OpenMP support in ROCm

OpenMP support in ROCm#

OpenMP release updates#

Additional optimizations for OpenMP offload
Host-exec services for printing on-device and doing malloc/free from device
Improved support for OMPT, the OpenMP tools interface
Driver improvements for multi-image and Target ID features
OMPD support, implements OpenMP debugger interfaces
ASan support for OpenMP
MI300A Unified Shared Memory support
Heterogeneous Debugging - A prototype of debug-info supporting AMDGPU targets, affecting most parts of the compiler, is implemented as documented in docs/AMDGPULLVMExtensionsForHeterogeneousDebugging.rst but is an ongoing work-in-progress. Fundamental changes are expected as parts of the design are adapted for upstreaming.

OpenMP introduction#

The ROCm installation includes an LLVM-based implementation that fully supports the OpenMP 4.5 standard and a subset of OpenMP 5.0, 5.1, and 5.2 standards. Fortran, C/C++ compilers, and corresponding runtime libraries are included. Along with host APIs, the OpenMP compilers support offloading code and data onto GPU devices. This document briefly describes the installation location of the OpenMP toolchain, example usage of device offloading, and usage of rocprof with OpenMP applications. The GPUs supported are the same as those supported by this ROCm release. See the list of supported GPUs for Linux and Windows.

The ROCm OpenMP compiler is implemented using LLVM compiler technology. The following image illustrates the internal steps taken to translate a user’s application into an executable that can offload computation to the AMDGPU. The compilation is a two-pass process. Pass 1 compiles the application to generate the CPU code and Pass 2 links the CPU code to the AMDGPU device code.

OpenMP toolchain

Installation#

The OpenMP toolchain is automatically installed as part of the standard ROCm installation and is available under /opt/rocm-{version}/llvm. The sub-directories are:

bin: Compilers (flang and clang) and other binaries.
examples: The usage section below shows how to compile and run these programs.
include: Header files.
lib: Libraries including those required for target offload.
lib-debug: Debug versions of the above libraries.

OpenMP: usage#

The example programs can be compiled and run by pointing the environment variable ROCM_PATH to the ROCm install directory.

Example:

export ROCM_PATH=/opt/rocm-{version}
cd $ROCM_PATH/share/openmp-extras/examples/openmp/veccopy
sudo make run

Note

sudo is required since we are building inside the /opt directory. Alternatively, copy the files to your home directory first.

The above invocation of Make compiles and runs the program. Note the options that are required for target offload from an OpenMP program:

-fopenmp --offload-arch=<gpu-arch>

Note

The compiler also accepts the alternative offloading notation:

-fopenmp -fopenmp-targets=amdgcn-amd-amdhsa -Xopenmp-target=amdgcn-amd-amdhsa -march=<gpu-arch>

Obtain the value of gpu-arch by running the following command:

% /opt/rocm-{version}/bin/rocminfo | grep gfx

See the complete list of compiler command-line references.

Using `rocprof` with OpenMP#

The following steps describe a typical workflow for using rocprof with OpenMP code compiled with AOMP:

Run rocprof with the program command line:
```
% rocprof <application> <args>
```
This produces a results.csv file in the user’s current directory that shows basic stats such as kernel names, grid size, number of registers used, etc. The user can choose to specify the preferred output file name using the o option.
Add options for a detailed result:
```
--stats: % rocprof --stats <application> <args>
```
The stats option produces timestamps for the kernels. Look into the output CSV file for the field, DurationNs, which is useful in getting an understanding of the critical kernels in the code.

Apart from --stats, the option --timestamp on produces a timestamp for the kernels.
After learning about the required kernels, the user can take a detailed look at each one of them. rocprof has support for hardware counters: a set of basic and a set of derived ones. See the complete list of counters using options –list-basic and –list-derived. rocprof accepts either a text or an XML file as an input.

For more details on rocprof, refer to the ROCProfilerV1 User Manual.

Using tracing options#

Prerequisite: When using the --sys-trace option, compile the OpenMP program with:

    -Wl,-rpath,/opt/rocm-{version}/lib -lamdhip64

The following tracing options are widely used to generate useful information:

--hsa-trace: This option is used to get a JSON output file with the HSA API execution traces and a flat profile in a CSV file.
--sys-trace: This allows programmers to trace both HIP and HSA calls. Since this option results in loading libamdhip64.so, follow the prerequisite as mentioned above.

A CSV and a JSON file are produced by the above trace options. The CSV file presents the data in a tabular format, and the JSON file can be visualized using Google Chrome at chrome://tracing/ or Perfetto. Navigate to Chrome or Perfetto and load the JSON file to see the timeline of the HSA calls.

