[OpenMP] Add environment variables to change stack / heap size in the CUDA plugin

[jhuber6]

Summary

This patch adds support for two environment variables to configure the device.
`LIBOMPTARGET_STACK_SIZE` sets the amount of memory in bytes that each thread
has for its stack. `LIBOMPTARGET_HEAP_SIZE` sets the amount of heap memory
that can be allocated using malloc / free on the device.

[jdoerfert]

LG

[abhinavgaba]

openmp/libomptarget/plugins/cuda/src/rtl.cpp

652

These enums don't seem to be defined in cuda.h, or somewhere else. Can you please take a look?

[jdoerfert]

openmp/libomptarget/plugins/cuda/src/rtl.cpp

652

My cuda.h defines them.

/opt/cuda/targets/x86_64-linux/include/cuda.h
1130:    CU_LIMIT_STACK_SIZE                       = 0x00, /**< GPU thread stack size */

https://docs.nvidia.com/cuda/cuda-driver-api/group__CUDA__TYPES.html lists them as well.
Are you sure they are not there?

[abhinavgaba]

openmp/libomptarget/plugins/cuda/src/rtl.cpp

652

I meant in the compiler sources. The other enums used in libomptarget seem to be defined in openmp/libomptarget/plugins/cuda/dynamic_cuda/cuda.h. Absence of these definitions on a machine that doesn't have Cuda drivers installed, is causing a build fail for me with these errors:

llvm/openmp/libomptarget/plugins/cuda/src/rtl.cpp: In member function ‘int {anonymous}::DeviceRTLTy::initDevice(int)’:
llvm/openmp/libomptarget/plugins/cuda/src/rtl.cpp:649:25: error: ‘CU_LIMIT_STACK_SIZE’ was not declared in this scope
if (cuCtxSetLimit(CU_LIMIT_STACK_SIZE, StackLimit) != CUDA_SUCCESS)
                  ^~~~~~~~~~~~~~~~~~~

I looked at the buildbots, but they are skipping building the cuda plugin altogether, so don't report any fails. For instance, https://lab.llvm.org/buildbot/#/builders/84/builds/12107/steps/4/logs/stdio has this in the log:

-- Could NOT find LIBOMPTARGET_DEP_CUDA_DRIVER (missing: LIBOMPTARGET_DEP_CUDA_DRIVER_LIBRARIES) 
-- Could NOT find LIBOMPTARGET_DEP_VEO (missing: LIBOMPTARGET_DEP_VEO_LIBRARIES LIBOMPTARGET_DEP_VEOSINFO_LIBRARIES LIBOMPTARGET_DEP_VEO_INCLUDE_DIRS) 
-- LIBOMPTARGET: Building offloading runtime library libomptarget.
...
-- LIBOMPTARGET: Not building CUDA offloading plugin: libelf dependency not found.
...
-- check-libomptarget does nothing.

[abhinavgaba]

openmp/libomptarget/plugins/cuda/src/rtl.cpp

652

This patch makes the build pass for me, but I have no way to verify it. @jhuber6 can you please take a look?

diff --git a/openmp/libomptarget/plugins/cuda/dynamic_cuda/cuda.cpp b/openmp/libomptarget/plugins/cuda/dynamic_cuda/cuda.cpp
index c84b3814065e..235efd2728de 100644
--- a/openmp/libomptarget/plugins/cuda/dynamic_cuda/cuda.cpp
+++ b/openmp/libomptarget/plugins/cuda/dynamic_cuda/cuda.cpp
@@ -61,6 +61,9 @@ DLWRAP(cuDeviceCanAccessPeer, 3);
 DLWRAP(cuCtxEnablePeerAccess, 2);
 DLWRAP(cuMemcpyPeerAsync, 6);

+DLWRAP(cuCtxGetLimit, 2);
+DLWRAP(cuCtxSetLimit, 2);
+
 DLWRAP_FINALIZE();

 #ifndef DYNAMIC_CUDA_PATH
diff --git a/openmp/libomptarget/plugins/cuda/dynamic_cuda/cuda.h b/openmp/libomptarget/plugins/cuda/dynamic_cuda/cuda.h
index 045c39cacc97..17aa2a12ef6c 100644
--- a/openmp/libomptarget/plugins/cuda/dynamic_cuda/cuda.h
+++ b/openmp/libomptarget/plugins/cuda/dynamic_cuda/cuda.h
@@ -34,6 +34,17 @@ typedef enum CUstream_flags_enum {
   CU_STREAM_NON_BLOCKING = 0x1,
 } CUstream_flags;

