[libomptarget] Implement host plugin for amdgpu

[libomptarget] Implement host plugin for amdgpu

Replacement for D71384. Primary difference is inlining the dependency on atmi
followed by extensive simplification and bugfixes. This is the latest version
from https://github.com/ROCm-Developer-Tools/amd-llvm-project/tree/aomp12 with
minor patches and a rename from hsa to amdgpu, on the basis that this can't be
used by other implementations of hsa without additional work.

This will not build unless the ROCM_DIR variable is passed so won't break other
builds. That variable is used to locate two amdgpu specific libraries that ship
as part of rocm:
libhsakmt at https://github.com/RadeonOpenCompute/ROCT-Thunk-Interface
libhsa-runtime64 at https://github.com/RadeonOpenCompute/ROCR-Runtime
These libraries build from source. The build scripts in those repos are for
shared libraries, but can be adapted to statically link both into this plugin.

There are caveats.
- This works well enough to run various tests and benchmarks, and will be used
  to support the current clang bring up
- It is adequately thread safe for the above but there will be races remaining
- It is not stylistically correct for llvm, though has had clang-format run
- It has suboptimal memory management and locking strategies
- The debug printing / error handling is inconsistent

I would like to contribute this pretty much as-is and then improve it in-tree.
This would be advantagous because the aomp12 branch that was in use for fixing
this codebase has just been joined with the amd internal rocm dev process.

Reviewed By: jdoerfert

Differential Revision: https://reviews.llvm.org/D85742

openmp/libomptarget/plugins/CMakeLists.txt

@@ -66,6 +66,7 @@ endif()
 endmacro()
 
 add_subdirectory(aarch64)
+add_subdirectory(amdgpu)
 add_subdirectory(cuda)
 add_subdirectory(ppc64)
 add_subdirectory(ppc64le)

