diff mbox

[hsa,merge,03/10] HSA libgomp plugin

Message ID 20160113173925.498574898@virgil.suse.cz
State New
Headers show

Commit Message

Martin Jambor Jan. 13, 2016, 5:39 p.m. UTC
Hi,

the patch below adds the HSA-specific plugin for libgomp.  The plugin
implements the interface mandated by libgomp and takes care of finding
any available HSA devices, finalizing HSAIL code and running it on
HSA-capable GPUs.

This patch is a re-post of
https://gcc.gnu.org/ml/gcc-patches/2015-12/msg00716.html with a number
of modifications requested by Jakub.

Thanks,

Martin


2016-01-13  Martin Jambor  <mjambor@suse.cz>
	    Martin Liska  <mliska@suse.cz>

	* plugin/plugin-hsa.c: New file.

Comments

Jakub Jelinek Jan. 14, 2016, 12:37 p.m. UTC | #1
On Wed, Jan 13, 2016 at 06:39:28PM +0100, Martin Jambor wrote:
> the patch below adds the HSA-specific plugin for libgomp.  The plugin
> implements the interface mandated by libgomp and takes care of finding
> any available HSA devices, finalizing HSAIL code and running it on
> HSA-capable GPUs.
> 
> This patch is a re-post of
> https://gcc.gnu.org/ml/gcc-patches/2015-12/msg00716.html with a number
> of modifications requested by Jakub.

Ok.

