From 010253430f3c94bcbf0665811899d304653d1f63 Mon Sep 17 00:00:00 2001
From: kjayapra-amd <karthik.jayaprakash@amd.com>
Date: Fri, 7 Feb 2025 14:51:33 -0500
Subject: [PATCH] SWDEV-516303 - Remove SDMA retainer logic to select the
 engine.

Change-Id: I818129444131825cdb87e06cb495afa3e5cdb683


[ROCm/clr commit: 1f583a687090577b4c2544d98f64c595b1215338]
---
 projects/clr/rocclr/device/rocm/rocblit.cpp | 18 ++----------------
 projects/clr/rocclr/device/rocm/rocblit.hpp |  2 --
 2 files changed, 2 insertions(+), 18 deletions(-)

diff --git a/projects/clr/rocclr/device/rocm/rocblit.cpp b/projects/clr/rocclr/device/rocm/rocblit.cpp
index 0634120106..1471cc4187 100644
--- a/projects/clr/rocclr/device/rocm/rocblit.cpp
+++ b/projects/clr/rocclr/device/rocm/rocblit.cpp
@@ -32,8 +32,7 @@ DmaBlitManager::DmaBlitManager(VirtualGPU& gpu, Setup setup)
     : HostBlitManager(gpu, setup),
       MinSizeForPinnedTransfer(dev().settings().pinnedMinXferSize_),
       completeOperation_(false),
-      context_(nullptr),
-      sdmaEngineRetainCount_(0) {
+      context_(nullptr) {
         dev().getSdmaRWMasks(&sdmaEngineReadMask_, &sdmaEngineWriteMask_);
       }
 
@@ -489,16 +488,9 @@ inline bool DmaBlitManager::rocrCopyBuffer(address dst, hsa_agent_t& dstAgent,
   if ((srcAgent.handle == dev().getCpuAgent().handle) &&
       (dstAgent.handle != dev().getCpuAgent().handle)) {
     engine = HwQueueEngine::SdmaWrite;
-    // Track the HtoD copies and increment the count. The last used SDMA engine might be busy
-    // and using it everytime can cause contention. When the count exceeds the threshold,
-    // reset it so as to check the engine status and fetch the new mask.
-    sdmaEngineRetainCount_ = (sdmaEngineRetainCount_ > kRetainCountThreshold)
-                              ? 0 : (sdmaEngineRetainCount_ + 1);
   } else if ((srcAgent.handle != dev().getCpuAgent().handle) &&
              (dstAgent.handle == dev().getCpuAgent().handle)) {
     engine = HwQueueEngine::SdmaRead;
-    // Track the DtoH copies and decrement the count.
-    sdmaEngineRetainCount_--;
   }
 
   if (engine == HwQueueEngine::Unknown && forceSDMA) {
@@ -510,13 +502,7 @@ inline bool DmaBlitManager::rocrCopyBuffer(address dst, hsa_agent_t& dstAgent,
   hsa_signal_t active = gpu().Barriers().ActiveSignal(kInitSignalValueOne, gpu().timestamp());
 
   if (!kUseRegularCopyApi && engine != HwQueueEngine::Unknown) {
-    if (sdmaEngineRetainCount_ > 0) {
-      // Check if there a recently used SDMA engine for the stream
-      copyMask = gpu().getLastUsedSdmaEngine();
-      ClPrint(amd::LOG_DEBUG, amd::LOG_COPY, "Last copy mask 0x%x", copyMask);
-      copyMask &= (engine == HwQueueEngine::SdmaRead ?
-                   sdmaEngineReadMask_ : sdmaEngineWriteMask_);
-    }
+    copyMask = gpu().getLastUsedSdmaEngine();
     if (copyMask == 0) {
       // Check SDMA engine status
       status = hsa_amd_memory_copy_engine_status(dstAgent, srcAgent, &freeEngineMask);
diff --git a/projects/clr/rocclr/device/rocm/rocblit.hpp b/projects/clr/rocclr/device/rocm/rocblit.hpp
index 5153523b6d..4d9e295f38 100644
--- a/projects/clr/rocclr/device/rocm/rocblit.hpp
+++ b/projects/clr/rocclr/device/rocm/rocblit.hpp
@@ -240,8 +240,6 @@ class DmaBlitManager : public device::HostBlitManager {
   const size_t MinSizeForPinnedTransfer;
   bool completeOperation_;                    //!< DMA blit manager must complete operation
   amd::Context* context_;                     //!< A dummy context
-  mutable int64_t sdmaEngineRetainCount_;     //!< Keeps track of memcopies to either get the last
-                                              //!< used SDMA engine or fetch the new mask
   uint32_t sdmaEngineReadMask_;               //!< SDMA Engine Read Mask
   uint32_t sdmaEngineWriteMask_;              //!< SDMA Engine Write Mask