[AMDGPU][Libomptarget] Remove g_atmi_machine global

Turns out the only purpose of this class was verify if device ID
was in range or not which could be done easily by using g_atl_machine.

Still getting rid of g_atl_machine is pending which would be done in
a later patch.

Reviewed By: JonChesterfield

Differential Revision: https://reviews.llvm.org/D103443

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

@@ -10,13 +10,6 @@
 #include <hsa_ext_amd.h>
 #include <memory>
 
-/*
- * Machine Info
- */
-atmi_machine_t *atmi_machine_get_info() {
-  return core::Runtime::GetMachineInfo();
-}
-
 /*
  * Data
  */

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

@@ -5,6 +5,11 @@
  *===------------------------------------------------------------------------*/
 #include "atmi_interop_hsa.h"
 #include "internal.h"
+#include "machine.h"
+
+// TODO: need to get rid of this as well
+
+extern ATLMachine g_atl_machine;
 
 hsa_status_t atmi_interop_hsa_get_symbol_info(
     const std::map<std::string, atl_symbol_info_t> &SymbolInfoTable,
@@ -18,11 +23,10 @@ hsa_status_t atmi_interop_hsa_get_symbol_info(
      atmi_memcpy(signal, host_add, var_addr, var_size);
   */
 
-  atmi_machine_t *machine = atmi_machine_get_info();
-  if (!symbol || !var_addr || !var_size || !machine)
+  if (!symbol || !var_addr || !var_size)
     return HSA_STATUS_ERROR;
   if (DeviceId < 0 ||
-      DeviceId >= machine->device_count_by_type[ATMI_DEVTYPE_GPU])
+      DeviceId >= g_atl_machine.processors<ATLGPUProcessor>().size())
     return HSA_STATUS_ERROR;
 
   // get the symbol info
@@ -52,11 +56,10 @@ hsa_status_t atmi_interop_hsa_get_kernel_info(
                                   &val);
   */
 
-  atmi_machine_t *machine = atmi_machine_get_info();
-  if (!kernel_name || !value || !machine)
+  if (!kernel_name || !value)
     return HSA_STATUS_ERROR;
   if (DeviceId < 0 ||
-      DeviceId >= machine->device_count_by_type[ATMI_DEVTYPE_GPU])
+      DeviceId >= g_atl_machine.processors<ATLGPUProcessor>().size())
     return HSA_STATUS_ERROR;
 
   hsa_status_t status = HSA_STATUS_SUCCESS;

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

@@ -58,23 +58,6 @@ hsa_status_t atmi_module_register_from_memory_to_place(
 
 /** @} */
 
-/** \defgroup machine ATMI Machine
- * @{
- */
-/**
- * @brief ATMI's device discovery function to get the current machine's
- * topology.
- *
- * @detail The @p atmi_machine_t structure is a tree-based representation of the
- * compute and memory elements in the current node. Once ATMI is initialized,
- * this function can be called to retrieve the pointer to this global structure.
- *
- * @return Returns a pointer to a global structure of tyoe @p atmi_machine_t.
- * Returns NULL if ATMI is not initialized.
- */
-atmi_machine_t *atmi_machine_get_info();
-/** @} */
-
 hsa_status_t atmi_memcpy_h2d(hsa_signal_t signal, void *deviceDest,
                              const void *hostSrc, size_t size,
                              hsa_agent_t agent);

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

@@ -49,8 +49,6 @@ public:
     return instance;
   }
 
-  // machine info
-  static atmi_machine_t *GetMachineInfo();
   // modules
   static hsa_status_t RegisterModuleFromMemory(
       void *, size_t, atmi_place_t,

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

@@ -140,8 +140,6 @@ static const std::map<std::string, KernelArgMD::ValueKind> ArgValueKind = {
     {"hidden_hostcall_buffer", KernelArgMD::ValueKind::HiddenHostcallBuffer},
 };
 
-// global variables. TODO: Get rid of these
-atmi_machine_t g_atmi_machine;
 ATLMachine g_atl_machine;
 
 std::vector<hsa_amd_memory_pool_t> atl_gpu_kernarg_pools;
@@ -154,12 +152,6 @@ std::vector<hsa_amd_memory_pool_t> atl_gpu_kernarg_pools;
 atl_context_t atlc = {.struct_initialized = false};
 
 namespace core {
-/* Machine Info */
-atmi_machine_t *Runtime::GetMachineInfo() {
-  if (!atlc.g_hsa_initialized)
-    return NULL;
-  return &g_atmi_machine;
-}
 
 hsa_status_t allow_access_to_all_gpu_agents(void *ptr) {
   std::vector<ATLGPUProcessor> &gpu_procs =
@@ -362,12 +354,7 @@ static hsa_status_t init_compute_and_memory() {
     }
   }
 
-  g_atmi_machine.device_count_by_type[ATMI_DEVTYPE_CPU] = cpu_procs.size();
-  g_atmi_machine.device_count_by_type[ATMI_DEVTYPE_GPU] = gpu_procs.size();
-
   size_t num_procs = cpu_procs.size() + gpu_procs.size();
-  // g_atmi_machine.devices = (atmi_device_t *)malloc(num_procs *
-  // sizeof(atmi_device_t));
   atmi_device_t *all_devices = reinterpret_cast<atmi_device_t *>(
       malloc(num_procs * sizeof(atmi_device_t)));
   int num_iGPUs = 0;
@@ -385,17 +372,10 @@ static hsa_status_t init_compute_and_memory() {
   DEBUG_PRINT("dGPU Agents: %d\n", num_dGPUs);
   DEBUG_PRINT("GPU Agents: %lu\n", gpu_procs.size());
 
-  g_atmi_machine.device_count_by_type[ATMI_DEVTYPE_iGPU] = num_iGPUs;
-  g_atmi_machine.device_count_by_type[ATMI_DEVTYPE_dGPU] = num_dGPUs;
-
   int cpus_begin = 0;
   int cpus_end = cpu_procs.size();
   int gpus_begin = cpu_procs.size();
   int gpus_end = cpu_procs.size() + gpu_procs.size();
-  g_atmi_machine.devices_by_type[ATMI_DEVTYPE_CPU] = &all_devices[cpus_begin];
-  g_atmi_machine.devices_by_type[ATMI_DEVTYPE_GPU] = &all_devices[gpus_begin];
-  g_atmi_machine.devices_by_type[ATMI_DEVTYPE_iGPU] = &all_devices[gpus_begin];
-  g_atmi_machine.devices_by_type[ATMI_DEVTYPE_dGPU] = &all_devices[gpus_begin];
   int proc_index = 0;
   for (int i = cpus_begin; i < cpus_end; i++) {
     all_devices[i].type = cpu_procs[proc_index].type();