	Jakub
diff mbox

Patch

diff --git a/libgomp/plugin/plugin-hsa.c b/libgomp/plugin/plugin-hsa.c
new file mode 100644
index 0000000..d888493
--- /dev/null
+++ b/libgomp/plugin/plugin-hsa.c
@@ -0,0 +1,1493 @@ 
+/* Plugin for HSAIL execution.
+
+   Copyright (C) 2013-2016 Free Software Foundation, Inc.
+
+   Contributed by Martin Jambor <mjambor@suse.cz> and
+   Martin Liska <mliska@suse.cz>.
+
+   This file is part of the GNU Offloading and Multi Processing Library
+   (libgomp).
+
+   Libgomp is free software; you can redistribute it and/or modify it
+   under the terms of the GNU General Public License as published by
+   the Free Software Foundation; either version 3, or (at your option)
+   any later version.
+
+   Libgomp is distributed in the hope that it will be useful, but WITHOUT ANY
+   WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+   FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
+   more details.
+
+   Under Section 7 of GPL version 3, you are granted additional
+   permissions described in the GCC Runtime Library Exception, version
+   3.1, as published by the Free Software Foundation.
+
+   You should have received a copy of the GNU General Public License and
+   a copy of the GCC Runtime Library Exception along with this program;
+   see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
+   <http://www.gnu.org/licenses/>.  */
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <pthread.h>
+#include <hsa.h>
+#include <hsa_ext_finalize.h>
+#include <dlfcn.h>
+#include "libgomp-plugin.h"
+#include "gomp-constants.h"
+
+/* Keep the following GOMP prefixed structures in sync with respective parts of
+   the compiler.  */
+
+/* Structure describing the run-time and grid properties of an HSA kernel
+   lauch.  */
+
+struct GOMP_kernel_launch_attributes
+{
+  /* Number of dimensions the workload has.  Maximum number is 3.  */
+  uint32_t ndim;
+  /* Size of the grid in the three respective dimensions.  */
+  uint32_t gdims[3];
+  /* Size of work-groups in the respective dimensions.  */
+  uint32_t wdims[3];
+};
+
+/* Collection of information needed for a dispatch of a kernel from a
+   kernel.  */
+
+struct GOMP_hsa_kernel_dispatch
+{
+  /* Pointer to a command queue associated with a kernel dispatch agent.  */
+  void *queue;
+  /* Pointer to reserved memory for OMP data struct copying.  */
+  void *omp_data_memory;
+  /* Pointer to a memory space used for kernel arguments passing.  */
+  void *kernarg_address;
+  /* Kernel object.  */
+  uint64_t object;
+  /* Synchronization signal used for dispatch synchronization.  */
+  uint64_t signal;
+  /* Private segment size.  */
+  uint32_t private_segment_size;
+  /* Group segment size.  */
+  uint32_t group_segment_size;
+  /* Number of children kernel dispatches.  */
+  uint64_t kernel_dispatch_count;
+  /* Debug purpose argument.  */
+  uint64_t debug;
+  /* Levels-var ICV.  */
+  uint64_t omp_level;
+  /* Kernel dispatch structures created for children kernel dispatches.  */
+  struct GOMP_hsa_kernel_dispatch **children_dispatches;
+  /* Number of threads.  */
+  uint32_t omp_num_threads;
+};
+
+/* Part of the libgomp plugin interface.  Return the name of the accelerator,
+   which is "hsa".  */
+
+const char *
+GOMP_OFFLOAD_get_name (void)
+{
+  return "hsa";
+}
+
+/* Part of the libgomp plugin interface.  Return the specific capabilities the
+   HSA accelerator have.  */
+
+unsigned int
+GOMP_OFFLOAD_get_caps (void)
+{
+  return GOMP_OFFLOAD_CAP_SHARED_MEM | GOMP_OFFLOAD_CAP_OPENMP_400;
+}
+
+/* Part of the libgomp plugin interface.  Identify as HSA accelerator.  */
+
+int
+GOMP_OFFLOAD_get_type (void)
+{
+  return OFFLOAD_TARGET_TYPE_HSA;
+}
+
+/* Return the libgomp version number we're compatible with.  There is
+   no requirement for cross-version compatibility.  */
+
+unsigned
+GOMP_OFFLOAD_version (void)
+{
+  return GOMP_VERSION;
+}
+
+/* Flag to decide whether print to stderr information about what is going on.
+   Set in init_debug depending on environment variables.  */
+
+static bool debug;
+
+/* Flag to decide if the runtime should suppress a possible fallback to host
+   execution.  */
+
+static bool suppress_host_fallback;
+
+/* Initialize debug and suppress_host_fallback according to the environment.  */
+
+static void
+init_enviroment_variables (void)
+{
+  if (getenv ("HSA_DEBUG"))
+    debug = true;
+  else
+    debug = false;
+
+  if (getenv ("HSA_SUPPRESS_HOST_FALLBACK"))
+    suppress_host_fallback = true;
+  else
+    suppress_host_fallback = false;
+}
+
+/* Print a logging message with PREFIX to stderr if HSA_DEBUG value
+   is set to true.  */
+
+#define HSA_LOG(prefix, ...) \
+  do \
+  { \
+    if (debug) \
+      { \
+	fprintf (stderr, prefix); \
+	fprintf (stderr, __VA_ARGS__); \
+      } \
+  } \
+  while (false);
+
+/* Print a debugging message to stderr.  */
+
+#define HSA_DEBUG(...) HSA_LOG ("HSA debug: ", __VA_ARGS__)
+
+/* Print a warning message to stderr.  */
+
+#define HSA_WARNING(...) HSA_LOG ("HSA warning: ", __VA_ARGS__)
+
+/* Print HSA warning STR with an HSA STATUS code.  */
+
+static void
+hsa_warn (const char *str, hsa_status_t status)
+{
+  if (!debug)
+    return;
+
+  const char *hsa_error;
+  hsa_status_string (status, &hsa_error);
+
+  fprintf (stderr, "HSA warning: %s\nRuntime message: %s", str, hsa_error);
+}
+
+/* Report a fatal error STR together with the HSA error corresponding to STATUS
+   and terminate execution of the current process.  */
+
+static void
+hsa_fatal (const char *str, hsa_status_t status)
+{
+  const char *hsa_error;
+  hsa_status_string (status, &hsa_error);
+  GOMP_PLUGIN_fatal ("HSA fatal error: %s\nRuntime message: %s", str,
+		     hsa_error);
+}
+
+struct hsa_kernel_description
+{
+  const char *name;
+  unsigned omp_data_size;
+  bool gridified_kernel_p;
+  unsigned kernel_dependencies_count;
+  const char **kernel_dependencies;
+};
+
+struct global_var_info
+{
+  const char *name;
+  void *address;
+};
+
+/* Data passed by the static initializer of a compilation unit containing BRIG
+   to GOMP_offload_register.  */
+
+struct brig_image_desc
+{
+  hsa_ext_module_t brig_module;
+  const unsigned kernel_count;
+  struct hsa_kernel_description *kernel_infos;
+  const unsigned global_variable_count;
+  struct global_var_info *global_variables;
+};
+
+struct agent_info;
+
+/* Information required to identify, finalize and run any given kernel.  */
+
+struct kernel_info
+{
+  /* Name of the kernel, required to locate it within the brig module.  */
+  const char *name;
+  /* Size of memory space for OMP data.  */
+  unsigned omp_data_size;
+  /* The specific agent the kernel has been or will be finalized for and run
+     on.  */
+  struct agent_info *agent;
+  /* The specific module where the kernel takes place.  */
+  struct module_info *module;
+  /* Mutex enforcing that at most once thread ever initializes a kernel for
+     use.  A thread should have locked agent->modules_rwlock for reading before
