From 621f70b3724654e1f810de44a080c60cc4e9012a Mon Sep 17 00:00:00 2001
From: foreman <dl.constructicon@amd.com>
Date: Wed, 9 Nov 2016 10:55:17 -0500
Subject: [PATCH] P4 to Git Change 1338735 by gandryey@gera-w8 on 2016/11/09
 10:46:19

	SWDEV-86035 - Add PAL backend to OpenCL
	- Enable split logic in PAL

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palsettings.cpp#9 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palvirtual.cpp#38 edit
---
 rocclr/runtime/device/pal/palsettings.cpp | 38 +++++++++++++++++++++--
 rocclr/runtime/device/pal/palvirtual.cpp  | 12 +++++--
 2 files changed, 45 insertions(+), 5 deletions(-)

diff --git a/rocclr/runtime/device/pal/palsettings.cpp b/rocclr/runtime/device/pal/palsettings.cpp
index 12ff991110..c76d835ded 100644
--- a/rocclr/runtime/device/pal/palsettings.cpp
+++ b/rocclr/runtime/device/pal/palsettings.cpp
@@ -9,6 +9,11 @@
 
 #include <algorithm>
 
+#if defined(_WIN32)
+#include "Windows.h"
+#include "VersionHelpers.h"
+#endif
+
 namespace pal {
 
 /*! \brief information for adjusting maximum workload time
@@ -20,6 +25,9 @@ struct ModifyMaxWorkload
 {
     uint32_t time;          //!< max work load time (10x ms)
     uint32_t minorVersion;  //!< OS minor version
+#if defined(_WIN32)
+    BYTE comparisonOps; //!< Comparison option
+#endif
 };
 
 
@@ -104,7 +112,7 @@ Settings::Settings()
     numComputeRings_ = 0;
 
     minWorkloadTime_ = 1;       // 0.1 ms
-    maxWorkloadTime_ = 5000;    // 500 ms
+    maxWorkloadTime_ = 500000;  // 500 ms
 
     // Controls tiled images in persistent
     //!@note IOL for Linux doesn't setup tiling aperture in CMM/QS
@@ -171,6 +179,13 @@ Settings::create(
         if (!aiPlus_) {
             // APU systems for VI
             apuSystem_  = true;
+            // Fix BSOD/TDR issues observed on Stoney Win7/8.1/10
+            minWorkloadTime_ = 1000;
+            modifyMaxWorkload.time = 1000;         // Decided by experiment
+            modifyMaxWorkload.minorVersion = 1; // Win 7
+#if defined(_WIN32)
+            modifyMaxWorkload.comparisonOps = VER_EQUAL; // Limit to Win 7 only
+#endif
         }
     case Pal::AsicRevision::Iceland:
     case Pal::AsicRevision::Tonga:
@@ -192,8 +207,11 @@ Settings::create(
             // APU systems for CI
             apuSystem_  = true;
             // Fix BSOD/TDR issues observed on Kaveri Win7 (EPR#416903)
-            modifyMaxWorkload.time = 2500;      // 250ms
+            modifyMaxWorkload.time = 250000;      // 250ms
             modifyMaxWorkload.minorVersion = 1; // Win 7
+#if defined(_WIN32)
+            modifyMaxWorkload.comparisonOps = VER_EQUAL; // limit to Win 7
+#endif
         }
         // Fall through ...
     case Pal::AsicRevision::Bonaire:
@@ -258,6 +276,22 @@ Settings::create(
         return false;
     }
 
+#if defined(_WIN32)
+    if (modifyMaxWorkload.time > 0) {
+        OSVERSIONINFOEX versionInfo = { 0 };
+        versionInfo.dwOSVersionInfoSize = sizeof(OSVERSIONINFOEX);
+        versionInfo.dwMajorVersion = 6;
+        versionInfo.dwMinorVersion = modifyMaxWorkload.minorVersion;
+
+        DWORDLONG conditionMask = 0;
+        VER_SET_CONDITION(conditionMask, VER_MAJORVERSION, modifyMaxWorkload.comparisonOps);
+        VER_SET_CONDITION(conditionMask, VER_MINORVERSION, modifyMaxWorkload.comparisonOps);
+        if (VerifyVersionInfo(&versionInfo, VER_MAJORVERSION | VER_MINORVERSION, conditionMask)) {
+            maxWorkloadTime_ = modifyMaxWorkload.time;
+        }
+    }
+#endif // defined(_WIN32)
+
     // Enable atomics support
     enableExtension(ClKhrInt64BaseAtomics);
     enableExtension(ClKhrInt64ExtendedAtomics);
diff --git a/rocclr/runtime/device/pal/palvirtual.cpp b/rocclr/runtime/device/pal/palvirtual.cpp
index fc8b368131..917e7ecc64 100644
--- a/rocclr/runtime/device/pal/palvirtual.cpp
+++ b/rocclr/runtime/device/pal/palvirtual.cpp
@@ -468,7 +468,7 @@ VirtualGPU::DmaFlushMgmt::DmaFlushMgmt(const Device& dev)
     aluCnt_ = dev.info().simdPerCU_ * dev.info().simdWidth_ * dev.info().maxComputeUnits_;
     maxDispatchWorkload_ = static_cast<uint64_t>(dev.info().maxClockFrequency_) *
         // find time in us
-        100 * dev.settings().maxWorkloadTime_ *
+        dev.settings().maxWorkloadTime_ *
         aluCnt_;
     resetCbWorkload(dev);
 }
@@ -479,7 +479,7 @@ VirtualGPU::DmaFlushMgmt::resetCbWorkload(const Device& dev)
     cbWorkload_ = 0;
     maxCbWorkload_ = static_cast<uint64_t>(dev.info().maxClockFrequency_) *
         // find time in us
-        100 * dev.settings().minWorkloadTime_ * aluCnt_;
+        dev.settings().minWorkloadTime_ * aluCnt_;
 }
 
 void
@@ -1995,6 +1995,12 @@ VirtualGPU::submitKernelInternal(
         dbgManager->allocParamMemList(numParams);
     }
 
+    bool needFlush = false;
+    dmaFlushMgmt_.findSplitSize(dev(), sizes.global().product(), hsaKernel.aqlCodeSize());
+    if (dmaFlushMgmt().dispatchSplitSize() != 0) {
+        needFlush = true;
+    }
+
     size_t newOffset[3] = {0, 0, 0};
     size_t newGlobalSize[3] = {0, 0, 0};
 
@@ -2296,7 +2302,7 @@ VirtualGPU::submitKernelInternal(
         }
 
         // Update the global GPU event
-        setGpuEvent(gpuEvent);
+        setGpuEvent(gpuEvent, needFlush);
 
         if (!printfDbgHSA().output(*this, printfEnabled, hsaKernel.printfInfo())) {
             LogError("Couldn't read printf data from the buffer!\n");