Skip to content

Latest commit

 

History

History
945 lines (735 loc) · 44.3 KB

README.md

File metadata and controls

945 lines (735 loc) · 44.3 KB
Raw

Using the IBM Z Deep Learning Compiler Container Images

Table of contents


Overview

The IBM Z Deep Learning Compiler uses ONNX-MLIR to compile .onnx deep learning AI models into shared libaries. The shared libaries can then be integrated into C, C++, Java, or Python applications.

The compiled models take advantage of IBM zSystems technologies including SIMD on IBM z13 and later and the Integrated Accelerator for AI available on IBM z16 without changes to the original model.

ONNX is an open format for representing AI models. It is open source and vendor neutral. Some AI frameworks directly support exporting to .onnx format. For other frameworks, open source converters are readily available. ONNX Support Tools has links to steps and converters for many popular AI frameworks.

See Verfied ONNX Model Zoo models for the list of models from the ONNX Model Zoo that have been built and verified with the IBM Z Deep Learning Compiler.

These are the general end-to-end steps to use IBM zDLC:

  1. Create, convert, or download an ONNX model.
  2. Download the zdlc image from IBM Z and LinuxOne Container Registry.
  3. Use the image to compile a shared library of the model for your desired language.
  4. Import the compiled model into your application.
  5. Run your application.

Download the IBM Z Deep Learning Compiler container image

Downloading the IBM Z Deep Learning Compiler container image requires credentials for the icr.io registry. Information on obtaining the credentials is located at IBM Z and LinuxONE Container Registry.

Determine the desired version of the zdlc image to download from the IBM Z and LinuxOne Container Registry.

Set ZDLC_IMAGE based on the desired IBM zDLC version:

ZDLC_IMAGE=icr.io/ibmz/zdlc:4.3.0

Pull the image as shown:

docker pull ${ZDLC_IMAGE}
Variable Description
ZDLC_IMAGE=icr.io/ibmz/zdlc:[version] Set [version] based on the desired version in IBM Z and LinuxONE Container Registry.

Download the IBM Z Deep Learning Compiler examples


The code examples are located in this GitHub repository. The easiest way to follow the examples is to clone the example code repository to your local system.

To clone the example repository to a new subdirectory called zDLC:

git clone https://github.com/IBM/zDLC

Set ZDLC_DIR to where you cloned this example repository:

ZDLC_DIR=$(pwd)/zDLC

This assumes you cloned the repository to the current working directory using the git clone command above. If you cloned the repository to another location, make sure to set this variable accordingly.

Variable Description
ZDLC_DIR=$(pwd)/zDLC $(pwd) resolves to the current working directory.
zDLC is the name of this repository. The zDLC directory is created automatically by git clone.

Environment variables

Set the environment variables for use with the IBM Z Deep Learning Compiler container image. The environment variables will simplify the container commands throughout the rest of this document. See the description in the table below for additional information.

NOTE: ZDLC_IMAGE and ZDLC_DIR are based on your local system. To set these environment variables, see:

GCC_IMAGE_ID=icr.io/ibmz/gcc:12
JDK_IMAGE_ID=icr.io/ibmz/openjdk:11
ZDLC_CODE_DIR=${ZDLC_DIR}/code
ZDLC_LIB_DIR=${ZDLC_DIR}/lib
ZDLC_BUILD_DIR=${ZDLC_DIR}/build
ZDLC_MODEL_DIR=${ZDLC_DIR}/models
ZDLC_MODEL_NAME=mnist-12
if [ -z ${ZDLC_IMAGE} ]; then echo ERROR: ZDLC_IMAGE must be set first; fi
if [ -z ${ZDLC_DIR} ] || [ ! -d ${ZDLC_DIR} ]; then echo ERROR: ZDLC_DIR must be set to an existing zDLC example directory first; fi
Variable Description
GCC_IMAGE_ID=icr.io/ibmz/gcc:12 Used in:
• Building C++ programs to call the model
• Compiling models to utilize the IBM Z Integrated Accelerator for AI
JDK_IMAGE_ID=icr.io/ibmz/openjdk:11 Used in:
• Building Java programs to call the model
ZDLC_CODE_DIR=${ZDLC_DIR}/code Used in:
• Building C++ programs to call the model
• Building Java programs to call the model
• Running the Python example
• Compiling models to utilize the IBM Z Integrated Accelerator for AI
ZDLC_LIB_DIR=${ZDLC_DIR}/lib Used in:
• Running the Python example
ZDLC_BUILD_DIR=${ZDLC_DIR}/build Used in:
• Building C++ programs to call the model
• Building Java programs to call the model
ZDLC_MODEL_DIR=${ZDLC_DIR}/models Used in:
• Building the code samples
• Building a model .so using the IBM Z Deep Learning Compiler
• Building C++ programs to call the model
• Building a model .jar file using the IBM zDLC
• Building Java programs to call the model
• Running the Python example
• Compiling models to utilize the IBM Z Integrated Accelerator for AI
• Obtaining IBM Z Deep Learning Compiler debug instrumentation
• Device placement of targets at compile time
ZDLC_MODEL_NAME=mnist-12 Used in:
• Building the code samples
• Building a model .so using the IBM Z Deep Learning Compiler
• Building C++ programs to call the model
• Compiling models to utilize the IBM Z Integrated Accelerator for AI
• Building a model .jar file using the IBM zDLC
• Building Java programs to call the model
• Running the Python example
• Obtaining IBM Z Deep Learning Compiler debug instrumentation
• Device placement of targets at compile time
if ... fi Simple tests to confirm ZDLC_IMAGE and ZDLC_DIR were set. If they were not set, set them and then reset the other variables.