+     acquiring it.  */
+  pthread_mutex_t init_mutex;
+  /* Flag indicating whether the kernel has been initialized and all fields
+     below it contain valid data.  */
+  bool initialized;
+  /* Flag indicating that the kernel has a problem that blocks an execution.  */
+  bool initialization_failed;
+  /* The object to be put into the dispatch queue.  */
+  uint64_t object;
+  /* Required size of kernel arguments.  */
+  uint32_t kernarg_segment_size;
+  /* Required size of group segment.  */
+  uint32_t group_segment_size;
+  /* Required size of private segment.  */
+  uint32_t private_segment_size;
+  /* List of all kernel dependencies.  */
+  const char **dependencies;
+  /* Number of dependencies.  */
+  unsigned dependencies_count;
+  /* Maximum OMP data size necessary for kernel from kernel dispatches.  */
+  unsigned max_omp_data_size;
+  /* True if the kernel is gridified.  */
+  bool gridified_kernel_p;
+};
+
+/* Information about a particular brig module, its image and kernels.  */
+
+struct module_info
+{
+  /* The next and previous module in the linked list of modules of an agent.  */
+  struct module_info *next, *prev;
+  /* The description with which the program has registered the image.  */
+  struct brig_image_desc *image_desc;
+
+  /* Number of kernels in this module.  */
+  int kernel_count;
+  /* An array of kernel_info structures describing each kernel in this
+     module.  */
+  struct kernel_info kernels[];
+};
+
+/* Information about shared brig library.  */
+
+struct brig_library_info
+{
+  char *file_name;
+  hsa_ext_module_t image;
+};
+
+/* Description of an HSA GPU agent and the program associated with it.  */
+
+struct agent_info
+{
+  /* The HSA ID of the agent.  Assigned when hsa_context is initialized.  */
+  hsa_agent_t id;
+  /* Whether the agent has been initialized.  The fields below are usable only
+     if it has been.  */
+  bool initialized;
+  /* The HSA ISA of this agent.  */
+  hsa_isa_t isa;
+  /* Command queue of the agent.  */
+  hsa_queue_t *command_q;
+  /* Kernel from kernel dispatch command queue.  */
+  hsa_queue_t *kernel_dispatch_command_q;
+  /* The HSA memory region from which to allocate kernel arguments.  */
+  hsa_region_t kernarg_region;
+
+  /* Read-write lock that protects kernels which are running or about to be run
+     from interference with loading and unloading of images.  Needs to be
+     locked for reading while a kernel is being run, and for writing if the
+     list of modules is manipulated (and thus the HSA program invalidated).  */
+  pthread_rwlock_t modules_rwlock;
+  /* The first module in a linked list of modules associated with this
+     kernel.  */
+  struct module_info *first_module;
+
+  /* Mutex enforcing that only one thread will finalize the HSA program.  A
+     thread should have locked agent->modules_rwlock for reading before
+     acquiring it.  */
+  pthread_mutex_t prog_mutex;
+  /* Flag whether the HSA program that consists of all the modules has been
+     finalized.  */
+  bool prog_finalized;
+  /* Flag whether the program was finalized but with a failure.  */
+  bool prog_finalized_error;
+  /* HSA executable - the finalized program that is used to locate kernels.  */
+  hsa_executable_t executable;
+  /* List of BRIG libraries.  */
+  struct brig_library_info **brig_libraries;
+  /* Number of loaded shared BRIG libraries.  */
+  unsigned brig_libraries_count;
+};
+
+/* Information about the whole HSA environment and all of its agents.  */
+
+struct hsa_context_info
+{
+  /* Whether the structure has been initialized.  */
+  bool initialized;
+  /* Number of usable GPU HSA agents in the system.  */
+  int agent_count;
+  /* Array of agent_info structures describing the individual HSA agents.  */
+  struct agent_info *agents;
+};
+
+/* Information about the whole HSA environment and all of its agents.  */
+
+static struct hsa_context_info hsa_context;
+
+/* Find kernel for an AGENT by name provided in KERNEL_NAME.  */
+
+static struct kernel_info *
+get_kernel_for_agent (struct agent_info *agent, const char *kernel_name)
+{
+  struct module_info *module = agent->first_module;
+
+  while (module)
+    {
+      for (unsigned i = 0; i < module->kernel_count; i++)
+	if (strcmp (module->kernels[i].name, kernel_name) == 0)
+	  return &module->kernels[i];
+
+      module = module->next;
+    }
+
+  return NULL;
+}
+
+/* Return true if the agent is a GPU and acceptable of concurrent submissions
+   from different threads.  */
+
+static bool
+suitable_hsa_agent_p (hsa_agent_t agent)
+{
+  hsa_device_type_t device_type;
+  hsa_status_t status
+    = hsa_agent_get_info (agent, HSA_AGENT_INFO_DEVICE, &device_type);
+  if (status != HSA_STATUS_SUCCESS || device_type != HSA_DEVICE_TYPE_GPU)
+    return false;
+
+  uint32_t features = 0;
+  status = hsa_agent_get_info (agent, HSA_AGENT_INFO_FEATURE, &features);
+  if (status != HSA_STATUS_SUCCESS
+      || !(features & HSA_AGENT_FEATURE_KERNEL_DISPATCH))
+    return false;
+  hsa_queue_type_t queue_type;
+  status = hsa_agent_get_info (agent, HSA_AGENT_INFO_QUEUE_TYPE, &queue_type);
+  if (status != HSA_STATUS_SUCCESS
+      || (queue_type != HSA_QUEUE_TYPE_MULTI))
+    return false;
+
+  return true;
+}
+
+/* Callback of hsa_iterate_agents, if AGENT is a GPU device, increment
+   agent_count in hsa_context.  */
+
+static hsa_status_t
+count_gpu_agents (hsa_agent_t agent, void *data __attribute__ ((unused)))
+{
+  if (suitable_hsa_agent_p (agent))
+    hsa_context.agent_count++;
+  return HSA_STATUS_SUCCESS;
+}
+
+/* Callback of hsa_iterate_agents, if AGENT is a GPU device, assign the agent
+   id to the describing structure in the hsa context.  The index of the
+   structure is pointed to by DATA, increment it afterwards.  */
+
+static hsa_status_t
+assign_agent_ids (hsa_agent_t agent, void *data)
+{
+  if (suitable_hsa_agent_p (agent))
+    {
+      int *agent_index = (int *) data;
+      hsa_context.agents[*agent_index].id = agent;
+      ++*agent_index;
+    }
+  return HSA_STATUS_SUCCESS;
+}
+
+/* Initialize hsa_context if it has not already been done.  */
+
+static void
+init_hsa_context (void)
+{
+  hsa_status_t status;
+  int agent_index = 0;
+
+  if (hsa_context.initialized)
+    return;
+  init_enviroment_variables ();
+  status = hsa_init ();
+  if (status != HSA_STATUS_SUCCESS)
+    hsa_fatal ("Run-time could not be initialized", status);
+  HSA_DEBUG ("HSA run-time initialized\n");
+  status = hsa_iterate_agents (count_gpu_agents, NULL);
+  if (status != HSA_STATUS_SUCCESS)
+    hsa_fatal ("HSA GPU devices could not be enumerated", status);
+  HSA_DEBUG ("There are %i HSA GPU devices.\n", hsa_context.agent_count);
+
+  hsa_context.agents
+    = GOMP_PLUGIN_malloc_cleared (hsa_context.agent_count
+				  * sizeof (struct agent_info));
+  status = hsa_iterate_agents (assign_agent_ids, &agent_index);
+  if (agent_index != hsa_context.agent_count)