+typedef enum CUlimit_enum {
+  CU_LIMIT_STACK_SIZE = 0x0,
+  CU_LIMIT_PRINTF_FIFO_SIZE = 0x1,
+  CU_LIMIT_MALLOC_HEAP_SIZE = 0x2,
+  CU_LIMIT_DEV_RUNTIME_SYNC_DEPTH = 0x3,
+  CU_LIMIT_DEV_RUNTIME_PENDING_LAUNCH_COUNT = 0x4,
+  CU_LIMIT_MAX_L2_FETCH_GRANULARITY = 0x5,
+  CU_LIMIT_PERSISTING_L2_CACHE_SIZE = 0x6,
+  CU_LIMIT_MAX
+} CUlimit;
+
 typedef enum CUdevice_attribute_enum {
   CU_DEVICE_ATTRIBUTE_MAX_BLOCK_DIM_X = 2,
   CU_DEVICE_ATTRIBUTE_MAX_GRID_DIM_X = 5,
@@ -100,4 +111,7 @@ CUresult cuCtxEnablePeerAccess(CUcontext, unsigned);
 CUresult cuMemcpyPeerAsync(CUdeviceptr, CUcontext, CUdeviceptr, CUcontext,
                            size_t, CUstream);

+CUresult cuCtxGetLimit(size_t *, CUlimit);
+CUresult cuCtxSetLimit(CUlimit, size_t);
+

[jdoerfert]

openmp/libomptarget/plugins/cuda/src/rtl.cpp

652

Right, that makes sense.

The above looks good to me. Could you commit it?

[abhinavgaba]

Johannes,

I do not have the access rights needed for git push. Could you please help commit the fix?

Thanks,
Abhinav

• libomptarget_cuda_build_fix.patch1 KBDownload

[abhinavgaba]

In D106627#2899112, @abhinavgaba wrote:

Johannes,

I do not have the access rights needed for git push. Could you please help commit the fix?

Thanks,
Abhinav

• libomptarget_cuda_build_fix.patch1 KBDownload

@pengfei kindly agreed to help commit it. Thanks, Pengfei.

[pengfei]

There're three fails in my local check-all. Is this expected?

Failed Tests (3):
  libomptarget :: x86_64-pc-linux-gnu :: offloading/memory_manager.cpp
  libomptarget :: x86_64-pc-linux-gnu :: offloading/parallel_offloading_map.cpp
  libomptarget :: x86_64-pc-linux-gnu :: offloading/taskloop_offload_nowait.cpp

[abhinavgaba]

In D106627#2899301, @pengfei wrote:

There're three fails in my local check-all. Is this expected?

Failed Tests (3):
  libomptarget :: x86_64-pc-linux-gnu :: offloading/memory_manager.cpp
  libomptarget :: x86_64-pc-linux-gnu :: offloading/parallel_offloading_map.cpp
  libomptarget :: x86_64-pc-linux-gnu :: offloading/taskloop_offload_nowait.cpp

Pengfei, these fails are from a different part of code, that should not be affected by this change.

I see these fails in my environment (RHEL 8.2, gcc 8.3.1) for commit 4a76bd0e (previous buildable commit) as well.

[pengfei]

In D106627#2899368, @abhinavgaba wrote:

In D106627#2899301, @pengfei wrote:

There're three fails in my local check-all. Is this expected?

Failed Tests (3):
  libomptarget :: x86_64-pc-linux-gnu :: offloading/memory_manager.cpp
  libomptarget :: x86_64-pc-linux-gnu :: offloading/parallel_offloading_map.cpp
  libomptarget :: x86_64-pc-linux-gnu :: offloading/taskloop_offload_nowait.cpp

Pengfei, these fails are from a different part of code, that should not be affected by this change.

I see these fails in my environment (RHEL 8.2, gcc 8.3.1) for commit 4a76bd0e (previous buildable commit) as well.

I see. Then I commited it rGf7c92995c0e1f95.

[abhinavgaba]

I see. Then I commited it rGf7c92995c0e1f95.

Thanks, Pengfei.

[jhuber6]

In D106627#2899398, @pengfei wrote:

In D106627#2899368, @abhinavgaba wrote:

In D106627#2899301, @pengfei wrote:

There're three fails in my local check-all. Is this expected?

Failed Tests (3):
  libomptarget :: x86_64-pc-linux-gnu :: offloading/memory_manager.cpp
  libomptarget :: x86_64-pc-linux-gnu :: offloading/parallel_offloading_map.cpp
  libomptarget :: x86_64-pc-linux-gnu :: offloading/taskloop_offload_nowait.cpp

Pengfei, these fails are from a different part of code, that should not be affected by this change.

I see these fails in my environment (RHEL 8.2, gcc 8.3.1) for commit 4a76bd0e (previous buildable commit) as well.

I see. Then I commited it rGf7c92995c0e1f95.

Thanks for fixing it, I forgot about building without the SDK.

openmp/docs/design/Runtimes.rst

 library at runtime. This allows the user to obtain useful runtime information as
 well as enable or disable certain features. A full list of supported environment
 variables is defined below.
 