For more details on tracing, refer to the ROCProfilerV1 User Manual.

Environment variables#

Environment Variable	Purpose
`OMP_NUM_TEAMS`	To set the number of teams for kernel launch, which is otherwise chosen by the implementation by default. You can set this number (subject to implementation limits) for performance tuning.
`LIBOMPTARGET_KERNEL_TRACE`	To print useful statistics for device operations. Setting it to 1 and running the program emits the name of every kernel launched, the number of teams and threads used, and the corresponding register usage. Setting it to 2 additionally emits timing information for kernel launches and data transfer operations between the host and the device.
`LIBOMPTARGET_INFO`	To print informational messages from the device runtime as the program executes. Setting it to a value of 1 or higher, prints fine-grain information and setting it to -1 prints complete information.
`LIBOMPTARGET_DEBUG`	To get detailed debugging information about data transfer operations and kernel launch when using a debug version of the device library. Set this environment variable to 1 to get the detailed information from the library.
`GPU_MAX_HW_QUEUES`	To set the number of HSA queues in the OpenMP runtime. The HSA queues are created on demand up to the maximum value as supplied here. The queue creation starts with a single initialized queue to avoid unnecessary allocation of resources. The provided value is capped if it exceeds the recommended, device-specific value.
`LIBOMPTARGET_AMDGPU_MAX_ASYNC_COPY_BYTES`	To set the threshold size up to which data transfers are initiated asynchronously. The default threshold size is 110241024 bytes (1MB).
`OMPX_FORCE_SYNC_REGIONS`	To force the runtime to execute all operations synchronously, i.e., wait for an operation to complete immediately. This affects data transfers and kernel execution. While it is mainly designed for debugging, it may have a minor positive effect on performance in certain situations.

OpenMP: features#

The OpenMP programming model is greatly enhanced with the following new features implemented in the past releases.

Asynchronous behavior in OpenMP target regions#

Controlling Asynchronous Behavior

The OpenMP offloading runtime executes in an asynchronous fashion by default, allowing multiple data transfers to start concurrently. However, if the data to be transferred becomes larger than the default threshold of 1MB, the runtime falls back to a synchronous data transfer. The buffers that have been locked already are always executed asynchronously. You can overrule this default behavior by setting LIBOMPTARGET_AMDGPU_MAX_ASYNC_COPY_BYTES and OMPX_FORCE_SYNC_REGIONS. See the Environment Variables table for details.

Multithreaded Offloading on the Same Device

The libomptarget plugin for GPU offloading allows creation of separate configurable HSA queues per chiplet, which enables two or more threads to concurrently offload to the same device.

Parallel Memory Copy Invocations

Implicit asynchronous execution of single target region enables parallel memory copy invocations.

Unified shared memory#

Unified Shared Memory (USM) provides a pointer-based approach to memory management. To implement USM, fulfill the following system requirements along with Xnack capability.

Prerequisites#

Linux Kernel versions above 5.14
Latest KFD driver packaged in ROCm stack
Xnack, as USM support can only be tested with applications compiled with Xnack capability

Xnack capability#

When enabled, Xnack capability allows GPU threads to access CPU (system) memory, allocated with OS-allocators, such as malloc, new, and mmap. Xnack must be enabled both at compile- and run-time. To enable Xnack support at compile-time, use:

--offload-arch=gfx908:xnack+

Or use another functionally equivalent option Xnack-any:

--offload-arch=gfx908

To enable Xnack functionality at runtime on a per-application basis, use environment variable:

HSA_XNACK=1

When Xnack support is not needed:

Build the applications to maximize resource utilization using:

--offload-arch=gfx908:xnack-

At runtime, set the HSA_XNACK environment variable to 0.

Unified shared memory pragma#

This OpenMP pragma is available on MI200 through xnack+ support.

omp requires unified_shared_memory

As stated in the OpenMP specifications, this pragma makes the map clause on target constructs optional. By default, on MI200, all memory allocated on the host is fine grain. Using the map clause on a target clause is allowed, which transforms the access semantics of the associated memory to coarse grain.