+    GOMP_PLUGIN_fatal ("Failed to assign IDs to all HSA agents");
+  hsa_context.initialized = true;
+}
+
+/* Callback of dispatch queues to report errors.  */
+
+static void
+queue_callback (hsa_status_t status,
+		hsa_queue_t *queue __attribute__ ((unused)),
+		void *data __attribute__ ((unused)))
+{
+  hsa_fatal ("Asynchronous queue error", status);
+}
+
+/* Callback of hsa_agent_iterate_regions.  Determine if a memory REGION can be
+   used for kernarg allocations and if so write it to the memory pointed to by
+   DATA and break the query.  */
+
+static hsa_status_t
+get_kernarg_memory_region (hsa_region_t region, void *data)
+{
+  hsa_status_t status;
+  hsa_region_segment_t segment;
+
+  status = hsa_region_get_info (region, HSA_REGION_INFO_SEGMENT, &segment);
+  if (status != HSA_STATUS_SUCCESS)
+    return status;
+  if (segment != HSA_REGION_SEGMENT_GLOBAL)
+    return HSA_STATUS_SUCCESS;
+
+  uint32_t flags;
+  status = hsa_region_get_info (region, HSA_REGION_INFO_GLOBAL_FLAGS, &flags);
+  if (status != HSA_STATUS_SUCCESS)
+    return status;
+  if (flags & HSA_REGION_GLOBAL_FLAG_KERNARG)
+    {
+      hsa_region_t *ret = (hsa_region_t *) data;
+      *ret = region;
+      return HSA_STATUS_INFO_BREAK;
+    }
+  return HSA_STATUS_SUCCESS;
+}
+
+/* Part of the libgomp plugin interface.  Return the number of HSA devices on
+   the system.  */
+
+int
+GOMP_OFFLOAD_get_num_devices (void)
+{
+  init_hsa_context ();
+  return hsa_context.agent_count;
+}
+
+/* Part of the libgomp plugin interface.  Initialize agent number N so that it
+   can be used for computation.  */
+
+void
+GOMP_OFFLOAD_init_device (int n)
+{
+  init_hsa_context ();
+  if (n >= hsa_context.agent_count)
+    GOMP_PLUGIN_fatal ("Request to initialize non-existing HSA device %i", n);
+  struct agent_info *agent = &hsa_context.agents[n];
+
+  if (agent->initialized)
+    return;
+
+  if (pthread_rwlock_init (&agent->modules_rwlock, NULL))
+    GOMP_PLUGIN_fatal ("Failed to initialize an HSA agent rwlock");
+  if (pthread_mutex_init (&agent->prog_mutex, NULL))
+    GOMP_PLUGIN_fatal ("Failed to initialize an HSA agent program mutex");
+
+  uint32_t queue_size;
+  hsa_status_t status;
+  status = hsa_agent_get_info (agent->id, HSA_AGENT_INFO_QUEUE_MAX_SIZE,
+			       &queue_size);
+  if (status != HSA_STATUS_SUCCESS)
+    hsa_fatal ("Error requesting maximum queue size of the HSA agent", status);
+  status = hsa_agent_get_info (agent->id, HSA_AGENT_INFO_ISA, &agent->isa);
+  if (status != HSA_STATUS_SUCCESS)
+    hsa_fatal ("Error querying the ISA of the agent", status);
+  status = hsa_queue_create (agent->id, queue_size, HSA_QUEUE_TYPE_MULTI,
+			     queue_callback, NULL, UINT32_MAX, UINT32_MAX,
+			     &agent->command_q);
+  if (status != HSA_STATUS_SUCCESS)
+    hsa_fatal ("Error creating command queue", status);
+
+  status = hsa_queue_create (agent->id, queue_size, HSA_QUEUE_TYPE_MULTI,
+			     queue_callback, NULL, UINT32_MAX, UINT32_MAX,
+			     &agent->kernel_dispatch_command_q);
+  if (status != HSA_STATUS_SUCCESS)
+    hsa_fatal ("Error creating kernel dispatch command queue", status);
+
+  agent->kernarg_region.handle = (uint64_t) -1;
+  status = hsa_agent_iterate_regions (agent->id, get_kernarg_memory_region,
+				      &agent->kernarg_region);
+  if (agent->kernarg_region.handle == (uint64_t) -1)
+    GOMP_PLUGIN_fatal ("Could not find suitable memory region for kernel "
+		       "arguments");
+  HSA_DEBUG ("HSA agent initialized, queue has id %llu\n",
+	     (long long unsigned) agent->command_q->id);
+  HSA_DEBUG ("HSA agent initialized, kernel dispatch queue has id %llu\n",
+	     (long long unsigned) agent->kernel_dispatch_command_q->id);
+  agent->initialized = true;
+}
+
+/* Verify that hsa_context has already been initialized and return the
+   agent_info structure describing device number N.  */
+
+static struct agent_info *
+get_agent_info (int n)
+{
+  if (!hsa_context.initialized)
+    GOMP_PLUGIN_fatal ("Attempt to use uninitialized HSA context.");
+  if (n >= hsa_context.agent_count)
+    GOMP_PLUGIN_fatal ("Request to operate on anon-existing HSA device %i", n);
+  if (!hsa_context.agents[n].initialized)
+    GOMP_PLUGIN_fatal ("Attempt to use an uninitialized HSA agent.");
+  return &hsa_context.agents[n];
+}
+
+/* Insert MODULE to the linked list of modules of AGENT.  */
+
+static void
+add_module_to_agent (struct agent_info *agent, struct module_info *module)
+{
+  if (agent->first_module)
+    agent->first_module->prev = module;
+  module->next = agent->first_module;
+  module->prev = NULL;
+  agent->first_module = module;
+}
+
+/* Remove MODULE from the linked list of modules of AGENT.  */
+
+static void
+remove_module_from_agent (struct agent_info *agent, struct module_info *module)
+{
+  if (agent->first_module == module)
+    agent->first_module = module->next;
+  if (module->prev)
+    module->prev->next = module->next;
+  if (module->next)
+    module->next->prev = module->prev;
+}
+
+/* Free the HSA program in agent and everything associated with it and set
+   agent->prog_finalized and the initialized flags of all kernels to false.  */
+
+static void
+destroy_hsa_program (struct agent_info *agent)
+{
+  if (!agent->prog_finalized || agent->prog_finalized_error)
+    return;
+
+  hsa_status_t status;
+
+  HSA_DEBUG ("Destroying the current HSA program.\n");
+
+  status = hsa_executable_destroy (agent->executable);
+  if (status != HSA_STATUS_SUCCESS)
+    hsa_fatal ("Could not destroy HSA executable", status);
+
+  struct module_info *module;
+  for (module = agent->first_module; module; module = module->next)
+    {
+      int i;
+      for (i = 0; i < module->kernel_count; i++)
+	module->kernels[i].initialized = false;
+    }
+  agent->prog_finalized = false;
+}
+
+/* Part of the libgomp plugin interface.  Load BRIG module described by struct
+   brig_image_desc in TARGET_DATA and return references to kernel descriptors
+   in TARGET_TABLE.  */
+
+int
+GOMP_OFFLOAD_load_image (int ord, unsigned version, void *target_data,
+			 struct addr_pair **target_table)
+{
+  if (GOMP_VERSION_DEV (version) > GOMP_VERSION_HSA)
+    GOMP_PLUGIN_fatal ("Offload data incompatible with HSA plugin"
+		       " (expected %u, received %u)",
+		       GOMP_VERSION_HSA, GOMP_VERSION_DEV (version));
+
+  struct brig_image_desc *image_desc = (struct brig_image_desc *) target_data;
+  struct agent_info *agent;
+  struct addr_pair *pair;
+  struct module_info *module;
+  struct kernel_info *kernel;
+  int kernel_count = image_desc->kernel_count;
+
+  agent = get_agent_info (ord);
+  if (pthread_rwlock_wrlock (&agent->modules_rwlock))
+    GOMP_PLUGIN_fatal ("Unable to write-lock an HSA agent rwlock");
+  if (agent->prog_finalized)
+    destroy_hsa_program (agent);
+
+  HSA_DEBUG ("Encountered %d kernels in an image\n", kernel_count);
+  pair = GOMP_PLUGIN_malloc (kernel_count * sizeof (struct addr_pair));
+  *target_table = pair;
+  module = (struct module_info *)
