From 0ecb1b804fc9530cd393813a4491e66b23db27ab Mon Sep 17 00:00:00 2001
From: Kamil Diedrich <kamil.diedrich@intel.com>
Date: Fri, 4 Dec 2020 12:53:01 +0100
Subject: [PATCH] combine timestamps implementation

Signed-off-by: Kamil Diedrich <kamil.diedrich@intel.com>
---
 .../command_queue/gpgpu_walker_bdw_plus.inl   | 53 +++++++++++--------
 .../source/gen12lp/gpgpu_walker_gen12lp.cpp   | 34 ------------
 2 files changed, 31 insertions(+), 56 deletions(-)

diff --git a/opencl/source/command_queue/gpgpu_walker_bdw_plus.inl b/opencl/source/command_queue/gpgpu_walker_bdw_plus.inl
index ddeafd683b..bcc8964b30 100644
--- a/opencl/source/command_queue/gpgpu_walker_bdw_plus.inl
+++ b/opencl/source/command_queue/gpgpu_walker_bdw_plus.inl
@@ -243,16 +243,18 @@ void GpgpuWalkerHelper<GfxFamily>::dispatchProfilingCommandsStart(
         hwInfo,
         args);
 
-    //MI_STORE_REGISTER_MEM for context local timestamp
-    timeStampAddress = hwTimeStamps.getGpuAddress() + offsetof(HwTimeStamps, ContextStartTS);
+    if (!HwHelper::get(hwInfo.platform.eRenderCoreFamily).useOnlyGlobalTimestamps()) {
+        //MI_STORE_REGISTER_MEM for context local timestamp
+        timeStampAddress = hwTimeStamps.getGpuAddress() + offsetof(HwTimeStamps, ContextStartTS);
 
-    //low part
-    auto pMICmdLow = commandStream->getSpaceForCmd<MI_STORE_REGISTER_MEM>();
-    MI_STORE_REGISTER_MEM cmd = GfxFamily::cmdInitStoreRegisterMem;
-    adjustMiStoreRegMemMode(&cmd);
-    cmd.setRegisterAddress(GP_THREAD_TIME_REG_ADDRESS_OFFSET_LOW);
-    cmd.setMemoryAddress(timeStampAddress);
-    *pMICmdLow = cmd;
+        //low part
+        auto pMICmdLow = commandStream->getSpaceForCmd<MI_STORE_REGISTER_MEM>();
+        MI_STORE_REGISTER_MEM cmd = GfxFamily::cmdInitStoreRegisterMem;
+        adjustMiStoreRegMemMode(&cmd);
+        cmd.setRegisterAddress(GP_THREAD_TIME_REG_ADDRESS_OFFSET_LOW);
+        cmd.setMemoryAddress(timeStampAddress);
+        *pMICmdLow = cmd;
+    }
 }
 
 template <typename GfxFamily>
@@ -263,21 +265,28 @@ void GpgpuWalkerHelper<GfxFamily>::dispatchProfilingCommandsEnd(
     using MI_STORE_REGISTER_MEM = typename GfxFamily::MI_STORE_REGISTER_MEM;
 
     // PIPE_CONTROL for global timestamp
-    auto pPipeControlCmd = commandStream->getSpaceForCmd<PIPE_CONTROL>();
-    PIPE_CONTROL cmdPipeControl = GfxFamily::cmdInitPipeControl;
-    cmdPipeControl.setCommandStreamerStallEnable(true);
-    *pPipeControlCmd = cmdPipeControl;
+    uint64_t timeStampAddress = hwTimeStamps.getGpuAddress() + offsetof(HwTimeStamps, GlobalEndTS);
+    PipeControlArgs args;
+    MemorySynchronizationCommands<GfxFamily>::addPipeControlAndProgramPostSyncOperation(
+        *commandStream,
+        PIPE_CONTROL::POST_SYNC_OPERATION_WRITE_TIMESTAMP,
+        timeStampAddress,
+        0llu,
+        hwInfo,
+        args);
 
