=true -r <result_dir> -- <application>

Perform a Basic Hotspots analysis, enabling the enable-gpu-usage knob to analyze GPU usage of a processor graphics engine, using the Overview gpu-counters-mode counter set, which is available only on a supported platform with an Intel® graphics driver installed. Enable tracing of OpenCL kernels execution with the enable-gpu-runtimes option.

For example, to collect GPU usage as well as GPU hardware metrics and trace OpenCL kernels on the BitonicSort application (-g is the option of the application), enter:

$ amplxe-cl -collect hotspots -knob gpu-counters-mode=overview -knob enable-gpu-usage=true -knob enable-gpu-runtimes=true -- BitonicSort -g

The summary information sent to stdout includes basic details on both the CPU and GPU.

Collection and Platform Info
------------------------------------------------------------------------
Parameter                 r010hs

------------------------  ----------------------------------------------
Application Command Line  BitonicSort "-g"
Operating System          Ubuntu 11.04 
Computer Name             <…>

Result Size               1367805


CPU
---
Parameter          r010hs
-----------------  -----------------------------------------------------
Name               3rd generation Intel® Core™ Processor family
Frequency          3500000000
Logical CPU Count  8

GPU
---
Parameter                      r010hs
-----------------------------  -----------------------------------------
Name                           Intel® HD Graphics 4000
Vendor                         Intel Corporation
Driver                         <…>
Stepping                       9
EU Count                       16
Max EU Thread Count            8
Max Core Frequency             1150000000
Graphics Performance Analysis  enabled

GPU OpenCL Info
---------------
Parameter            r010hs
-------------------  -------------
Version              OpenCL C 1.2
Max Compute Units    16
Max Work Group Size  512
Local Memory         65536

Summary
-------
Elapsed Time:  1.792
CPU Time:      2.282

Viewing GPU Analysis Data Collected from Command Line

When your analysis is completed, open the collected result in the VTune Amplifier GUI:

For CPU/GPU Concurrency data, use the Platform View viewpoint and focus on the Platform window.
For GPU analysis, use your analysis default viewpoint and focus on the Graphics window to analyze GPU activity. For OpenCL applications running on Windows targets, you can also use the gpu-computing-tasks command line report to view computing tasks or their instances annotated with total and average time, instance count and average values of GPU hardware metrics (if applicable).

Examples:

View the collected data per OpenCL kernels submitted and executed on Intel Processor Graphics using the gpu-computing-tasks report type:
```
$ amplxe-cl -R gpu-computing-tasks -r r010hs
```

Filter or group the collected data by OpenCL kernel instances:

$ amplxe-cl -R gpu-computing-tasks -r r010hs -group-by computing-instance

Running GPU Analysis from Command Line

Platform Analysis Syntax

GPU Analysis Syntax

Configuring GPU Analysis from Command Line

Note

Viewing GPU Analysis Data Collected from Command Line

See Also