IBM Z Deep Learning Compiler command line interface help

Running the IBM Z Deep Learning Compiler container image with no parameters shows the complete help for the IBM Z Deep Learning Compiler.

docker run --rm ${ZDLC_IMAGE}

Note the command line entry point for the IBM Z Deep Learning Compiler is the zdlc command. The IBM Z Deep Learning Compiler is invoked by running the zdlc image with the docker run command.

Command
and
Parameters		 Description
docker run Run the container image.
--rm Delete the container after running the command.

The help for the IBM Z Deep Learning Compiler can also be displayed by adding the --help option to the command line.


Building the code samples

If you are using the default mnist example model from the ONNX Model Zoo, you can download it using:

wget --directory-prefix $ZDLC_MODEL_DIR https://github.com/onnx/models/raw/main/validated/vision/classification/mnist/model/$ZDLC_MODEL_NAME.onnx

or see Obtaining the models to download the other models from the model zoo. The examples use $ZDLC_MODEL_DIR as the directory and $ZDLC_MODEL_NAME specifies the model name (without the .onnx) in that directory.


Building a model .so using the IBM Z Deep Learning Compiler

Use the --EmitLib option to build a .so shared library of the model specified by ZDLC_MODEL_NAME in Environment variables:

docker run --rm -v ${ZDLC_MODEL_DIR}:/workdir:z ${ZDLC_IMAGE} --EmitLib --O3 --mcpu=z14 --mtriple=s390x-ibm-loz ${ZDLC_MODEL_NAME}.onnx
Command
and
Parameters		 Description
ZDLC_MODEL_NAME Name of the model to compile without ending suffix.
docker run Run the container image.
--rm Delete the container after running the command.
-v ${ZDLC_MODEL_DIR}:/workdir:z The host bind mount points to the directory with the model ONNX file. :z is required to share the volume if SELinux is installed.
--EmitLib Build the .so shared library of the model.
--O3 Optimize to the highest level.
--mcpu=z14 The minimum CPU architecture (for generated code instructions).
--mtriple=s390x-ibm-loz The target architecture for generated code.
${ZDLC_MODEL_NAME}.onnx Builds the .so shared library from the specified ONNX file.

The built .so shared library is written to the host bind mount location.

The ONNX models for the examples can be found in the ONNX Model Zoo.


Building C++ programs to call the model

The example program is written in the C++ programming language and compiled with the g++ compiler. The example program calls the IBM Z Deep Learning Compiler APIs built into the .so shared library. The source code for the example program is at C++ example.

Some setup steps are required before building the programs to call the model. The ONNX-MLIR Runtime API files first need to be copied from the container image. Run these commands from the command line to copy files.

mkdir -p ${ZDLC_BUILD_DIR}
docker run --rm -v ${ZDLC_BUILD_DIR}:/files:z --entrypoint '/usr/bin/bash' ${ZDLC_IMAGE} -c "cp -r /usr/local/{include,lib} /files"
Command
and
Parameters		 Description
docker run Run the container image.
--rm Delete the container after running the command.
-v ${ZDLC_BUILD_DIR}:/files:z The host bind mount points to the directory to copy the build files from IBM. :z is required to share the volume if SELinux is installed.
cp Run the copy command to copy the build files from IBM into the host bind mount.

Run this optional step to see the files that were copied.

ls -laR ${ZDLC_BUILD_DIR}

Next pull a Docker image with the g++ compiler tools installed.

docker pull ${GCC_IMAGE_ID}

The setup steps have been completed. Use the g++ image and the ONNX-MLIR C++ Runtime API files to build the program.

cp ${ZDLC_MODEL_DIR}/${ZDLC_MODEL_NAME}.so ${ZDLC_CODE_DIR}
docker run --rm -v ${ZDLC_CODE_DIR}:/code:z -v ${ZDLC_BUILD_DIR}:/build:z ${GCC_IMAGE_ID} g++ -std=c++11 -O3 -I /build/include /code/deep_learning_compiler_run_model_example.cpp -l:${ZDLC_MODEL_NAME}.so -L/code -Wl,-rpath='$ORIGIN' -o /code/deep_learning_compiler_run_model_example

The following table explains the command line:

Command
and
Parameters		 Description
docker run Run the container image.
--rm Delete the container after running the command.
-v ${ZDLC_CODE_DIR}:/code:z The /code host bind mount points to the directory with the calling program. :z is required to share the volume if SELinux is installed.
-v ${ZDLC_BUILD_DIR}:/build:z The /build host bind mount points to the directory containing the build files from IBM. :z is required to share the volume if SELinux is installed.