-    //MI_STORE_REGISTER_MEM for context local timestamp
-    uint64_t timeStampAddress = hwTimeStamps.getGpuAddress() + offsetof(HwTimeStamps, ContextEndTS);
+    if (!HwHelper::get(hwInfo.platform.eRenderCoreFamily).useOnlyGlobalTimestamps()) {
+        //MI_STORE_REGISTER_MEM for context local timestamp
+        uint64_t timeStampAddress = hwTimeStamps.getGpuAddress() + offsetof(HwTimeStamps, ContextEndTS);
 
-    //low part
-    auto pMICmdLow = commandStream->getSpaceForCmd<MI_STORE_REGISTER_MEM>();
-    MI_STORE_REGISTER_MEM cmd = GfxFamily::cmdInitStoreRegisterMem;
-    adjustMiStoreRegMemMode(&cmd);
-    cmd.setRegisterAddress(GP_THREAD_TIME_REG_ADDRESS_OFFSET_LOW);
-    cmd.setMemoryAddress(timeStampAddress);
-    *pMICmdLow = cmd;
+        //low part
+        auto pMICmdLow = commandStream->getSpaceForCmd<MI_STORE_REGISTER_MEM>();
+        MI_STORE_REGISTER_MEM cmd = GfxFamily::cmdInitStoreRegisterMem;
+        adjustMiStoreRegMemMode(&cmd);
+        cmd.setRegisterAddress(GP_THREAD_TIME_REG_ADDRESS_OFFSET_LOW);
+        cmd.setMemoryAddress(timeStampAddress);
+        *pMICmdLow = cmd;
+    }
 }
 
 template <typename GfxFamily>
diff --git a/opencl/source/gen12lp/gpgpu_walker_gen12lp.cpp b/opencl/source/gen12lp/gpgpu_walker_gen12lp.cpp
index ab33b249a8..4781b2d355 100644
--- a/opencl/source/gen12lp/gpgpu_walker_gen12lp.cpp
+++ b/opencl/source/gen12lp/gpgpu_walker_gen12lp.cpp
@@ -20,40 +20,6 @@ void GpgpuWalkerHelper<TGLLPFamily>::adjustMiStoreRegMemMode(MI_STORE_REG_MEM<TG
     storeCmd->setMmioRemapEnable(true);
 }
 
-template <>
-void GpgpuWalkerHelper<TGLLPFamily>::dispatchProfilingCommandsStart(
-    TagNode<HwTimeStamps> &hwTimeStamps,
-    LinearStream *commandStream,
-    const HardwareInfo &hwInfo) {
-    // PIPE_CONTROL for global timestamp
-    uint64_t timeStampAddress = hwTimeStamps.getGpuAddress() + offsetof(HwTimeStamps, GlobalStartTS);
-    PipeControlArgs args;
-    MemorySynchronizationCommands<TGLLPFamily>::addPipeControlAndProgramPostSyncOperation(
-        *commandStream,
-        PIPE_CONTROL::POST_SYNC_OPERATION_WRITE_TIMESTAMP,
-        timeStampAddress,
-        0llu,
-        hwInfo,
-        args);
-}
-
-template <>
-void GpgpuWalkerHelper<TGLLPFamily>::dispatchProfilingCommandsEnd(
-    TagNode<HwTimeStamps> &hwTimeStamps,
-    LinearStream *commandStream,
-    const HardwareInfo &hwInfo) {
-    // PIPE_CONTROL for global timestamp
-    uint64_t timeStampAddress = hwTimeStamps.getGpuAddress() + offsetof(HwTimeStamps, GlobalEndTS);
-    PipeControlArgs args;
-    MemorySynchronizationCommands<TGLLPFamily>::addPipeControlAndProgramPostSyncOperation(
-        *commandStream,
-        PIPE_CONTROL::POST_SYNC_OPERATION_WRITE_TIMESTAMP,
-        timeStampAddress,
-        0llu,
-        hwInfo,
-        args);
-}
-
 template <>
 void HardwareInterface<TGLLPFamily>::dispatchWorkarounds(
     LinearStream *commandStream,