From 0141e6809f6cfbf3e803d321f341a8d13c4d148e Mon Sep 17 00:00:00 2001
From: Saleel Kudchadker <Saleel.Kudchadker@amd.com>
Date: Mon, 8 May 2023 16:43:29 -0700
Subject: [PATCH] SWDEV-398151 - Partly relax static engine allocation

Change-Id: I4903b51a34b597a2e84d771b52cf629f877dba05


[ROCm/clr commit: 0b475284e93fae11979d892a18fee90e74f7131c]
---
 projects/clr/rocclr/device/rocm/rocblit.cpp   |  9 ++------
 projects/clr/rocclr/device/rocm/rocdevice.cpp | 21 +++++++++----------
 projects/clr/rocclr/device/rocm/rocdevice.hpp |  5 ++---
 3 files changed, 14 insertions(+), 21 deletions(-)

diff --git a/projects/clr/rocclr/device/rocm/rocblit.cpp b/projects/clr/rocclr/device/rocm/rocblit.cpp
index 1c0f8741ef..7284ad1b09 100644
--- a/projects/clr/rocclr/device/rocm/rocblit.cpp
+++ b/projects/clr/rocclr/device/rocm/rocblit.cpp
@@ -692,15 +692,10 @@ bool DmaBlitManager::hsaCopy(const Memory& srcMemory, const Memory& dstMemory,
   }
 
   if (engine != HwQueueEngine::Unknown) {
-    if (copyMask == Device::kSkipQueryStatus) {
-      // Do not query engine status or take copy_on_engine path
-      status = HSA_STATUS_ERROR_OUT_OF_RESOURCES;
-    }
-
     if (copyMask == 0) {
       // Check SDMA engine status
       status = hsa_amd_memory_copy_engine_status(dstAgent, srcAgent, &freeEngineMask);
-      ClPrint(amd::LOG_DEBUG, amd::LOG_COPY, "Query copy engine status %x, free_engine mask %x",
+      ClPrint(amd::LOG_DEBUG, amd::LOG_COPY, "Query copy engine status %x, free_engine mask 0x%x",
               status, freeEngineMask);
       // Return a mask with the rightmost bit set
       copyMask = freeEngineMask - (freeEngineMask & (freeEngineMask - 1));
@@ -729,7 +724,7 @@ bool DmaBlitManager::hsaCopy(const Memory& srcMemory, const Memory& dstMemory,
     auto wait_events = gpu().Barriers().WaitingSignal(engine);
     hsa_signal_t active = gpu().Barriers().ActiveSignal(kInitSignalValueOne, gpu().timestamp());
     ClPrint(amd::LOG_DEBUG, amd::LOG_COPY,
-        "HSA Async Blit Copy dst=0x%zx, src=0x%zx, size=%ld, wait_event=0x%zx, "
+        "HSA Async Copy dst=0x%zx, src=0x%zx, size=%ld, wait_event=0x%zx, "
         "completion_signal=0x%zx",
         dst, src, size[0], (wait_events.size() != 0) ? wait_events[0].handle : 0,
         active.handle);
diff --git a/projects/clr/rocclr/device/rocm/rocdevice.cpp b/projects/clr/rocclr/device/rocm/rocdevice.cpp
index bb791d4c6f..f4c2ebc766 100644
--- a/projects/clr/rocclr/device/rocm/rocdevice.cpp
+++ b/projects/clr/rocclr/device/rocm/rocdevice.cpp
@@ -1243,17 +1243,17 @@ bool Device::populateOCLDeviceConstants() {
 
   // Find SDMA read mask
   if (HSA_STATUS_SUCCESS != hsa_amd_memory_copy_engine_status(getCpuAgent(), getBackendDevice(),
-                                                              &maxSdmaReadMask)) {
+                                                              &maxSdmaReadMask_)) {
     return false;
   }
-  assert(maxSdmaReadMask > 0 && "No SDMA engines available for Read");
+  assert(maxSdmaReadMask_ > 0 && "No SDMA engines available for Read");
 
   // Find SDMA write mask
   if (HSA_STATUS_SUCCESS != hsa_amd_memory_copy_engine_status(getBackendDevice(), getCpuAgent(),
-                                                              &maxSdmaWriteMask)) {
+                                                              &maxSdmaWriteMask_)) {
     return false;
   }
-  assert(maxSdmaWriteMask > 0 && "No SDMA engines available for Write");
+  assert(maxSdmaWriteMask_ > 0 && "No SDMA engines available for Write");
 
   info_.localMemSizePerCU_ = group_segment_size;
   info_.localMemSize_ = group_segment_size;
@@ -1657,8 +1657,12 @@ bool Device::populateOCLDeviceConstants() {
     LogError("HSA_AMD_AGENT_INFO_SVM_DIRECT_HOST_ACCESS query failed.");
   }
 
-  ClPrint(amd::LOG_INFO, amd::LOG_INIT, "HMM support: %d, xnack: %d, direct host access: %d",
+  ClPrint(amd::LOG_INFO, amd::LOG_INIT, "Gfx Major/Minor/Stepping: %d/%d/%d", isa().versionMajor(),
+  isa().versionMinor(), isa().versionStepping());
+  ClPrint(amd::LOG_INFO, amd::LOG_INIT, "HMM support: %d, XNACK: %d, Direct host access: %d",
     info_.hmmSupported_, info_.hmmCpuMemoryAccessible_, info_.hmmDirectHostAccess_);
+  ClPrint(amd::LOG_INFO, amd::LOG_INIT, "Max SDMA Read Mask: 0x%x, Max SDMA Write Mask: 0x%x",
+          maxSdmaReadMask_, maxSdmaWriteMask_);
 
   info_.globalCUMask_ = {};
   info_.virtualMemoryManagement_ = false;
@@ -3338,12 +3342,7 @@ uint32_t Device::fetchSDMAMask(const device::BlitManager* handle, bool readEngin
     }
   }
 
-  uint32_t mask = (readEngine ? maxSdmaReadMask : maxSdmaWriteMask) & engine;
-  if (engine != 0 && mask == 0 ) {
-    return kSkipQueryStatus;
-  } else {
-    return mask;
-  }
+  return (readEngine ? maxSdmaReadMask_ : maxSdmaWriteMask_) & engine;
 }
 
 // ================================================================================================
diff --git a/projects/clr/rocclr/device/rocm/rocdevice.hpp b/projects/clr/rocclr/device/rocm/rocdevice.hpp
index 10a02b9b3d..e23c25a1e3 100644
--- a/projects/clr/rocclr/device/rocm/rocdevice.hpp
+++ b/projects/clr/rocclr/device/rocm/rocdevice.hpp
@@ -624,13 +624,12 @@ class Device : public NullDevice {
   std::vector<std::map<hsa_queue_t*, QueueInfo>> queueWithCUMaskPool_;
 
   //! Read and Write mask for device<->host
-  uint32_t maxSdmaReadMask;
-  uint32_t maxSdmaWriteMask;
+  uint32_t maxSdmaReadMask_;
+  uint32_t maxSdmaWriteMask_;
   //! Map of SDMA engineId<->stream
   mutable std::map<uint32_t, const device::BlitManager*> engineAssignMap_;
 
  public:
-  constexpr static uint32_t kSkipQueryStatus = 1 << 31;
   std::atomic<uint> numOfVgpus_;  //!< Virtual gpu unique index
 
   //! enum for keeping the total and available queue priorities