The following table explains the g++ command line:

Command
and
Parameters		 Description
g++ Run the g++ compiler from the container command line.
-std=c++11 -O3 g++ compiler options (See the man g++ help for additional information.).
-I /build/include This is the location of the include header files.
/code/deep_learning_compiler_run_model_example.cpp The example program to build.
-l:${ZDLC_MODEL_NAME}.so The model .so shared library that was previously built.
-L/code Tell the g++ linker where to find the model .so shared library.
-Wl,-rpath='$ORIGIN' (This is a very important parameter for correctly building the C++ example program.) The GNU loader (LD) uses the rpath to locate the model .so file when the program is run. (See the man ld.so help for additional information.)
-o /code/deep_learning_compiler_run_model_example Tell the g++ linker the name of the built program.

The program is now ready to be run from the command line. When run, the program will inference the model with randomly generated test data values.

docker run --rm -v ${ZDLC_CODE_DIR}:/code:z ${GCC_IMAGE_ID} /code/deep_learning_compiler_run_model_example

With this example, the program is linked to the built model and is run in the container. The expected program output is ten random float values (because the input was random) from the model.

Building a model .jar file using the IBM zDLC compiler

Use the --EmitJNI option to build a jar file of the model specified by ZDLC_MODEL_NAME in Environment variables.

docker run --rm -v ${ZDLC_MODEL_DIR}:/workdir:z ${ZDLC_IMAGE} --EmitJNI --O3 --mcpu=z14 --mtriple=s390x-ibm-loz ${ZDLC_MODEL_NAME}.onnx
Command
and
Parameters		 Description
ZDLC_MODEL_NAME Name of the model to compile without ending suffix.
docker run Run the container image.
--rm Delete the container after running the command.
-v ${ZDLC_MODEL_DIR}:/workdir:z The host bind mount points to the directory with the model ONNX file. :z is required to share the volume if SELinux is installed.
--EmitJNI Build the jar file of the model.
${ZDLC_MODEL_NAME}.onnx Builds the .jar shared library from the specified ONNX file.

The built jar file is written to the host bind mount location.


Building Java programs to call the model

The example program is written in the Java programming language and compiled with a Java JDK. The example program calls the ONNX-MLIR Java Runtime APIs through the JNI interfaces built in the model jar file. The source code for the example program is at Java example.

Some setup steps are required before building the programs to call the model. The ONNX-MLIR Runtime API files first need to be copied from the container image. Run these commands from the command line to copy files.

mkdir -p ${ZDLC_BUILD_DIR}
docker run --rm -v ${ZDLC_BUILD_DIR}:/files:z --entrypoint '/usr/bin/bash' ${ZDLC_IMAGE} -c "cp -r /usr/local/{include,lib} /files"
Command
and
Parameters		 Description
docker run Run the container image.
--rm Delete the container after running the command.
-v ${ZDLC_BUILD_DIR}:/files:z The host bind mount points to the directory to copy the build files from IBM. :z is required to share the volume if SELinux is installed.
cp Run the copy command to copy the build files from IBM into the host bind mount.

Run this optional step to see the files that were copied.

ls -laR ${ZDLC_BUILD_DIR}

Pull a Java JDK image to build and run the Java example:

docker pull ${JDK_IMAGE_ID}

Build the Java calling program using the javac command.

mkdir -p ${ZDLC_CODE_DIR}/class
docker run --rm -v ${ZDLC_CODE_DIR}:/code:z -v ${ZDLC_BUILD_DIR}:/build:z ${JDK_IMAGE_ID} javac -classpath /build/lib/javaruntime.jar -d /code/class /code/deep_learning_compiler_run_model_example.java
Command
and
Parameters		 Description
docker run Run the container image.
--rm Delete the container after running the command.
-v ${ZDLC_CODE_DIR}:/code:z The /code host bind mount points to the directory with the calling program. :z is required to share the volume if SELinux is installed.
-v ${ZDLC_BUILD_DIR}:/build:z The /build host bind mount points to the directory containing the build files from IBM. :z is required to share the volume if SELinux is installed.
javac Run the JDK Java compiler from the container command line.
-classpath /build/lib/javaruntime.jar Need to specify the path to the run-time jar from IBM.
-d /code/class The build class files are stored at ${ZDLC_CODE_DIR}/class.

The program is now ready to be run from the command line. When run, the program will inference the model with randomly generated test data values.

cp ${ZDLC_MODEL_DIR}/${ZDLC_MODEL_NAME}.jar ${ZDLC_CODE_DIR}
docker run --rm -v ${ZDLC_CODE_DIR}:/code:z ${JDK_IMAGE_ID} java -classpath /code/class:/code/${ZDLC_MODEL_NAME}.jar deep_learning_compiler_run_model_example

With this example, the Java classpath contains the paths for the host bind mounts when run within the container. The classpath needs to be adjusted if the Java program is run directly from the command line. The expected program output is a list of random float values (because the input was random) from the model.


