SWDEV-451166 - Disable kernel args for non-XGMI if HDP flush register is invalid

Change-Id: I227e046e2b9cb25476a50240f5d070adbd558f21


[ROCm/clr commit: 96f5c44851]

projects/clr/hipamd/src/hip_graph_internal.cpp

@@ -396,9 +396,17 @@ hipError_t GraphExec::CaptureAQLPackets() {
     }
 
     if (device_kernarg_pool_ && !device->isXgmi()) {
-      *device->info().hdpMemFlushCntl = 1u;
-      if (*device->info().hdpMemFlushCntl != UINT32_MAX) {
-        LogError("Unexpected HDP Register readback value!");
+      if (device->info().hdpMemFlushCntl != nullptr) {
+        *device->info().hdpMemFlushCntl = 1u;
+        if (*device->info().hdpMemFlushCntl != UINT32_MAX) {
+          LogError("Unexpected HDP Register readback value!");
+        }
+      } else {
+        amd::Command* command = new amd::Marker(*capture_stream_, true);
+        if (command != nullptr) {
+          command->enqueue();
+          command->release();
+        }
       }
     }
 

projects/clr/rocclr/device/rocm/rocdevice.cpp

@@ -705,6 +705,27 @@ bool Device::create() {
     return false;
   }
 
+  setupCpuAgent();
+
+  // Get Agent HDP Flush Register Memory
+  hsa_amd_hdp_flush_t hdpInfo;
+  if (HSA_STATUS_SUCCESS !=
+      hsa_agent_get_info(bkendDevice_,
+        static_cast<hsa_agent_info_t>(HSA_AMD_AGENT_INFO_HDP_FLUSH), &hdpInfo)) {
+    LogPrintfError("Unable to determine HDP flush info for HSA device %s", agent_name);
+    return false;
+  }
+
+  info_.hdpMemFlushCntl = hdpInfo.HDP_MEM_FLUSH_CNTL;
+  info_.hdpRegFlushCntl = hdpInfo.HDP_REG_FLUSH_CNTL;
+
+  bool device_kernel_args = true;
+  if (!isXgmi_ && ((info_.hdpMemFlushCntl == nullptr) || (info_.hdpRegFlushCntl == nullptr))) {
+    LogWarning("Unable to determine HDP flush register address. "
+      "Device kernel arguments are not supported");
+    device_kernel_args = false;
+  }
+
   // Create HSA settings
   assert(!settings_);
   roc::Settings* hsaSettings = new roc::Settings();
@@ -713,7 +734,7 @@ bool Device::create() {
       !hsaSettings->create((agent_profile_ == HSA_PROFILE_FULL), isa->versionMajor(),
                            isa->versionMinor(), isa->versionStepping(),
                            isa->xnack() == amd::Isa::Feature::Enabled,
-                           coop_groups)) {
+                           coop_groups, device_kernel_args)) {
     LogPrintfError("Unable to create settings for HSA device %s (PCI ID %x)", agent_name,
                    pciDeviceId_);
     return false;
@@ -754,17 +775,6 @@ bool Device::create() {
   }
   info_.pciDomainID = pci_domain_id;
 
-  // Get Agent HDP Flush Register Memory
-  hsa_amd_hdp_flush_t hdpInfo;
-  if (HSA_STATUS_SUCCESS !=
-      hsa_agent_get_info(bkendDevice_,
-        static_cast<hsa_agent_info_t>(HSA_AMD_AGENT_INFO_HDP_FLUSH), &hdpInfo)) {
-    LogPrintfError("Unable to determine HDP flush info for HSA device %s", agent_name);
-    return false;
-  }
-  info_.hdpMemFlushCntl = hdpInfo.HDP_MEM_FLUSH_CNTL;
-  info_.hdpRegFlushCntl = hdpInfo.HDP_REG_FLUSH_CNTL;
-
   if (populateOCLDeviceConstants() == false) {
     LogPrintfError("populateOCLDeviceConstants failed for HSA device %s (PCI ID %x)", agent_name,
                    pciDeviceId_);
@@ -1257,7 +1267,6 @@ bool Device::populateOCLDeviceConstants() {
     engineAssignMap_[1 << i] = 0;
   }
 
-  setupCpuAgent();
 
   checkAtomicSupport();
 

projects/clr/rocclr/device/rocm/rocsettings.cpp

@@ -102,7 +102,8 @@ Settings::Settings() {
 
 // ================================================================================================
 bool Settings::create(bool fullProfile, uint32_t gfxipMajor, uint32_t gfxipMinor,
-                      uint32_t gfxStepping, bool enableXNACK, bool coop_groups) {
+                      uint32_t gfxStepping, bool enableXNACK, bool coop_groups,
+                      bool device_kernel_args) {
   customHostAllocator_ = false;
 
   if (fullProfile) {
@@ -168,7 +169,7 @@ bool Settings::create(bool fullProfile, uint32_t gfxipMajor, uint32_t gfxipMinor
   // Enable device kernel args for MI300* for now
   if (gfxipMajor == 9 && gfxipMinor == 4 &&
       (gfxStepping == 0 || gfxStepping == 1 || gfxStepping == 2)) {
-    device_kernel_args_ = HIP_FORCE_DEV_KERNARG;
+    device_kernel_args_ = HIP_FORCE_DEV_KERNARG && device_kernel_args;
   }
 
   if (gfxipMajor >= 10) {

projects/clr/rocclr/device/rocm/rocsettings.hpp

@@ -84,7 +84,8 @@ class Settings : public device::Settings {
 
   //! Creates settings
   bool create(bool fullProfile, uint32_t gfxipMajor, uint32_t gfxipMinor, uint32_t gfxStepping,
-              bool enableXNACK, bool coop_groups = false);
+              bool enableXNACK, bool coop_groups = false,
+              bool device_kernel_args = true);
 
  private:
   //! Disable copy constructor

projects/clr/rocclr/device/rocm/rocvirtual.cpp

@@ -3210,7 +3210,9 @@ bool VirtualGPU::submitKernelInternal(const amd::NDRangeContainer& sizes,
       }
     }
 
-    const auto pcieKernargs = !dev().isXgmi() && dev().settings().device_kernel_args_;
+    const auto pcieKernargs = !dev().isXgmi() &&
+                              dev().settings().device_kernel_args_ &&
+                              roc_device_.info().largeBar_;
     address argBuffer = hidden_arguments;
     bool isGraphCapture = vcmd != nullptr && vcmd->getCapturingState();