+    GOMP_PLUGIN_malloc_cleared (sizeof (struct module_info)
+				+ kernel_count * sizeof (struct kernel_info));
+  module->image_desc = image_desc;
+  module->kernel_count = kernel_count;
+
+  kernel = &module->kernels[0];
+
+  /* Allocate memory for kernel dependencies.  */
+  for (unsigned i = 0; i < kernel_count; i++)
+    {
+      pair->start = (uintptr_t) kernel;
+      pair->end = (uintptr_t) (kernel + 1);
+
+      struct hsa_kernel_description *d = &image_desc->kernel_infos[i];
+      kernel->agent = agent;
+      kernel->module = module;
+      kernel->name = d->name;
+      kernel->omp_data_size = d->omp_data_size;
+      kernel->gridified_kernel_p = d->gridified_kernel_p;
+      kernel->dependencies_count = d->kernel_dependencies_count;
+      kernel->dependencies = d->kernel_dependencies;
+      if (pthread_mutex_init (&kernel->init_mutex, NULL))
+	GOMP_PLUGIN_fatal ("Failed to initialize an HSA kernel mutex");
+
+      kernel++;
+      pair++;
+    }
+
+  add_module_to_agent (agent, module);
+  if (pthread_rwlock_unlock (&agent->modules_rwlock))
+    GOMP_PLUGIN_fatal ("Unable to unlock an HSA agent rwlock");
+  return kernel_count;
+}
+
+/* Add a shared BRIG library from a FILE_NAME to an AGENT.  */
+
+static struct brig_library_info *
+add_shared_library (const char *file_name, struct agent_info *agent)
+{
+  struct brig_library_info *library = NULL;
+
+  void *f = dlopen (file_name, RTLD_NOW);
+  void *start = dlsym (f, "__brig_start");
+  void *end = dlsym (f, "__brig_end");
+
+  if (start == NULL || end == NULL)
+    return NULL;
+
+  unsigned size = end - start;
+  char *buf = (char *) GOMP_PLUGIN_malloc (size);
+  memcpy (buf, start, size);
+
+  library = GOMP_PLUGIN_malloc (sizeof (struct agent_info));
+  library->file_name = (char *) GOMP_PLUGIN_malloc
+    ((strlen (file_name) + 1));
+  strcpy (library->file_name, file_name);
+  library->image = (hsa_ext_module_t) buf;
+
+  return library;
+}
+
+/* Release memory used for BRIG shared libraries that correspond
+   to an AGENT.  */
+
+static void
+release_agent_shared_libraries (struct agent_info *agent)
+{
+  for (unsigned i = 0; i < agent->brig_libraries_count; i++)
+    if (agent->brig_libraries[i])
+      {
+	free (agent->brig_libraries[i]->file_name);
+	free (agent->brig_libraries[i]->image);
+	free (agent->brig_libraries[i]);
+      }
+
+  free (agent->brig_libraries);
+}
+
+/* Create and finalize the program consisting of all loaded modules.  */
+
+static void
+create_and_finalize_hsa_program (struct agent_info *agent)
+{
+  hsa_status_t status;
+  hsa_ext_program_t prog_handle;
+  int mi = 0;
+
+  if (pthread_mutex_lock (&agent->prog_mutex))
+    GOMP_PLUGIN_fatal ("Could not lock an HSA agent program mutex");
+  if (agent->prog_finalized)
+    goto final;
+
+  status = hsa_ext_program_create (HSA_MACHINE_MODEL_LARGE, HSA_PROFILE_FULL,
+				   HSA_DEFAULT_FLOAT_ROUNDING_MODE_DEFAULT,
+				   NULL, &prog_handle);
+  if (status != HSA_STATUS_SUCCESS)
+    hsa_fatal ("Could not create an HSA program", status);
+
+  HSA_DEBUG ("Created a finalized program\n");
+
+  struct module_info *module = agent->first_module;
+  while (module)
+    {
+      status = hsa_ext_program_add_module (prog_handle,
+					   module->image_desc->brig_module);
+      if (status != HSA_STATUS_SUCCESS)
+	hsa_fatal ("Could not add a module to the HSA program", status);
+      module = module->next;
+      mi++;
+    }
+
+  /* Load all shared libraries.  */
+  const char *libraries[] = { "libhsamath.so", "libhsastd.so" };
+  const unsigned libraries_count = sizeof (libraries) / sizeof (const char *);
+
+  agent->brig_libraries_count = libraries_count;
+  agent->brig_libraries = GOMP_PLUGIN_malloc_cleared
+    (sizeof (struct brig_library_info) * libraries_count);
+
+  for (unsigned i = 0; i < libraries_count; i++)
+    {
+      struct brig_library_info *library = add_shared_library (libraries[i],
+							      agent);
+      if (library == NULL)
+	{
+	  HSA_WARNING ("Could not open a shared BRIG library: %s\n",
+		       libraries[i]);
+	  continue;
+	}
+
+      status = hsa_ext_program_add_module (prog_handle, library->image);
+      if (status != HSA_STATUS_SUCCESS)
+	hsa_warn ("Could not add a shared BRIG library the HSA program",
+		  status);
+      else
+	HSA_DEBUG ("a shared BRIG library has been added to a program: %s\n",
+		   libraries[i]);
+    }
+
+  hsa_ext_control_directives_t control_directives;
+  memset (&control_directives, 0, sizeof (control_directives));
+  hsa_code_object_t code_object;
+  status = hsa_ext_program_finalize (prog_handle, agent->isa,
+				     HSA_EXT_FINALIZER_CALL_CONVENTION_AUTO,
+				     control_directives, "",
+				     HSA_CODE_OBJECT_TYPE_PROGRAM,
+				     &code_object);
+  if (status != HSA_STATUS_SUCCESS)
+    {
+      hsa_warn ("Finalization of the HSA program failed", status);
+      goto failure;
+    }
+
+  HSA_DEBUG ("Finalization done\n");
+  hsa_ext_program_destroy (prog_handle);
+
+  status
+    = hsa_executable_create (HSA_PROFILE_FULL, HSA_EXECUTABLE_STATE_UNFROZEN,
+			     "", &agent->executable);
+  if (status != HSA_STATUS_SUCCESS)
+    hsa_fatal ("Could not create HSA executable", status);
+
+  module = agent->first_module;
+  while (module)
+    {
+      /* Initialize all global variables declared in the module.  */
+      for (unsigned i = 0; i < module->image_desc->global_variable_count; i++)
+	{
+	  struct global_var_info *var;
+	  var = &module->image_desc->global_variables[i];
+	  status
+	    = hsa_executable_global_variable_define (agent->executable,
+						     var->name, var->address);
+
+	  HSA_DEBUG ("Defining global variable: %s, address: %p\n", var->name,
+		     var->address);
+
+	  if (status != HSA_STATUS_SUCCESS)
+	    hsa_fatal ("Could not define a global variable in the HSA program",
+		       status);
+	}
+
+      module = module->next;
+    }
+
+  status = hsa_executable_load_code_object (agent->executable, agent->id,
+					    code_object, "");
+  if (status != HSA_STATUS_SUCCESS)
+    hsa_fatal ("Could not add a code object to the HSA executable", status);
+  status = hsa_executable_freeze (agent->executable, "");
+  if (status != HSA_STATUS_SUCCESS)
+    hsa_fatal ("Could not freeze the HSA executable", status);
+
+  HSA_DEBUG ("Froze HSA executable with the finalized code object\n");
+
+  /* If all goes good, jump to final.  */
+  goto final;
+
+failure:
+  agent->prog_finalized_error = true;
+
+final:
+  agent->prog_finalized = true;
+
+  if (pthread_mutex_unlock (&agent->prog_mutex))
+    GOMP_PLUGIN_fatal ("Could not unlock an HSA agent program mutex");
+}
+
+/* Create kernel dispatch data structure for given KERNEL.  */
+
+static struct GOMP_hsa_kernel_dispatch *
+create_single_kernel_dispatch (struct kernel_info *kernel,
+			       unsigned omp_data_size)
+{
+  struct agent_info *agent = kernel->agent;
+  struct GOMP_hsa_kernel_dispatch *shadow
+    = GOMP_PLUGIN_malloc_cleared (sizeof (struct GOMP_hsa_kernel_dispatch));