Running the Python example

This example program is written in Python and runs using the Python runtime. The example program calls the ONNX-MLIR Runtime APIs by leveraging pybind and PyExecutionSession which is best described in sections Using PyRuntime and PyRuntime Module in the linked documentation.

If not already compiled, compile the model specified by ZDLC_MODEL_NAME in Environment variables to a .so shared library as described previously.

Next, copy the PyRuntime library out of the docker container using:

mkdir -p ${ZDLC_LIB_DIR}
docker run --rm -v ${ZDLC_LIB_DIR}:/files:z --entrypoint '/usr/bin/bash' ${ZDLC_IMAGE} -c "cp /usr/local/lib/PyRuntime* /files"
Command
and
Parameters		 Description
docker run Run the container image.
--rm Delete the container after running the command.
-v ${ZDLC_LIB_DIR}:/files:z The /files host bind mount points to the directory we want to contain the PyRuntime library. :z is required to share the volume if SELinux is installed.
--entrypoint '/usr/bin/bash' The user will enter the container with /usr/bin/bash as the starting process.
-c "cp" Tell the entrypoint bash process to copy the PyRuntime library outside of the container into the directory bind mounted at /files.

Run this optional step to see the files that were copied.

ls -laR ${ZDLC_LIB_DIR}

Two configuration approaches are described in onnx-mlir's Configuring and using PyRuntime, but we'll prefer the PYTHONPATH approach so we avoid creating symbolic links for this example.

Build the example Python image with the following command:

docker build -f ${ZDLC_DIR}/docker/Dockerfile.python -t zdlc-python-example .
Command
and
Parameters		 Description
docker build Build the container image.
-f docker/Dockerfile.python Use docker/Dockerfile.python as the Dockerfile for this container build.
-t zdlc-python-example Build the image with the image:tag specification of zdlc-python-example:latest.

Finally, run the Python client with the following command:

docker run --rm -v ${ZDLC_LIB_DIR}:/build/lib:z -v ${ZDLC_CODE_DIR}:/code:z -v ${ZDLC_MODEL_DIR}:/models:z --env PYTHONPATH=/build/lib zdlc-python-example:latest /code/deep_learning_compiler_run_model_python.py /models/${ZDLC_MODEL_NAME}.so
Command
and
Parameters		 Description
docker run Run the container image.
--rm Delete the container after running the command.
-v ${ZDLC_LIB_DIR}:/build/lib:z The /build/lib host bind mount points to the directory containing the PyRuntime library. :z is required to share the volume if SELinux is installed.
-v ${ZDLC_CODE_DIR}:/code:z The /code host bind mount points to the directory with the calling program. :z is required to share the volume if SELinux is installed.
-v ${ZDLC_MODEL_DIR}:/model:z The /model host bind mount points to the directory with the model .so file. :z is required to share the volume if SELinux is installed.
--env PYTHONPATH=/build/lib When the container is launched, the PYTHONPATH environment variable is setup to point to /build/lib directory containing the PyRuntime library needed for execution.

Once complete, you'll see output like the following:

The input tensor dimensions are:
[1, 3, 224, 224]
A brief overview of the output tensor is:
[[-2.4883294   0.4591511   1.1298141  ... -2.8113475  -1.3842212
   2.6721394 ]
 [-5.064701    0.17290297 -1.866698   ...  0.39307398 -4.6048536
   2.116905  ]
 [-3.6744304   1.906144   -2.4807017  ... -0.96054727 -3.919518
   0.92789984]]
The dimensions of the output tensor are:
(3, 1000)

Note that the output values will be random since the input values are random.


IBM Z Integrated Accelerator for AI

IBM z16 systems include a new Integrated Accelerator for AI to enable real-time AI for transaction processing at scale. The IBM Z Deep Learning Compiler helps your new and existing deep learning models take advantage of this new accelerator.

Any IBM zSystem can be used to compile models to take advantage of the Integrated Accelerator for AI, including IBM z15 and older machines. However, if acceleration is enabled at compile time, the compiled model will only run on IBM zSystems which have the accelerator. Machines which have an accelerator can run models compiled without acceleration but those models will not take advantage of the accelerator.


Compiling models to utilize the IBM Z Integrated Accelerator for AI

Like other compilers, the IBM zDLC's default settings compile models so that they run on as many systems as possible. To use machine specific features, such as the Integrated Accelerator for AI, you must specify an additional option when compiling the model.

When set, supported ONNX Operators are directed to the accelerator instead of the CPU. The compile process handles routing the operations between the CPU and accelerator and any required data conversion. No changes are required to your model.

