From 280a458d0c5a7b20393f09cda892d1a0d9faa698 Mon Sep 17 00:00:00 2001
From: Sean Keely <Sean.Keely@amd.com>
Date: Fri, 20 Aug 2021 00:56:04 -0500
Subject: [PATCH] Workaround gfx90a SDMA0 quirk.

Because of sharing ports with other engines, the
hardware design team has advised that SDMA0 on gfx90a
should only be used for host-to-device data transfers.
The recommendation is to use SDMA1 for any device-to-device
or device-to-host data transfers.

A driver change will ensure that, for each gfx90a
device, only the first PCIe SDMA queue a process
requests will possibly be from SDMA0. This patch ensures
that the first PCIe queue requested (which may be from
SDMA0) is always set up for host-to-device.

Change-Id: I6793ca95596dedaed9d5be1dbd9469ceef2a5c33
---
 .../hsa-runtime/core/runtime/amd_gpu_agent.cpp    | 15 +++++++++++----
 1 file changed, 11 insertions(+), 4 deletions(-)

diff --git a/runtime/hsa-runtime/core/runtime/amd_gpu_agent.cpp b/runtime/hsa-runtime/core/runtime/amd_gpu_agent.cpp
index 6dc93a4403..cb6c56d88c 100644
--- a/runtime/hsa-runtime/core/runtime/amd_gpu_agent.cpp
+++ b/runtime/hsa-runtime/core/runtime/amd_gpu_agent.cpp
@@ -599,7 +599,7 @@ void GpuAgent::InitDma() {
   queues_[QueueUtility].reset(queue_lambda);
 
   // Decide which engine to use for blits.
-  auto blit_lambda = [this](bool use_xgmi, lazy_ptr<core::Queue>& queue) {
+  auto blit_lambda = [this](bool use_xgmi, lazy_ptr<core::Queue>& queue, bool isHostToDev) {
     Flag::SDMA_OVERRIDE sdma_override = core::Runtime::runtime_singleton_->flag().enable_sdma();
 
     // User SDMA queues are unstable on gfx8 and unsupported on gfx1013.
@@ -608,6 +608,12 @@ void GpuAgent::InitDma() {
     if (sdma_override != Flag::SDMA_DEFAULT) use_sdma = (sdma_override == Flag::SDMA_ENABLE);
 
     if (use_sdma && (HSA_PROFILE_BASE == profile_)) {
+      // On gfx90a ensure that HostToDevice queue is created first and so is placed on SDMA0.
+      if ((!use_xgmi) && (!isHostToDev) && (isa_->GetMajorVersion() == 9) &&
+          (isa_->GetMinorVersion() == 0) && (isa_->GetStepping() == 10)) {
+        *blits_[BlitHostToDev];
+      }
+
       auto ret = CreateBlitSdma(use_xgmi);
       if (ret != nullptr) return ret;
     }
@@ -643,13 +649,14 @@ void GpuAgent::InitDma() {
     return ret;
   });
   blits_[BlitHostToDev].reset(
-      [blit_lambda, this]() { return blit_lambda(false, queues_[QueueBlitOnly]); });
+      [blit_lambda, this]() { return blit_lambda(false, queues_[QueueBlitOnly], true); });
   blits_[BlitDevToHost].reset(
-      [blit_lambda, this]() { return blit_lambda(false, queues_[QueueUtility]); });
+      [blit_lambda, this]() { return blit_lambda(false, queues_[QueueUtility], false); });
 
   // XGMI engines.
   for (uint32_t idx = DefaultBlitCount; idx < blit_cnt_; idx++) {
-    blits_[idx].reset([blit_lambda, this]() { return blit_lambda(true, queues_[QueueUtility]); });
+    blits_[idx].reset(
+        [blit_lambda, this]() { return blit_lambda(true, queues_[QueueUtility], false); });
   }
 
   // GWS queues.