A simple program demonstrating the use of this feature is:
$ cat parallel_for.cpp
#include <stdlib.h>
#include <stdio.h>

#define N 64
#pragma omp requires unified_shared_memory
int main() {
  int n = N;
  int *a = new int[n];
  int *b = new int[n];

  for(int i = 0; i < n; i++)
    b[i] = i;

  #pragma omp target parallel for map(to:b[:n])
  for(int i = 0; i < n; i++)
    a[i] = b[i];

  for(int i = 0; i < n; i++)
    if(a[i] != i)
      printf("error at %d: expected %d, got %d\n", i, i+1, a[i]);

  return 0;
}
$ clang++ -O2 -target x86_64-pc-linux-gnu -fopenmp --offload-arch=gfx90a:xnack+ parallel_for.cpp
$ HSA_XNACK=1 ./a.out

In the above code example, pointer “a” is not mapped in the target region, while pointer “b” is. Both are valid pointers on the GPU device and passed by-value to the kernel implementing the target region. This means the pointer values on the host and the device are the same.

The difference between the memory pages pointed to by these two variables is that the pages pointed by “a” are in fine-grain memory, while the pages pointed to by “b” are in coarse-grain memory during and after the execution of the target region. This is accomplished in the OpenMP runtime library with calls to the ROCr runtime to set the pages pointed by “b” as coarse grain.

OMPT target support#

The OpenMP runtime in ROCm implements a subset of the OMPT device APIs, as described in the OpenMP specification document. These APIs allow first-party tools to examine the profile and kernel traces that execute on a device. A tool can register callbacks for data transfer and kernel dispatch entry points or use APIs to start and stop tracing for device-related activities such as data transfer and kernel dispatch timings and associated metadata. If device tracing is enabled, trace records for device activities are collected during program execution and returned to the tool using the APIs described in the specification.

The following example demonstrates how a tool uses the supported OMPT target APIs. The README in /opt/rocm/llvm/examples/tools/ompt outlines the steps to be followed, and the provided example can be run as shown below:

cd $ROCM_PATH/share/openmp-extras/examples/tools/ompt/veccopy-ompt-target-tracing
sudo make run

The file veccopy-ompt-target-tracing.c simulates how a tool initiates device activity tracing. The file callbacks.h shows the callbacks registered and implemented by the tool.

Floating point atomic operations#

The MI200-series GPUs support the generation of hardware floating-point atomics using the OpenMP atomic pragma. The support includes single- and double-precision floating-point atomic operations. The programmer must ensure that the memory subjected to the atomic operation is in coarse-grain memory by mapping it explicitly with the help of map clauses when not implicitly mapped by the compiler as per the OpenMP specifications. This makes these hardware floating-point atomic instructions “fast,” as they are faster than using a default compare-and-swap loop scheme, but at the same time “unsafe,” as they are not supported on fine-grain memory. The operation in unified_shared_memory mode also requires programmers to map the memory explicitly when not implicitly mapped by the compiler.

To request fast floating-point atomic instructions at the file level, use compiler flag -munsafe-fp-atomics or a hint clause on a specific pragma:

double a = 0.0;
#pragma omp atomic hint(AMD_fast_fp_atomics)
a = a + 1.0;

Note

AMD_unsafe_fp_atomics is an alias for AMD_fast_fp_atomics, and AMD_safe_fp_atomics is implemented with a compare-and-swap loop.

To disable the generation of fast floating-point atomic instructions at the file level, build using the option -msafe-fp-atomics or use a hint clause on a specific pragma:

double a = 0.0;
#pragma omp atomic hint(AMD_safe_fp_atomics)
a = a + 1.0;

The hint clause value always has a precedence over the compiler flag, which allows programmers to create atomic constructs with a different behavior than the rest of the file.

See the example below, where the user builds the program using -msafe-fp-atomics to select a file-wide “safe atomic” compilation. However, the fast atomics hint clause over variable “a” takes precedence and operates on “a” using a fast/unsafe floating-point atomic, while the variable “b” in the absence of a hint clause is operated upon using safe floating-point atomics as per the compiler flag.

double a = 0.0;.
#pragma omp atomic hint(AMD_fast_fp_atomics)
a = a + 1.0;

double b = 0.0;
#pragma omp atomic
b = b + 1.0;

AddressSanitizer tool#

AddressSanitizer (ASan) is a memory error detector tool utilized by applications to detect various errors ranging from spatial issues such as out-of-bound access to temporal issues such as use-after-free. The AOMP compiler supports ASan for AMD GPUs with applications written in both HIP and OpenMP.

Features supported on host platform (Target x86_64):

Use-after-free
Buffer overflows
Heap buffer overflow
Stack buffer overflow
Global buffer overflow
Use-after-return
Use-after-scope
Initialization order bugs

Features supported on AMDGPU platform (amdgcn-amd-amdhsa):

Heap buffer overflow
Global buffer overflow

Software (kernel/OS) requirements: Unified Shared Memory support with Xnack capability. See the section on Unified Shared Memory for prerequisites and details on Xnack.