To compile a model to use the Integrated Accelerator for AI, The --maccel=NNPA option needs to be specified on the command line. Additionally, since the accelerator is only available for IBM z16 and greater, it is recommended to also use --mcpu=16.

Using the .so shared library example, the command line to compile models that take advantage of the Integrated Accelerator for AI is:

docker run --rm -v ${ZDLC_MODEL_DIR}:/workdir:z ${ZDLC_IMAGE} --EmitLib --O3 --mcpu=z16 --mtriple=s390x-ibm-loz --maccel=NNPA ${ZDLC_MODEL_NAME}.onnx

Once the model is built to use the IBM Z Integrated Accelerator for AI, no changes are required on the command line to run the model:

cp ${ZDLC_MODEL_DIR}/${ZDLC_MODEL_NAME}.so ${ZDLC_CODE_DIR}
docker run --rm -v ${ZDLC_CODE_DIR}:/code:z ${GCC_IMAGE_ID} /code/deep_learning_compiler_run_model_example

The same flags are required for compiling shared libraries for any language including Java and Python. Likewise, no additional steps are required when running the shared libraries.


Performance tips for IBM Z Integrated Accelerator for AI

When compiling models for the IBM Z Integrated Accelerator for AI, IBM zDLC optimizes models to take advantage of the accelerator when possible. In order to support a wide range of models, IBM zDLC will compile models so operators not supported by the accelerator, or operators with unsupported settings, run on the CPU.


Specifying input tensor dimensions

When running models with multiple dynamic dimensions (i.e. models with multiple -1 in their input signatures), using the --shapeInformation flag to set those dimensions to static values may improve model runtime performance. For some models, this allows the IBM zDLC to better determine at compile time which operations will be compatible with the accelerator.

For example, if a vision model has an input tensor with shape (-1, -1, -1, 3) representing (batch, height, width, channels), you may see increased performance by specifying the height and width dimensions at compile time. To do so, add --shapeInformation 0:-1x640x480x3 when compiling the model. If the model has mutliple input tensors, those can also be specified using --shapeInformation 0:-1x640x480x3,1:-1x100,2:... .

The --shapeInformation flag can be used with --onnx-op-stats to determine if specifying the shape enables more operations to run on the IBM Z Integrated Accelerator for AI. See View operation targets at compile time.


View operation targets at compile time

The IBM Z Deep Learning Compiler can optionally report the number of Operators that will run on CPU vs the IBM Z Integrated Accelerator for AI at compile time.

When compiling the model, add --onnx-op-stats [TXT|JSON]. Operations that begin with onnx.* will execute on CPU and operations that begin with zhigh.* are related to the IBM Z Integrated Accelerator for AI.


Device placement of operations

When compiling using the IBM Z Deep Learning Compiler, the target for operations can be obtained or specified. The target for operations is either on the CPU or the IBM Z Integrated Accelerator (NNPA) for AI.

Note that specifying a target of NNPA for an operation does not guarantee it will run on NNPA. This can happen for reasons such as:

  • NNPA is not supported for the operation.
  • NNPA does not support the tensor size used in the model.
  • The operation will run faster on the CPU.

For details on obtaining or specifying the target for device placement see:

Examples

  1. Save default device placement json to a file while compiling a model.

    docker run --rm -v ${ZDLC_MODEL_DIR}:/workdir:z ${ZDLC_IMAGE} --EmitLib --O3 --mcpu=z16 --mtriple=s390x-ibm-loz --maccel=NNPA --nnpa-save-device-placement-file=${ZDLC_MODEL_NAME}.json ${ZDLC_MODEL_NAME}.onnx
    • --EmitLib specifies to build a .so shared library of the model.
    • --nnpa-save-device-placement-file option specifies to save the device placement for the model's operations to a json file.

    Sample content of the model's output device placement json file. Notice the onnx.Gemm node with name Plus214-Times212_2 is targeted to device nnpa:

    {
    "device_placement": [
        {
            "device": "nnpa",
            "node_type": "onnx.Conv",
            "onnx_node_name": "Plus30-Convolution28-Initializer_Parameter6_0"
        },
        {
            "device": "nnpa",
            "node_type": "onnx.Relu",
            "onnx_node_name": "ReLU32"
        },
        {
            "device": "nnpa",
            "node_type": "onnx.MaxPoolSingleOut",
            "onnx_node_name": "Pooling66"
        },
        {
            "device": "nnpa",
            "node_type": "onnx.Conv",
            "onnx_node_name": "Plus112-Convolution110-Initializer_Parameter88_1"
        },
        {
            "device": "nnpa",
            "node_type": "onnx.Relu",
            "onnx_node_name": "ReLU114"
        },
        {
            "device": "nnpa",
            "node_type": "onnx.MaxPoolSingleOut",
            "onnx_node_name": "Pooling160"
        },
        {
            "device": "",
            "node_type": "onnx.Reshape",
            "onnx_node_name": "Times212_reshape0"
        },
        {
            "device": "nnpa",
            "node_type": "onnx.Gemm",
            "onnx_node_name": "Plus214-Times212_2"
        }
    ]
}