openmp/libomptarget/plugins/amdgpu/CMakeLists.txt

@@ -0,0 +1,84 @@
+##===----------------------------------------------------------------------===##
+#
+#                     The LLVM Compiler Infrastructure
+#
+# This file is dual licensed under the MIT and the University of Illinois Open
+# Source Licenses. See LICENSE.txt for details.
+#
+##===----------------------------------------------------------------------===##
+#
+# Build a plugin for an AMDGPU machine if available.
+#
+##===----------------------------------------------------------------------===##
+
+################################################################################
+
+if(NOT LIBOMPTARGET_DEP_LIBELF_FOUND)
+  libomptarget_say("Not building AMDGPU plugin: LIBELF not found")
+  return()
+endif()
+
+if(NOT ROCM_DIR)
+  libomptarget_say("Not building AMDGPU plugin: ROCM_DIR is not set")
+  return()
+endif()
+
+set(LIBOMPTARGET_DEP_LIBHSA_INCLUDE_DIRS ${ROCM_DIR}/hsa/include ${ROCM_DIR}/hsa/include/hsa)
+set(LIBOMPTARGET_DEP_LIBHSA_LIBRARIES_DIRS ${ROCM_DIR}/hsa/lib)
+set(LIBOMPTARGET_DEP_LIBHSAKMT_LIBRARIES_DIRS ${ROCM_DIR}/lib)
+
+mark_as_advanced( LIBOMPTARGET_DEP_LIBHSA_INCLUDE_DIRS LIBOMPTARGET_DEP_LIBHSA_LIBRARIES_DIRS)
+
+if(NOT CMAKE_SYSTEM_PROCESSOR MATCHES "(x86_64)|(ppc64le)|(aarch64)$" AND CMAKE_SYSTEM_NAME MATCHES "Linux")
+  libomptarget_say("Not building amdgpu plugin: only support amdgpu in Linux x86_64, ppc64le, or aarch64 hosts.")
+  return()
+endif()
+libomptarget_say("Building amdgpu offloading plugin using ROCM_DIR = ${ROCM_DIR}")
+
+libomptarget_say("LIBOMPTARGET_DEP_LIBHSA_INCLUDE_DIRS: ${LIBOMPTARGET_DEP_LIBHSA_INCLUDE_DIRS}")
+libomptarget_say("LIBOMPTARGET_DEP_LIBHSA_LIBRARIES_DIRS ${LIBOMPTARGET_DEP_LIBHSA_LIBRARIES_DIRS}")
+libomptarget_say("LIBOMPTARGET_DEP_LIBHSAKMT_LIBRARIES_DIRS: ${LIBOMPTARGET_DEP_LIBHSAKMT_LIBRARIES_DIRS}")
+
+################################################################################
+# Define the suffix for the runtime messaging dumps.
+add_definitions(-DTARGET_NAME=AMDGPU)
+if(CMAKE_SYSTEM_PROCESSOR MATCHES "(ppc64le)|(aarch64)$")
+   add_definitions(-DLITTLEENDIAN_CPU=1)
+endif()
+
+if(CMAKE_BUILD_TYPE MATCHES Debug)
+  add_definitions(-DDEBUG)
+endif()
+
+include_directories(
+  ${LIBOMPTARGET_DEP_LIBHSA_INCLUDE_DIRS}
+  ${CMAKE_CURRENT_SOURCE_DIR}/impl
+)
+
+add_library(omptarget.rtl.amdgpu SHARED
+      impl/atmi.cpp
+      impl/atmi_interop_hsa.cpp
+      impl/data.cpp
+      impl/machine.cpp
+      impl/system.cpp
+      impl/utils.cpp
+      impl/msgpack.cpp
+      src/rtl.cpp 
+      )
+
+# Install plugin under the lib destination folder.
+# When we build for debug, OPENMP_LIBDIR_SUFFIX get set to -debug
+install(TARGETS omptarget.rtl.amdgpu LIBRARY DESTINATION "lib${OPENMP_LIBDIR_SUFFIX}")
+
+target_link_libraries(
+  omptarget.rtl.amdgpu
+  -lpthread -ldl -Wl,-rpath,${OPENMP_INSTALL_LIBDIR}
+  -L${LIBOMPTARGET_DEP_LIBHSA_LIBRARIES_DIRS} -L${LIBOMPTARGET_DEP_LIBHSAKMT_LIBRARIES_DIRS} -lhsa-runtime64 -lhsakmt -Wl,-rpath,${LIBOMPTARGET_DEP_LIBHSA_LIBRARIES_DIRS},-rpath,${LIBOMPTARGET_DEP_LIBHSAKMT_LIBRARIES_DIRS}
+  -lelf
+  "-Wl,--version-script=${CMAKE_CURRENT_SOURCE_DIR}/../exports"
+  "-Wl,-z,defs"
+  )
+
+# Report to the parent scope that we are building a plugin for amdgpu
+set(LIBOMPTARGET_SYSTEM_TARGETS "${LIBOMPTARGET_SYSTEM_TARGETS} amdgcn-amd-amdhsa" PARENT_SCOPE)
+

openmp/libomptarget/plugins/amdgpu/impl/atmi.cpp

@@ -0,0 +1,44 @@
+/*===--------------------------------------------------------------------------
+ *              ATMI (Asynchronous Task and Memory Interface)
+ *
+ * This file is distributed under the MIT License. See LICENSE.txt for details.
+ *===------------------------------------------------------------------------*/
+#include "rt.h"
+/*
+ * Initialize/Finalize
+ */
+atmi_status_t atmi_init() { return core::Runtime::Initialize(); }
+
+atmi_status_t atmi_finalize() { return core::Runtime::Finalize(); }
+
+/*
+ * Machine Info
+ */
+atmi_machine_t *atmi_machine_get_info() {
+  return core::Runtime::GetMachineInfo();
+}
+
+/*
+ * Modules
+ */
+atmi_status_t atmi_module_register_from_memory_to_place(
+    void *module_bytes, size_t module_size, atmi_place_t place,
+    atmi_status_t (*on_deserialized_data)(void *data, size_t size,
+                                          void *cb_state),
+    void *cb_state) {
+  return core::Runtime::getInstance().RegisterModuleFromMemory(
+      module_bytes, module_size, place, on_deserialized_data, cb_state);
+}
+
+/*
+ * Data
+ */
+atmi_status_t atmi_memcpy(void *dest, const void *src, size_t size) {
+  return core::Runtime::Memcpy(dest, src, size);
+}
+
+atmi_status_t atmi_free(void *ptr) { return core::Runtime::Memfree(ptr); }
+
+atmi_status_t atmi_malloc(void **ptr, size_t size, atmi_mem_place_t place) {
+  return core::Runtime::Malloc(ptr, size, place);
+}