+
+  shadow->queue = agent->command_q;
+  shadow->omp_data_memory
+    = omp_data_size > 0 ? GOMP_PLUGIN_malloc (omp_data_size) : NULL;
+  unsigned dispatch_count = kernel->dependencies_count;
+  shadow->kernel_dispatch_count = dispatch_count;
+
+  shadow->children_dispatches
+    = GOMP_PLUGIN_malloc (dispatch_count * sizeof (shadow));
+
+  shadow->object = kernel->object;
+
+  hsa_signal_t sync_signal;
+  hsa_status_t status = hsa_signal_create (1, 0, NULL, &sync_signal);
+  if (status != HSA_STATUS_SUCCESS)
+    hsa_fatal ("Error creating the HSA sync signal", status);
+
+  shadow->signal = sync_signal.handle;
+  shadow->private_segment_size = kernel->private_segment_size;
+  shadow->group_segment_size = kernel->group_segment_size;
+
+  status
+    = hsa_memory_allocate (agent->kernarg_region, kernel->kernarg_segment_size,
+			   &shadow->kernarg_address);
+  if (status != HSA_STATUS_SUCCESS)
+    hsa_fatal ("Could not allocate memory for HSA kernel arguments", status);
+
+  return shadow;
+}
+
+/* Release data structure created for a kernel dispatch in SHADOW argument.  */
+
+static void
+release_kernel_dispatch (struct GOMP_hsa_kernel_dispatch *shadow)
+{
+  HSA_DEBUG ("Released kernel dispatch: %p has value: %lu (%p)\n", shadow,
+	     shadow->debug, (void *) shadow->debug);
+
+  hsa_memory_free (shadow->kernarg_address);
+
+  hsa_signal_t s;
+  s.handle = shadow->signal;
+  hsa_signal_destroy (s);
+
+  free (shadow->omp_data_memory);
+
+  for (unsigned i = 0; i < shadow->kernel_dispatch_count; i++)
+    release_kernel_dispatch (shadow->children_dispatches[i]);
+
+  free (shadow->children_dispatches);
+  free (shadow);
+}
+
+/* Initialize a KERNEL without its dependencies.  MAX_OMP_DATA_SIZE is used
+   to calculate maximum necessary memory for OMP data allocation.  */
+
+static void
+init_single_kernel (struct kernel_info *kernel, unsigned *max_omp_data_size)
+{
+  hsa_status_t status;
+  struct agent_info *agent = kernel->agent;
+  hsa_executable_symbol_t kernel_symbol;
+  status = hsa_executable_get_symbol (agent->executable, NULL, kernel->name,
+				      agent->id, 0, &kernel_symbol);
+  if (status != HSA_STATUS_SUCCESS)
+    {
+      hsa_warn ("Could not find symbol for kernel in the code object", status);
+      goto failure;
+    }
+  HSA_DEBUG ("Located kernel %s\n", kernel->name);
+  status
+    = hsa_executable_symbol_get_info (kernel_symbol,
+				      HSA_EXECUTABLE_SYMBOL_INFO_KERNEL_OBJECT,
+				      &kernel->object);
+  if (status != HSA_STATUS_SUCCESS)
+    hsa_fatal ("Could not extract a kernel object from its symbol", status);
+  status = hsa_executable_symbol_get_info
+    (kernel_symbol, HSA_EXECUTABLE_SYMBOL_INFO_KERNEL_KERNARG_SEGMENT_SIZE,
+     &kernel->kernarg_segment_size);
+  if (status != HSA_STATUS_SUCCESS)
+    hsa_fatal ("Could not get info about kernel argument size", status);
+  status = hsa_executable_symbol_get_info
+    (kernel_symbol, HSA_EXECUTABLE_SYMBOL_INFO_KERNEL_GROUP_SEGMENT_SIZE,
+     &kernel->group_segment_size);
+  if (status != HSA_STATUS_SUCCESS)
+    hsa_fatal ("Could not get info about kernel group segment size", status);
+  status = hsa_executable_symbol_get_info
+    (kernel_symbol, HSA_EXECUTABLE_SYMBOL_INFO_KERNEL_PRIVATE_SEGMENT_SIZE,
+     &kernel->private_segment_size);
+  if (status != HSA_STATUS_SUCCESS)
+    hsa_fatal ("Could not get info about kernel private segment size",
+	       status);
+
+  HSA_DEBUG ("Kernel structure for %s fully initialized with "
+	     "following segment sizes: \n", kernel->name);
+  HSA_DEBUG ("  group_segment_size: %u\n",
+	     (unsigned) kernel->group_segment_size);
+  HSA_DEBUG ("  private_segment_size: %u\n",
+	     (unsigned) kernel->private_segment_size);
+  HSA_DEBUG ("  kernarg_segment_size: %u\n",
+	     (unsigned) kernel->kernarg_segment_size);
+  HSA_DEBUG ("  omp_data_size: %u\n", kernel->omp_data_size);
+  HSA_DEBUG ("  gridified_kernel_p: %u\n", kernel->gridified_kernel_p);
+
+  if (kernel->omp_data_size > *max_omp_data_size)
+    *max_omp_data_size = kernel->omp_data_size;
+
+  for (unsigned i = 0; i < kernel->dependencies_count; i++)
+    {
+      struct kernel_info *dependency
+	= get_kernel_for_agent (agent, kernel->dependencies[i]);
+
+      if (dependency == NULL)
+	{
+	  HSA_DEBUG ("Could not find a dependency for a kernel: %s, "
+		     "dependency name: %s\n", kernel->name,
+		     kernel->dependencies[i]);
+	  goto failure;
+	}
+
+      if (dependency->dependencies_count > 0)
+	{
+	  HSA_DEBUG ("HSA does not allow kernel dispatching code with "
+		     "a depth bigger than one\n")
+	  goto failure;
+	}
+
+      init_single_kernel (dependency, max_omp_data_size);
+    }
+
+  return;
+
+failure:
+  kernel->initialization_failed = true;
+}
+
+/* Indent stream F by INDENT spaces.  */
+
+static void
+indent_stream (FILE *f, unsigned indent)
+{
+  fprintf (f, "%*s", indent, "");
+}
+
+/* Dump kernel DISPATCH data structure and indent it by INDENT spaces.  */
+
+static void
+print_kernel_dispatch (struct GOMP_hsa_kernel_dispatch *dispatch, unsigned indent)
+{
+  indent_stream (stderr, indent);
+  fprintf (stderr, "this: %p\n", dispatch);
+  indent_stream (stderr, indent);
+  fprintf (stderr, "queue: %p\n", dispatch->queue);
+  indent_stream (stderr, indent);
+  fprintf (stderr, "omp_data_memory: %p\n", dispatch->omp_data_memory);
+  indent_stream (stderr, indent);
+  fprintf (stderr, "kernarg_address: %p\n", dispatch->kernarg_address);
+  indent_stream (stderr, indent);
+  fprintf (stderr, "object: %lu\n", dispatch->object);
+  indent_stream (stderr, indent);
+  fprintf (stderr, "signal: %lu\n", dispatch->signal);
+  indent_stream (stderr, indent);
+  fprintf (stderr, "private_segment_size: %u\n",
+	   dispatch->private_segment_size);
+  indent_stream (stderr, indent);
+  fprintf (stderr, "group_segment_size: %u\n",
+	   dispatch->group_segment_size);
+  indent_stream (stderr, indent);
+  fprintf (stderr, "children dispatches: %lu\n",
+	   dispatch->kernel_dispatch_count);
+  indent_stream (stderr, indent);
+  fprintf (stderr, "omp_num_threads: %u\n",
+	   dispatch->omp_num_threads);
+  fprintf (stderr, "\n");
+
+  for (unsigned i = 0; i < dispatch->kernel_dispatch_count; i++)
+    print_kernel_dispatch (dispatch->children_dispatches[i], indent + 2);
+}
+
+/* Create kernel dispatch data structure for a KERNEL and all its
+   dependencies.  */
+
+static struct GOMP_hsa_kernel_dispatch *
+create_kernel_dispatch (struct kernel_info *kernel, unsigned omp_data_size)
+{
+  struct GOMP_hsa_kernel_dispatch *shadow
+    = create_single_kernel_dispatch (kernel, omp_data_size);
+  shadow->omp_num_threads = 64;
+  shadow->debug = 0;
+  shadow->omp_level = kernel->gridified_kernel_p ? 1 : 0;
+
+  /* Create kernel dispatch data structures.  We do not allow to have
+     a kernel dispatch with depth bigger than one.  */
+  for (unsigned i = 0; i < kernel->dependencies_count; i++)
+    {
+      struct kernel_info *dependency