  2. Edit the ${ZDLC_MODEL_NAME}.json file and change target for onnx.Gemm operation Plus214-Times212_2.

    Sample content of the edited model's input device placement json file. Notice the onnx.Gemm node with name Plus214-Times212_2 has been modified to target device cpu instead of nnpa:

    {
    "device_placement": [
        {
            "device": "nnpa",
            "node_type": "onnx.Conv",
            "onnx_node_name": "Plus30-Convolution28-Initializer_Parameter6_0"
        },
        {
            "device": "nnpa",
            "node_type": "onnx.Relu",
            "onnx_node_name": "ReLU32"
        },
        {
            "device": "nnpa",
            "node_type": "onnx.MaxPoolSingleOut",
            "onnx_node_name": "Pooling66"
        },
        {
            "device": "nnpa",
            "node_type": "onnx.Conv",
            "onnx_node_name": "Plus112-Convolution110-Initializer_Parameter88_1"
        },
        {
            "device": "nnpa",
            "node_type": "onnx.Relu",
            "onnx_node_name": "ReLU114"
        },
        {
            "device": "nnpa",
            "node_type": "onnx.MaxPoolSingleOut",
            "onnx_node_name": "Pooling160"
        },
        {
            "device": "",
            "node_type": "onnx.Reshape",
            "onnx_node_name": "Times212_reshape0"
        },
        {
            "device": "cpu",
            "node_type": "onnx.Gemm",
            "onnx_node_name": "Plus214-Times212_2"
        }
    ]
}

  3. Build a model and loading the device placement specification for the model's operations from a json file.

    In this example the device placement json file is used to specify that the onnx.Gemm node with name Plus214-Times212_2 should target device cpu instead of nnpa as indicated in the device json file from the previous example.

    docker run --rm -v ${ZDLC_MODEL_DIR}:/workdir:z ${ZDLC_IMAGE} --EmitLib --O3 --mcpu=z16 --mtriple=s390x-ibm-loz --maccel=NNPA --nnpa-load-device-placement-file=${ZDLC_MODEL_NAME}.json ${ZDLC_MODEL_NAME}.onnx
    • --EmitLib specifies to build a .so shared library of the model.
    • --nnpa-load-device-placement-file option specifies to load the device placement specification for the model's operations from a json file.

Scope and Versioning Policy

Project Scope

IBM Z Deep Learning Compiler (IBM zDLC) follows a continuous release model with a cadence of 3-4 minor releases per year. Bug fixes are applied to the next minor release and are not ported to earlier major or minor releases.

Each release of IBM zDLC links to a specific version of ONNX-MLIR and supports CPU and IBM Z integrated AI Accelerator (NNPA) on IBM Z systems. Other ONNX-MLIR accelerators are not supported by IBM zDLC.

The following links lists supported operators, operator opset ranges, and any operator specific limitations. Operators that are not listed or usage of documented limitations are beyond IBM zDLC project scope:

Versioning Policy

IBM Z Deep Learning Compiler (IBM zDLC) follows the semantic versioning guidelines with a few deviations. These differences account for IBM zDLC’s nature as a compiler and are outlined below. Each zDLC release is versioned as: [MAJOR].[MINOR].[PATCH]

MAJOR / VERSION:

All releases with the same major number have runtime APIs that are compatible with earlier releases. That means programs designed to run model libraries generated by IBM zDLC X.0 will be able to run model libraries generated by later IBM zDLC releases that have the same major version, (i.e. X.0, X.1, X.2, etc.). Programs designed to run model libraries generated by IBM zDLC X.1 are able to run model libraries generated by IBM zDLC X.1, X.2, etc.

Changes in major releases indicate one or more of the following:

  • Changes related to runtime APIs that are incompatible with earlier releases. Programs that import model libraries generated by IBM zDLC might need to be updated between major releases depending on the affected runtime API languages.

  • Changes related to critical or general model compile time flags that are incompatible with earlier releases.

  • Significant feature additions beyond normal minor release updates.

Note: Prebuilt pybind11 PyRuntimes for versions of Python that have reached end of life can be removed without a major release increase change.

MINOR / FEATURE:

Minor releases typically contain new features, improvements, and bug fixes.

Generally, we strive to keep minor releases fully compatible with earlier releases. However, there may be cases where performance, debug, or accelerator specific compile time flags can be changed in incompatible ways during minor releases.

Support for end-of-life languages may be removed in minor releases.

PATCH / MAINTENANCE:

Patch releases contain only bug fixes, security updates or updates to non-IBM zDLC packages included in the IBM zDLC containers. IBM zDLC only introduces compatible changes in patch updates.


Obtaining IBM Z Deep Learning Compiler debug instrumentation

Instrumention debug information can be obtained during model runtime using three different methods during model compilation for the IBM Z Deep Learning Compiler.