     * ``LIBOMPTARGET_DEBUG=<Num>``
     * ``LIBOMPTARGET_PROFILE=<Filename>``
     * ``LIBOMPTARGET_MEMORY_MANAGER_THRESHOLD=<Num>``
     * ``LIBOMPTARGET_INFO=<Num>``
+    * ``LIBOMPTARGET_HEAP_SIZE=<Num>``
+    * ``LIBOMPTARGET_STACK_SIZE=<Num>``
 
 LIBOMPTARGET_DEBUG
 """"""""""""""""""
 
 ``LIBOMPTARGET_DEBUG`` controls whether or not debugging information will be
 displayed. This feature is only availible if ``libomptarget`` was built with
 ``-DOMPTARGET_DEBUG``. The debugging output provided is intended for use by
 ``libomptarget`` developers. More user-friendly output is presented when using
double sum(double *A, std::size_t N) {
       double sum = 0.0;
     #pragma omp target teams distribute parallel for reduction(+:sum) map(to:A[0 : N])
       for (int i = 0; i < N; ++i)
         sum += A[i];
     
       return sum;
     }
 
+LIBOMPTARGET_STACK_SIZE
+"""""""""""""""""""""""
+
+This environment variable sets the stack size in bytes for the CUDA plugin. This
+can be used to increase or decrease the standard amount of memory reserved for
+each thread's stack.
+
+LIBOMPTARGET_HEAP_SIZE
+"""""""""""""""""""""""
+
+This environment variable sets the amount of memory in bytes that can be
+allocated using ``malloc`` and ``free`` for the CUDA plugin. This is necessary
+for some applications that allocate too much memory either through the user or
+globalization.
+
 .. toctree::
    :hidden:
    :maxdepth: 1
 
    Offloading
 
 LLVM/OpenMP Target Host Runtime Plugins (``libomptarget.rtl.XXXX``)
 -------------------------------------------------------------------

openmp/libomptarget/plugins/cuda/src/rtl.cpp

 //===----RTLs/cuda/src/rtl.cpp - Target RTLs Implementation ------- C++ -*-===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 //
 // RTL for CUDA machine
 //
 //===----------------------------------------------------------------------===//
 
 #include <cassert>
 #include <cstddef>
 #include <cuda.h>
 #include <list>
 #include <memory>
 #include <mutex>
 #include <string>
 #include <unordered_map>
 #include <vector>
 
 #include "Debug.h"
int initDevice(const int DeviceId) {
 
     // Adjust teams to the env variables
     if (EnvTeamLimit > 0 && DeviceData[DeviceId].BlocksPerGrid > EnvTeamLimit) {
       DP("Capping max CUDA blocks per grid to OMP_TEAM_LIMIT=%d\n",
          EnvTeamLimit);
       DeviceData[DeviceId].BlocksPerGrid = EnvTeamLimit;
     }
 
+    size_t StackLimit;
+    size_t HeapLimit;
+    if (const char *EnvStr = getenv("LIBOMPTARGET_STACK_SIZE")) {
+      StackLimit = std::stol(EnvStr);
+      if (cuCtxSetLimit(CU_LIMIT_STACK_SIZE, StackLimit) != CUDA_SUCCESS)
+        return OFFLOAD_FAIL;
+    } else {
+      if (cuCtxGetLimit(&StackLimit, CU_LIMIT_STACK_SIZE) != CUDA_SUCCESS)
+        return OFFLOAD_FAIL;
+    }
+    if (const char *EnvStr = getenv("LIBOMPTARGET_HEAP_SIZE")) {
+      HeapLimit = std::stol(EnvStr);
+      if (cuCtxSetLimit(CU_LIMIT_MALLOC_HEAP_SIZE, HeapLimit) != CUDA_SUCCESS)
+        return OFFLOAD_FAIL;
+    } else {
+      if (cuCtxGetLimit(&HeapLimit, CU_LIMIT_MALLOC_HEAP_SIZE) != CUDA_SUCCESS)
+        return OFFLOAD_FAIL;
+    }
+
     INFO(OMP_INFOTYPE_PLUGIN_KERNEL, DeviceId,
          "Device supports up to %d CUDA blocks and %d threads with a "
          "warp size of %d\n",
          DeviceData[DeviceId].BlocksPerGrid,
          DeviceData[DeviceId].ThreadsPerBlock, DeviceData[DeviceId].WarpSize);
+    INFO(OMP_INFOTYPE_PLUGIN_KERNEL, DeviceId,
+         "Device heap size is %d Bytes, device stack size is %d Bytes per "
+         "thread\n",
+         (int)HeapLimit, (int)StackLimit);
 
     // Set default number of teams
     if (EnvNumTeams > 0) {
       DP("Default number of teams set according to environment %d\n",
          EnvNumTeams);
       DeviceData[DeviceId].NumTeams = EnvNumTeams;
     } else {
       DeviceData[DeviceId].NumTeams = DeviceRTLTy::DefaultNumTeams;