+	= get_kernel_for_agent (kernel->agent, kernel->dependencies[i]);
+      shadow->children_dispatches[i]
+	= create_single_kernel_dispatch (dependency, omp_data_size);
+      shadow->children_dispatches[i]->queue
+	= kernel->agent->kernel_dispatch_command_q;
+      shadow->children_dispatches[i]->omp_level = 1;
+    }
+
+  return shadow;
+}
+
+/* Do all the work that is necessary before running KERNEL for the first time.
+   The function assumes the program has been created, finalized and frozen by
+   create_and_finalize_hsa_program.  */
+
+static void
+init_kernel (struct kernel_info *kernel)
+{
+  if (pthread_mutex_lock (&kernel->init_mutex))
+    GOMP_PLUGIN_fatal ("Could not lock an HSA kernel initialization mutex");
+  if (kernel->initialized)
+    {
+      if (pthread_mutex_unlock (&kernel->init_mutex))
+	GOMP_PLUGIN_fatal ("Could not unlock an HSA kernel initialization "
+			   "mutex");
+
+      return;
+    }
+
+  /* Precomputed maximum size of OMP data necessary for a kernel from kernel
+     dispatch operation.  */
+  init_single_kernel (kernel, &kernel->max_omp_data_size);
+
+  if (!kernel->initialization_failed)
+    HSA_DEBUG ("\n");
+
+  kernel->initialized = true;
+  if (pthread_mutex_unlock (&kernel->init_mutex))
+    GOMP_PLUGIN_fatal ("Could not unlock an HSA kernel initialization "
+		       "mutex");
+}
+
+/* Parse the target attributes INPUT provided by the compiler and return true
+   if we should run anything all.  If INPUT is NULL, fill DEF with default
+   values, then store INPUT or DEF into *RESULT.  */
+
+static bool
+parse_target_attributes (void **input,
+			 struct GOMP_kernel_launch_attributes *def,
+			 struct GOMP_kernel_launch_attributes **result)
+{
+  if (!input)
+    GOMP_PLUGIN_fatal ("No target arguments provided");
+
+  bool attrs_found = false;
+  while (*input)
+    {
+      uintptr_t id = (uintptr_t) *input;
+      if ((id & GOMP_TARGET_ARG_DEVICE_MASK) == GOMP_DEVICE_HSA
+	  && ((id & GOMP_TARGET_ARG_ID_MASK)
+	      == GOMP_TARGET_ARG_HSA_KERNEL_ATTRIBUTES))
+	{
+	  input++;
+	  attrs_found = true;
+	  break;
+	}
+
+      if (id & GOMP_TARGET_ARG_SUBSEQUENT_PARAM)
+	input++;
+      input++;
+    }
+
+  if (!attrs_found)
+    {
+      def->ndim = 1;
+      def->gdims[0] = 1;
+      def->gdims[1] = 1;
+      def->gdims[2] = 1;
+      def->wdims[0] = 1;
+      def->wdims[1] = 1;
+      def->wdims[2] = 1;
+      *result = def;
+      HSA_DEBUG ("GOMP_OFFLOAD_run called with no launch attributes\n");
+      return true;
+    }
+
+  struct GOMP_kernel_launch_attributes *kla;
+  kla = (struct GOMP_kernel_launch_attributes *) *input;
+  *result = kla;
+  if (kla->ndim != 1)
+    GOMP_PLUGIN_fatal ("HSA does not yet support number of dimensions "
+		       "different from one.");
+  if (kla->gdims[0] == 0)
+    return false;
+
+  HSA_DEBUG ("GOMP_OFFLOAD_run called with grid size %u and group size %u\n",
+	     kla->gdims[0], kla->wdims[0]);
+
+  return true;
+}
+
+/* Return true if the HSA runtime can run function FN_PTR.  */
+
+bool
+GOMP_OFFLOAD_can_run (void *fn_ptr)
+{
+  struct kernel_info *kernel = (struct kernel_info *) fn_ptr;
+  struct agent_info *agent = kernel->agent;
+  create_and_finalize_hsa_program (agent);
+
+  if (agent->prog_finalized_error)
+    goto failure;
+
+  init_kernel (kernel);
+  if (kernel->initialization_failed)
+    goto failure;
+
+  return true;
+
+failure:
+  if (suppress_host_fallback)
+    GOMP_PLUGIN_fatal ("HSA host fallback has been suppressed");
+  HSA_DEBUG ("HSA target cannot be launched, doing a host fallback\n");
+  return false;
+}
+
+/* Part of the libgomp plugin interface.  Run a kernel on device N and pass it
+   an array of pointers in VARS as a parameter.  The kernel is identified by
+   FN_PTR which must point to a kernel_info structure.  */
+
+void
+GOMP_OFFLOAD_run (int n, void *fn_ptr, void *vars, void **args)
+{
+  struct kernel_info *kernel = (struct kernel_info *) fn_ptr;
+  struct agent_info *agent = kernel->agent;
+  struct GOMP_kernel_launch_attributes def;
+  struct GOMP_kernel_launch_attributes *kla;
+  if (!parse_target_attributes (args, &def, &kla))
+    {
+      HSA_DEBUG ("Will not run HSA kernel because the grid size is zero\n");
+      return;
+    }
+  if (pthread_rwlock_rdlock (&agent->modules_rwlock))
+    GOMP_PLUGIN_fatal ("Unable to read-lock an HSA agent rwlock");
+
+  if (!agent->initialized)
+    GOMP_PLUGIN_fatal ("Agent must be initialized");
+
+  if (!kernel->initialized)
+    GOMP_PLUGIN_fatal ("Called kernel must be initialized");
+
+  struct GOMP_hsa_kernel_dispatch *shadow
+    = create_kernel_dispatch (kernel, kernel->max_omp_data_size);
+
+  if (debug)
+    {
+      fprintf (stderr, "\nKernel has following dependencies:\n");
+      print_kernel_dispatch (shadow, 2);
+    }
+
+  uint64_t index = hsa_queue_add_write_index_release (agent->command_q, 1);
+  HSA_DEBUG ("Got AQL index %llu\n", (long long int) index);
+
+  /* Wait until the queue is not full before writing the packet.   */
+  while (index - hsa_queue_load_read_index_acquire (agent->command_q)
+	 >= agent->command_q->size)
+    ;
+
+  hsa_kernel_dispatch_packet_t *packet;
+  packet = ((hsa_kernel_dispatch_packet_t *) agent->command_q->base_address)
+	   + index % agent->command_q->size;
+
+  memset (((uint8_t *) packet) + 4, 0, sizeof (*packet) - 4);
+  packet->setup |= (uint16_t) 1 << HSA_KERNEL_DISPATCH_PACKET_SETUP_DIMENSIONS;
+  packet->grid_size_x = kla->gdims[0];
+  uint32_t wgs = kla->wdims[0];
+  if (wgs == 0)
+    /* TODO: Provide a default via environment.  */
+    wgs = 64;
+  else if (wgs > kla->gdims[0])
+    wgs = kla->gdims[0];
+  packet->workgroup_size_x = wgs;
+  packet->grid_size_y = 1;
+  packet->workgroup_size_y = 1;
+  packet->grid_size_z = 1;
+  packet->workgroup_size_z = 1;
+  packet->private_segment_size = kernel->private_segment_size;
+  packet->group_segment_size = kernel->group_segment_size;
+  packet->kernel_object = kernel->object;
+  packet->kernarg_address = shadow->kernarg_address;
+  hsa_signal_t s;
+  s.handle = shadow->signal;
+  packet->completion_signal = s;
+  hsa_signal_store_relaxed (s, 1);
+  memcpy (shadow->kernarg_address, &vars, sizeof (vars));
+
+  memcpy (shadow->kernarg_address + sizeof (vars), &shadow,
+	  sizeof (struct hsa_kernel_runtime *));
+
+  HSA_DEBUG ("Copying kernel runtime pointer to kernarg_address\n");
+
+  uint16_t header;
+  header = HSA_PACKET_TYPE_KERNEL_DISPATCH << HSA_PACKET_HEADER_TYPE;
+  header |= HSA_FENCE_SCOPE_SYSTEM << HSA_PACKET_HEADER_ACQUIRE_FENCE_SCOPE;
+  header |= HSA_FENCE_SCOPE_SYSTEM << HSA_PACKET_HEADER_RELEASE_FENCE_SCOPE;
+
+  HSA_DEBUG ("Going to dispatch kernel %s\n", kernel->name);
+
+  __atomic_store_n ((uint16_t *) (&packet->header), header, __ATOMIC_RELEASE);
+  hsa_signal_store_release (agent->command_q->doorbell_signal, index);
+
+  /* TODO: GPU agents in Carrizo APUs cannot properly update L2 cache for
+     signal wait and signal load operations on their own and we need to
