rocr: Workaround for peak SDMA b/w on gfx94x (#626)

* Ideally SDMA0/1/2 are the engines to use for H2D/D2H due to physical
  PCIE proximity
* Allow using same src/dst agent for SDMA query apis

projects/rocr-runtime/runtime/hsa-runtime/core/runtime/amd_gpu_agent.cpp

@@ -1244,11 +1244,30 @@ hsa_status_t GpuAgent::DmaCopyStatus(core::Agent& dst_agent, core::Agent& src_ag
 
 hsa_status_t GpuAgent::DmaPreferredEngine(core::Agent& dst_agent, core::Agent& src_agent,
                                           uint32_t *recommended_ids_mask) {
-  assert(((src_agent.device_type() == core::Agent::kAmdGpuDevice) ||
-          (dst_agent.device_type() == core::Agent::kAmdGpuDevice)) &&
-         ("Both devices are CPU agents which is not expected"));
+  // From the collected data, gfx94x performance is better only for first 3 SDMA engines
+  bool isGfx94x = (isa_->GetMajorVersion() == 9 &&
+                  (isa_->GetMinorVersion() == 4 || isa_->GetMinorVersion() == 5));
 
-  *recommended_ids_mask = rec_sdma_eng_id_peers_info_[dst_agent.public_handle().handle];
+  if (isGfx94x &&
+      ((src_agent.device_type() == core::Agent::kAmdCpuDevice &&
+        dst_agent.device_type() == core::Agent::kAmdGpuDevice) ||
+        (src_agent.device_type() == core::Agent::kAmdGpuDevice &&
+        dst_agent.device_type() == core::Agent::kAmdCpuDevice))) {
+
+    if (src_agent.device_type() == core::Agent::kAmdCpuDevice) {
+      // Host to Device: Use SDMA engine 0 if available
+      *recommended_ids_mask = HSA_AMD_SDMA_ENGINE_0;
+    } else {
+      // Device to Host: Use SDMA engines 1 and 2 if available
+      *recommended_ids_mask = HSA_AMD_SDMA_ENGINE_1;
+
+      if (properties_.NumSdmaEngines + properties_.NumSdmaXgmiEngines > 2) {
+        *recommended_ids_mask |= HSA_AMD_SDMA_ENGINE_2;
+      }
+    }
+  } else {
+    *recommended_ids_mask = rec_sdma_eng_id_peers_info_[dst_agent.public_handle().handle];
+  }
 
   return HSA_STATUS_SUCCESS;
 }

projects/rocr-runtime/runtime/hsa-runtime/core/runtime/runtime.cpp

@@ -605,10 +605,6 @@ hsa_status_t Runtime::CopyMemoryStatus(core::Agent* dst_agent, core::Agent* src_
   const bool src_gpu = (src_agent->device_type() == core::Agent::DeviceType::kAmdGpuDevice);
   core::Agent* copy_agent = (src_gpu) ? src_agent : dst_agent;
 
-  if (dst_agent == src_agent) {
-    return HSA_STATUS_ERROR_INVALID_AGENT;
-  }
-
   return copy_agent->DmaCopyStatus(*dst_agent, *src_agent, engine_ids_mask);
 }
 
@@ -617,10 +613,6 @@ hsa_status_t Runtime::GetPreferredEngine(core::Agent* dst_agent, core::Agent* sr
   const bool src_gpu = (src_agent->device_type() == core::Agent::DeviceType::kAmdGpuDevice);
   core::Agent* copy_agent = (src_gpu) ? src_agent : dst_agent;
 
-  if (dst_agent == src_agent) {
-    return HSA_STATUS_ERROR_INVALID_AGENT;
-  }
-
   return copy_agent->DmaPreferredEngine(*dst_agent, *src_agent, recommended_ids_mask);
 }
 

projects/rocr-runtime/runtime/hsa-runtime/inc/hsa_ext_amd.h

@@ -1810,8 +1810,6 @@ hsa_status_t HSA_API
  *
  * @retval ::HSA_STATUS_ERROR_OUT_OF_RESOURCES Agent does not have available SDMA engines.
  *
- * @retval ::HSA_STATUS_ERROR_INVALID_AGENT dst_agent and src_agent are the same as
- * dst_agent == src_agent is generally used for shader copies.
  */
 hsa_status_t HSA_API
 hsa_amd_memory_copy_engine_status(hsa_agent_t dst_agent, hsa_agent_t src_agent,
@@ -1828,8 +1826,6 @@ hsa_amd_memory_copy_engine_status(hsa_agent_t dst_agent, hsa_agent_t src_agent,
  *
  * @retval ::HSA_STATUS_SUCCESS For mask returned
  *
- * @retval ::HSA_STATUS_ERROR_INVALID_AGENT dst_agent and src_agent are the same as
- * dst_agent == src_agent is generally used for shader copies.
  */
 hsa_status_t HSA_API
 hsa_amd_memory_get_preferred_copy_engine(hsa_agent_t dst_agent, hsa_agent_t src_agent,