  1. Profile IR option
  2. Instrument options
  3. NNPAUnsupportedOps option

Profile IR Option

Spcifying the --profile-ir option for the IBM Z Deep Learning Compiler to cause instrumention debug information to be printed during model runtime.

The values for the --profile-ir option are as follows:

Option Value Description
None No profiling. This is the defualt.
Onnx Profile for onnx ops.
ZHigh Profile for NNPA zhigh ops.

Examples

  1. Profiling for onnx ops:

    Compiling the model:

    docker run --rm -v ${ZDLC_MODEL_DIR}:/workdir:z ${ZDLC_IMAGE} --EmitLib --O3 --mcpu=z16 --mtriple=s390x-ibm-loz --maccel=NNPA --profile-ir=Onnx ${ZDLC_MODEL_NAME}.onnx

    Sample output when running model:

    #0) before onnx.Constant Time elapsed: 1692618036.154512 accumulated: 1692618036.154512 (Times212_reshape1)
#1) after  onnx.Constant Time elapsed: 0.000005 accumulated: 1692618036.154517 (Times212_reshape1)
#2) before onnx.Constant Time elapsed: 0.000002 accumulated: 1692618036.154519 (Plus112-Convolution110-Initializer_Parameter88)
#3) after  onnx.Constant Time elapsed: 0.000004 accumulated: 1692618036.154523 (Plus112-Convolution110-Initializer_Parameter88)

  2. Profiling for zhigh ops:

    Compiling the model:

    docker run --rm -v ${ZDLC_MODEL_DIR}:/workdir:z ${ZDLC_IMAGE} --EmitLib --O3 --mcpu=z16 --mtriple=s390x-ibm-loz --maccel=NNPA --profile-ir=ZHigh ${ZDLC_MODEL_NAME}.onnx

    Sample output when running model:

    #8) before zhigh.StickifiedConstant Time elapsed: 0.000003 accumulated: 1692617935.010772 (ReLU32-Plus30-Convolution28-Initializer_Parameter6)
#9) after  zhigh.StickifiedConstant Time elapsed: 0.000004 accumulated: 1692617935.010776 (ReLU32-Plus30-Convolution28-Initializer_Parameter6)
#10) before zhigh.Conv2D Time elapsed: 0.000003 accumulated: 1692617935.010779 (ReLU32-Plus30-Convolution28-Initializer_Parameter6)
#11) after  zhigh.Conv2D Time elapsed: 0.000161 accumulated: 1692617935.010940 (ReLU32-Plus30-Convolution28-Initializer_Parameter6)

Notes:

  • The call to initialize instrumentation, OMInstrumentInit, must be done before loading the model shared library.
  • Runtime instrumenting will affect model performance due to the additional tracking and printing.

Instrument Options

Specifying the instrument options for the IBM Z Deep Learning Compiler to cause instrumention debug information to be printed during model runtime.

There are three types of instrument options that can be specified.

  1. The stage to be instrumented is specified using the --instrument-stage option with value:

    Option Value Description
    Onnx Get onnx-level profiling. If "--maccel=NNPA" is also specified then get profile onnx ops before lowering to zhigh.
    ZHigh Get NNPA profiling for onnx and zhigh ops.
    ZLow Get NNPA profiling for zlow ops.

  2. The operations to be instrumented are specified using the --instrument-ops option with value:

    Option Value Description
    NONE or "" No instrumentation.
    ops1,ops2, ... Multiple ops. e.g. onnx.Conv,onnx.Add for Conv and Add ops.
    ops.* Ops using * wildcard. e.g. onnx.* for all onnx operations.

  3. The instrumentation actions are specified using the following options:

    Option Description
    --InstrumentBeforeOp Insert instrument before op.
    --InstrumentAfterOp Insert instrument after op.
    --InstrumentReportTime Instrument runtime reports time usage.
    --InstrumentReportMemory Instrument runtime reports memory usage.

Examples

  1. Profiling time for onnx ops before lowering to zhigh ops:

    Compiling the model:

    docker run --rm -v ${ZDLC_MODEL_DIR}:/workdir:z ${ZDLC_IMAGE} --EmitLib --O3 --mcpu=z16 --mtriple=s390x-ibm-loz --maccel=NNPA --instrument-stage=Onnx --instrument-ops=onnx.* --InstrumentBeforeOp --InstrumentAfterOp --InstrumentReportTime ${ZDLC_MODEL_NAME}.onnx

    Sample output when running model:

    #  0) before onnx.Constant Time elapsed: 1691688479.493696 accumulated: 1691688479.493696
#  1) after  onnx.Constant Time elapsed: 0.000005 accumulated: 1691688479.493701
#  2) before onnx.Constant Time elapsed: 0.000004 accumulated: 1691688479.493705
#  3) after  onnx.Constant Time elapsed: 0.000004 accumulated: 1691688479.493709