openmp/libomptarget/plugins/amdgpu/impl/atmi.h

@@ -0,0 +1,203 @@
+/*===--------------------------------------------------------------------------
+ *              ATMI (Asynchronous Task and Memory Interface)
+ *
+ * This file is distributed under the MIT License. See LICENSE.txt for details.
+ *===------------------------------------------------------------------------*/
+#ifndef INCLUDE_ATMI_H_
+#define INCLUDE_ATMI_H_
+
+#define ROCM_VERSION_MAJOR 3
+#define ROCM_VERSION_MINOR 2
+
+/** \defgroup enumerations Enumerated Types
+ * @{
+ */
+
+/**
+ * @brief Status codes.
+ */
+typedef enum atmi_status_t {
+  /**
+   * The function has been executed successfully.
+   */
+  ATMI_STATUS_SUCCESS = 0,
+  /**
+   * A undocumented error has occurred.
+   */
+  ATMI_STATUS_UNKNOWN = 1,
+  /**
+   * A generic error has occurred.
+   */
+  ATMI_STATUS_ERROR = 2,
+} atmi_status_t;
+
+/**
+ * @brief Device Types.
+ */
+typedef enum atmi_devtype_s {
+  ATMI_DEVTYPE_CPU = 0x0001,
+  ATMI_DEVTYPE_iGPU = 0x0010,                               // Integrated GPU
+  ATMI_DEVTYPE_dGPU = 0x0100,                               // Discrete GPU
+  ATMI_DEVTYPE_GPU = ATMI_DEVTYPE_iGPU | ATMI_DEVTYPE_dGPU, // Any GPU
+  ATMI_DEVTYPE_ALL = 0x111 // Union of all device types
+} atmi_devtype_t;
+
+/**
+ * @brief Memory Access Type.
+ */
+typedef enum atmi_memtype_s {
+  ATMI_MEMTYPE_FINE_GRAINED = 0,
+  ATMI_MEMTYPE_COARSE_GRAINED = 1,
+  ATMI_MEMTYPE_ANY
+} atmi_memtype_t;
+
+/**
+ * @brief ATMI Memory Fences for Tasks.
+ */
+typedef enum atmi_task_fence_scope_s {
+  /**
+   * No memory fence applied; external fences have to be applied around the task
+   * launch/completion.
+   */
+  ATMI_FENCE_SCOPE_NONE = 0,
+  /**
+   * The fence is applied to the device.
+   */
+  ATMI_FENCE_SCOPE_DEVICE = 1,
+  /**
+   * The fence is applied to the entire system.
+   */
+  ATMI_FENCE_SCOPE_SYSTEM = 2
+} atmi_task_fence_scope_t;
+
+/** @} */
+
+/** \defgroup common Common ATMI Structures
+ *  @{
+ */
+
+/**
+ * @brief ATMI Compute Place
+ */
+typedef struct atmi_place_s {
+  /**
+   * The node in a cluster where computation should occur.
+   * Default is node_id = 0 for local computations.
+   */
+  unsigned int node_id;
+  /**
+   * Device type: CPU, GPU or DSP
+   */
+  atmi_devtype_t type;
+  /**
+   * The device ordinal number ordered by runtime; -1 for any
+   */
+  int device_id;
+} atmi_place_t;
+
+/**
+ * @brief ATMI Memory Place
+ */
+typedef struct atmi_mem_place_s {
+  /**
+   * The node in a cluster where computation should occur.
+   * Default is node_id = 0 for local computations.
+   */
+  unsigned int node_id;
+  /**
+   * Device type: CPU, GPU or DSP
+   */
+  atmi_devtype_t dev_type;
+  /**
+   * The device ordinal number ordered by runtime; -1 for any
+   */
+  int dev_id;
+  // atmi_memtype_t mem_type;        // Fine grained or Coarse grained
+  /**
+   * The memory space/region ordinal number ordered by runtime; -1 for any
+   */
+  int mem_id;
+} atmi_mem_place_t;