+     periodically call the hsa_signal_load_acquire on completion signals of
+     children kernels in the CPU to make that happen.  As soon the
+     limitation will be resolved, this workaround can be removed.  */
+
+  HSA_DEBUG ("Kernel dispatched, waiting for completion\n");
+
+  /* Root signal waits with 1ms timeout.  */
+  while (hsa_signal_wait_acquire (s, HSA_SIGNAL_CONDITION_LT, 1, 1000 * 1000,
+				  HSA_WAIT_STATE_BLOCKED) != 0)
+    for (unsigned i = 0; i < shadow->kernel_dispatch_count; i++)
+      {
+	hsa_signal_t child_s;
+	child_s.handle = shadow->children_dispatches[i]->signal;
+
+	HSA_DEBUG ("Waiting for children completion signal: %lu\n",
+		   shadow->children_dispatches[i]->signal);
+	hsa_signal_load_acquire (child_s);
+      }
+
+  release_kernel_dispatch (shadow);
+
+  if (pthread_rwlock_unlock (&agent->modules_rwlock))
+    GOMP_PLUGIN_fatal ("Unable to unlock an HSA agent rwlock");
+}
+
+/* Information to be passed to a thread running a kernel asycnronously.  */
+
+struct async_run_info
+{
+  int device;
+  void *tgt_fn;
+  void *tgt_vars;
+  void **args;
+  void *async_data;
+};
+
+/* Thread routine to run a kernel asynchronously.  */
+
+static void *
+run_kernel_asynchronously (void *thread_arg)
+{
+  struct async_run_info *info = (struct async_run_info *) thread_arg;
+  int device = info->device;
+  void *tgt_fn = info->tgt_fn;
+  void *tgt_vars = info->tgt_vars;
+  void **args = info->args;
+  void *async_data = info->async_data;
+
+  free (info);
+  GOMP_OFFLOAD_run (device, tgt_fn, tgt_vars, args);
+  GOMP_PLUGIN_target_task_completion (async_data);
+  return NULL;
+}
+
+/* Part of the libgomp plugin interface.  Run a kernel like GOMP_OFFLOAD_run
+   does, but asynchronously and call GOMP_PLUGIN_target_task_completion when it
+   has finished.  */
+
+void
+GOMP_OFFLOAD_async_run (int device, void *tgt_fn, void *tgt_vars,
+			void **args, void *async_data)
+{
+  pthread_t pt;
+  struct async_run_info *info;
+  HSA_DEBUG ("GOMP_OFFLOAD_async_run invoked\n")
+  info = GOMP_PLUGIN_malloc (sizeof (struct async_run_info));
+
+  info->device = device;
+  info->tgt_fn = tgt_fn;
+  info->tgt_vars = tgt_vars;
+  info->args = args;
+  info->async_data = async_data;
+
+  int err = pthread_create (&pt, NULL, &run_kernel_asynchronously, info);
+  if (err != 0)
+    GOMP_PLUGIN_fatal ("HSA asynchronous thread creation failed: %s",
+		       strerror (err));
+  err = pthread_detach (pt);
+  if (err != 0)
+    GOMP_PLUGIN_fatal ("Failed to detach a thread to run HSA kernel "
+		       "asynchronously: %s", strerror (err));
+}
+
+/* Deinitialize all information associated with MODULE and kernels within
+   it.  */
+
+void
+destroy_module (struct module_info *module)
+{
+  int i;
+  for (i = 0; i < module->kernel_count; i++)
+    if (pthread_mutex_destroy (&module->kernels[i].init_mutex))
+      GOMP_PLUGIN_fatal ("Failed to destroy an HSA kernel initialization "
+			 "mutex");
+}
+
+/* Part of the libgomp plugin interface.  Unload BRIG module described by
+   struct brig_image_desc in TARGET_DATA from agent number N.  */
+
+void
+GOMP_OFFLOAD_unload_image (int n, unsigned version, void *target_data)
+{
+  if (GOMP_VERSION_DEV (version) > GOMP_VERSION_HSA)
+    GOMP_PLUGIN_fatal ("Offload data incompatible with HSA plugin"
+		       " (expected %u, received %u)",
+		       GOMP_VERSION_HSA, GOMP_VERSION_DEV (version));
+
+  struct agent_info *agent;
+  agent = get_agent_info (n);
+  if (pthread_rwlock_wrlock (&agent->modules_rwlock))
+    GOMP_PLUGIN_fatal ("Unable to write-lock an HSA agent rwlock");
+
+  struct module_info *module = agent->first_module;
+  while (module)
+    {
+      if (module->image_desc == target_data)
+	break;
+      module = module->next;
+    }
+  if (!module)
+    GOMP_PLUGIN_fatal ("Attempt to unload an image that has never been "
+		       "loaded before");
+
+  remove_module_from_agent (agent, module);
+  destroy_module (module);
+  free (module);
+  destroy_hsa_program (agent);
+  if (pthread_rwlock_unlock (&agent->modules_rwlock))
+    GOMP_PLUGIN_fatal ("Unable to unlock an HSA agent rwlock");
+}
+
+/* Part of the libgomp plugin interface.  Deinitialize all information and
+   status associated with agent number N.  We do not attempt any
+   synchronization, assuming the user and libgomp will not attempt
+   deinitialization of a device that is in any way being used at the same
+   time.  */
+
+void
+GOMP_OFFLOAD_fini_device (int n)
+{
+  struct agent_info *agent = get_agent_info (n);
+  if (!agent->initialized)
+    return;
+
+  struct module_info *next_module = agent->first_module;
+  while (next_module)
+    {
+      struct module_info *module = next_module;
+      next_module = module->next;
+      destroy_module (module);
+      free (module);
+    }
+  agent->first_module = NULL;
+  destroy_hsa_program (agent);
+
+  release_agent_shared_libraries (agent);
+
+  hsa_status_t status = hsa_queue_destroy (agent->command_q);
+  if (status != HSA_STATUS_SUCCESS)
+    hsa_fatal ("Error destroying command queue", status);
+  status = hsa_queue_destroy (agent->kernel_dispatch_command_q);
+  if (status != HSA_STATUS_SUCCESS)
+    hsa_fatal ("Error destroying kernel dispatch command queue", status);
+  if (pthread_mutex_destroy (&agent->prog_mutex))
+    GOMP_PLUGIN_fatal ("Failed to destroy an HSA agent program mutex");
+  if (pthread_rwlock_destroy (&agent->modules_rwlock))
+    GOMP_PLUGIN_fatal ("Failed to destroy an HSA agent rwlock");
+  agent->initialized = false;
+}
+
+/* Part of the libgomp plugin interface.  Not implemented as it is not required
+   for HSA.  */
+
+void *
+GOMP_OFFLOAD_alloc (int ord, size_t size)
+{
+  GOMP_PLUGIN_fatal ("HSA GOMP_OFFLOAD_alloc is not implemented because "
+		     "it should never be called");
+}
+
+/* Part of the libgomp plugin interface.  Not implemented as it is not required
+   for HSA.  */
+
+void
+GOMP_OFFLOAD_free (int ord, void *ptr)
+{
+  GOMP_PLUGIN_fatal ("HSA GOMP_OFFLOAD_free is not implemented because "
+		     "it should never be called");
+}
+
+/* Part of the libgomp plugin interface.  Not implemented as it is not required
+   for HSA.  */
+
+void *
+GOMP_OFFLOAD_dev2host (int ord, void *dst, const void *src, size_t n)
+{
+  GOMP_PLUGIN_fatal ("HSA GOMP_OFFLOAD_dev2host is not implemented because "
+		     "it should never be called");
+}
+
+/* Part of the libgomp plugin interface.  Not implemented as it is not required
+   for HSA.  */
+
+void *
+GOMP_OFFLOAD_host2dev (int ord, void *dst, const void *src, size_t n)
+{
+  GOMP_PLUGIN_fatal ("HSA GOMP_OFFLOAD_host2dev is not implemented because "
+		     "it should never be called");
+}
+
+/* Part of the libgomp plugin interface.  Not implemented as it is not required
+   for HSA.  */
+
+void *
+GOMP_OFFLOAD_dev2dev (int ord, void *dst, const void *src, size_t n)
+{
+  GOMP_PLUGIN_fatal ("HSA GOMP_OFFLOAD_dev2dev is not implemented because "
+		     "it should never be called");
+}