  2. Profiling time for onnx and zhigh ops:

    Compiling the model:

    docker run --rm -v ${ZDLC_MODEL_DIR}:/workdir:z ${ZDLC_IMAGE} --EmitLib --O3 --mcpu=z16 --mtriple=s390x-ibm-loz --maccel=NNPA --instrument-stage=ZHigh --instrument-ops=onnx.*,zhigh.* --InstrumentBeforeOp --InstrumentAfterOp --InstrumentReportTime ${ZDLC_MODEL_NAME}.onnx

    Sample output when running model:

    # 24) before onnx.Reshape Time elapsed: 0.000002 accumulated: 1691688806.270982 (Times212_reshape0)
# 25) after  onnx.Reshape Time elapsed: 0.000001 accumulated: 1691688806.270983 (Times212_reshape0)
# 26) before zhigh.Stick Time elapsed: 0.000002 accumulated: 1691688806.270985
# 27) after  zhigh.Stick Time elapsed: 0.000003 accumulated: 1691688806.270988

  3. Profiling memory for zlow ops:

    Compiling the model:

    docker run --rm -v ${ZDLC_MODEL_DIR}:/workdir:z ${ZDLC_IMAGE} --EmitLib --O3 --mcpu=z16 --mtriple=s390x-ibm-loz --maccel=NNPA --instrument-stage=ZLow --instrument-ops=zlow.* --InstrumentBeforeOp --InstrumentAfterOp --InstrumentReportMemory ${ZDLC_MODEL_NAME}.onnx

    Sample output when running model:

    # 14) before zlow.matmul VMem:  5456
# 15) after  zlow.matmul VMem:  5456
# 16) before zlow.add VMem:  5456
# 17) after  zlow.add VMem:  5456

Notes:

  • The call to initialize instrumentation, OMInstrumentInit, must be done before loading the model shared library.
  • Runtime instrumenting will affect model performance due to the additional tracking and printing.

NNPAUnsupportedOps option

The --opt-report=NNPAUnsupportedOps flag is used to generate a report to inform the user why certain operations are not running on NNPA. The report is displayed in CSV format with the following headers ==NNPA-UNSUPPORTEDOPS-REPORT==, <operation name>, <node name>, <reason>. This is definitely a great option to utilize when further debugging models in zDLC.

Example

$  --EmitZHighIR --O3 --mcpu=z16 --mtriple=s390x-ibm-loz --maccel=NNPA model.onnx --opt-report=NNPAUnsupportedOps

==NNPA-UNSUPPORTEDOPS-REPORT==, onnx.Mul, Mul_28, Element type is not F16 or F32.
==NNPA-UNSUPPORTEDOPS-REPORT==, onnx.Add, Add_31, Element type is not F16 or F32.
==NNPA-UNSUPPORTEDOPS-REPORT==, onnx.Div, Div_122, Rank 0 is not supported. zAIU only supports rank in range of (0, 4].
==NNPA-UNSUPPORTEDOPS-REPORT==, onnx.Add, Add_125, Rank 0 is not supported. zAIU only supports rank in range of (0, 4].
==NNPA-UNSUPPORTEDOPS-REPORT==, onnx.Mul, Mul_128, Rank 0 is not supported. zAIU only supports rank in range of (0, 4].
==NNPA-UNSUPPORTEDOPS-REPORT==, onnx.Div, Div_216, Rank 0 is not supported. zAIU only supports rank in range of (0, 4].
==NNPA-UNSUPPORTEDOPS-REPORT==, onnx.Add, Add_219, Rank 0 is not supported. zAIU only supports rank in range of (0, 4].
==NNPA-UNSUPPORTEDOPS-REPORT==, onnx.Mul, Mul_222, Rank 0 is not supported. zAIU only supports rank in range of (0, 4].
==NNPA-UNSUPPORTEDOPS-REPORT==, onnx.Div, Div_310, Rank 0 is not supported. zAIU only supports rank in range of (0, 4].
==NNPA-UNSUPPORTEDOPS-REPORT==, onnx.Add, Add_313, Rank 0 is not supported. zAIU only supports rank in range of (0, 4].
==NNPA-UNSUPPORTEDOPS-REPORT==, onnx.Mul, Mul_316, Rank 0 is not supported. zAIU only supports rank in range of (0, 4].
==NNPA-UNSUPPORTEDOPS-REPORT==, onnx.Div, Div_404, Rank 0 is not supported. zAIU only supports rank in range of (0, 4].
==NNPA-UNSUPPORTEDOPS-REPORT==, onnx.Add, Add_407, Rank 0 is not supported. zAIU only supports rank in range of (0, 4].
==NNPA-UNSUPPORTEDOPS-REPORT==, onnx.Mul, Mul_410, Rank 0 is not supported. zAIU only supports rank in range of (0, 4].

Removing IBM Z Deep Learning Compiler


First, find the IMAGE ID for the container image.

docker images

Then delete the image using the IMAGE ID.

docker rmi IMAGE-ID

If an in-use error occurs while attempting to delete the container image, use the docker ps -a command to show any running containers. Use the docker stop and docker rm commands to remove the running instances of the container. Then re-run the docker rmi command.