+
+/**
+ * @brief ATMI Memory Space/region Structure
+ */
+typedef struct atmi_memory_s {
+  /**
+   * Memory capacity
+   */
+  unsigned long int capacity;
+  /**
+   * Memory type
+   */
+  atmi_memtype_t type;
+} atmi_memory_t;
+
+/**
+ * @brief ATMI Device Structure
+ */
+typedef struct atmi_device_s {
+  /**
+   * Device type: CPU, GPU or DSP
+   */
+  atmi_devtype_t type;
+  /**
+   * The number of compute cores
+   */
+  unsigned int core_count;
+  /**
+   * The number of memory spaces/regions that are accessible
+   * from this device
+   */
+  unsigned int memory_count;
+  /**
+   * Array of memory spaces/regions that are accessible
+   * from this device.
+   */
+  atmi_memory_t *memories;
+} atmi_device_t;
+
+/**
+ * @brief ATMI Machine Structure
+ */
+typedef struct atmi_machine_s {
+  /**
+   * The number of devices categorized by the device type
+   */
+  unsigned int device_count_by_type[ATMI_DEVTYPE_ALL];
+  /**
+   * The device structures categorized by the device type
+   */
+  atmi_device_t *devices_by_type[ATMI_DEVTYPE_ALL];
+} atmi_machine_t;
+
+// Below are some helper macros that can be used to setup
+// some of the ATMI data structures.
+#define ATMI_PLACE_CPU(node, cpu_id)                                           \
+  { .node_id = node, .type = ATMI_DEVTYPE_CPU, .device_id = cpu_id }
+#define ATMI_PLACE_GPU(node, gpu_id)                                           \
+  { .node_id = node, .type = ATMI_DEVTYPE_GPU, .device_id = gpu_id }
+#define ATMI_MEM_PLACE_CPU(node, cpu_id)                                       \
+  {                                                                            \
+    .node_id = node, .dev_type = ATMI_DEVTYPE_CPU, .dev_id = cpu_id,           \
+    .mem_id = -1                                                               \
+  }
+#define ATMI_MEM_PLACE_GPU(node, gpu_id)                                       \
+  {                                                                            \
+    .node_id = node, .dev_type = ATMI_DEVTYPE_GPU, .dev_id = gpu_id,           \
+    .mem_id = -1                                                               \
+  }
+#define ATMI_MEM_PLACE_CPU_MEM(node, cpu_id, cpu_mem_id)                       \
+  {                                                                            \
+    .node_id = node, .dev_type = ATMI_DEVTYPE_CPU, .dev_id = cpu_id,           \
+    .mem_id = cpu_mem_id                                                       \
+  }
+#define ATMI_MEM_PLACE_GPU_MEM(node, gpu_id, gpu_mem_id)                       \
+  {                                                                            \
+    .node_id = node, .dev_type = ATMI_DEVTYPE_GPU, .dev_id = gpu_id,           \
+    .mem_id = gpu_mem_id                                                       \
+  }
+#define ATMI_MEM_PLACE(d_type, d_id, m_id)                                     \
+  { .node_id = 0, .dev_type = d_type, .dev_id = d_id, .mem_id = m_id }
+
+#endif // INCLUDE_ATMI_H_