Example:

Heap buffer overflow

void  main() {
.......  // Some program statements
.......  // Some program statements
#pragma omp target map(to : A[0:N], B[0:N]) map(from: C[0:N])
{
#pragma omp parallel for
    for(int i =0 ; i < N; i++){
    C[i+10] = A[i] + B[i];
  }   // end of for loop
}
.......   // Some program statements
}// end of main

See the complete sample code for heap buffer overflow here.

Global buffer overflow

#pragma omp declare target
   int A[N],B[N],C[N];
#pragma omp end declare target
void main(){
......  // some program statements
......  // some program statements
#pragma omp target data map(to:A[0:N],B[0:N]) map(from: C[0:N])
{
#pragma omp target update to(A,B)
#pragma omp target parallel for
for(int i=0; i<N; i++){
    C[i]=A[i*100]+B[i+22];
} // end of for loop
#pragma omp target update from(C)
}
........  // some program statements
} // end of main

See the complete sample code for global buffer overflow here.

Clang compiler option for kernel optimization#

You can use the clang compiler option -fopenmp-target-fast for kernel optimization if certain constraints implied by its component options are satisfied. -fopenmp-target-fast enables the following options:

-fopenmp-target-ignore-env-vars: It enables code generation of specialized kernels including no-loop and Cross-team reductions.
-fopenmp-assume-no-thread-state: It enables the compiler to assume that no thread in a parallel region modifies an Internal Control Variable (ICV), thus potentially reducing the device runtime code execution.
-fopenmp-assume-no-nested-parallelism: It enables the compiler to assume that no thread in a parallel region encounters a parallel region, thus potentially reducing the device runtime code execution.
-O3 if no -O* is specified by the user.

Specialized kernels#

Clang will attempt to generate specialized kernels based on compiler options and OpenMP constructs. The following specialized kernels are supported:

No-loop
Big-jump-loop
Cross-team reductions

To enable the generation of specialized kernels, follow these guidelines:

Do not specify teams, threads, and schedule-related environment variables. The num_teams clause in an OpenMP target construct acts as an override and prevents the generation of the no-loop kernel. If the specification of num_teams clause is a user requirement then clang tries to generate the big-jump-loop kernel instead of the no-loop kernel.
Assert the absence of the teams, threads, and schedule-related environment variables by adding the command-line option -fopenmp-target-ignore-env-vars.
To automatically enable the specialized kernel generation, use -Ofast or -fopenmp-target-fast for compilation.
To disable specialized kernel generation, use -fno-openmp-target-ignore-env-vars.

No-loop kernel generation#

The no-loop kernel generation feature optimizes the compiler performance by generating a specialized kernel for certain OpenMP target constructs such as target teams distribute parallel for. The specialized kernel generation feature assumes every thread executes a single iteration of the user loop, which leads the runtime to launch a total number of GPU threads equal to or greater than the iteration space size of the target region loop. This allows the compiler to generate code for the loop body without an enclosing loop, resulting in reduced control-flow complexity and potentially better performance.

Big-jump-loop kernel generation#

A no-loop kernel is not generated if the OpenMP teams construct uses a num_teams clause. Instead, the compiler attempts to generate a different specialized kernel called the big-jump-loop kernel. The compiler launches the kernel with a grid size determined by the number of teams specified by the OpenMP num_teams clause and the blocksize chosen either by the compiler or specified by the corresponding OpenMP clause.

Cross-team optimized reduction kernel generation#

If the OpenMP construct has a reduction clause, the compiler attempts to generate optimized code by utilizing efficient cross-team communication. New APIs for cross-team reduction are implemented in the device runtime and are automatically generated by clang.

OpenMP support in ROCm

Contents

OpenMP support in ROCm#

OpenMP release updates#

OpenMP introduction#

Installation#

OpenMP: usage#

Using rocprof with OpenMP#

Using tracing options#

Environment variables#

OpenMP: features#

Asynchronous behavior in OpenMP target regions#

Unified shared memory#

Prerequisites#

Xnack capability#

Unified shared memory pragma#

OMPT target support#

Floating point atomic operations#

AddressSanitizer tool#

Clang compiler option for kernel optimization#

Specialized kernels#

No-loop kernel generation#

Big-jump-loop kernel generation#

Cross-team optimized reduction kernel generation#

Using `rocprof` with OpenMP#