Add setting to cache the GPUMemory mappings for pipeline upload

During pipeline upload the GpuMemory is allocated from the pool. This memory is mapped, pipeline is uploaded and then the memory is unmapped. In applications with lots of pipelines being created during rendering, this causes lots of map/unmap calls and slows down the pipeline upload. Add a setting that allows the cache of the GpuMemory mapping, so that it can be reused by another pipeline upload.
GPUOpen-Drivers · Jul 12, 2021 · 749cfb4 · 749cfb4
1 parent 1311c5c
commit 749cfb4
Show file tree

Hide file tree

Showing 6 changed files with 123 additions and 3 deletions.
diff --git a/src/core/g_palSettings.cpp b/src/core/g_palSettings.cpp
@@ -162,6 +162,7 @@ void SettingsLoader::SetupDefaults()
     m_settings.overlayReportMes = true;
     m_settings.mipGenUseFastPath = false;
     m_settings.useFp16GenMips = false;
+    m_settings.keepGPUMemPoolMappedForPipelineCreation = false;
     m_settings.tmzEnabled = true;
 #if PAL_DEVELOPER_BUILD
     m_settings.dbgHelperBits = 0x0;
@@ -610,6 +611,11 @@ void SettingsLoader::ReadSettings()
                            &m_settings.useFp16GenMips,
                            InternalSettingScope::PrivatePalKey);
 
+    static_cast<Pal::Device*>(m_pDevice)->ReadSetting(pkeepGPUMemPoolMappedForPipelineCreationStr,
+                           Util::ValueType::Boolean,
+                           &m_settings.keepGPUMemPoolMappedForPipelineCreation,
+                           InternalSettingScope::PrivatePalKey);
+
     static_cast<Pal::Device*>(m_pDevice)->ReadSetting(pTmzEnabledStr,
                            Util::ValueType::Boolean,
                            &m_settings.tmzEnabled,
@@ -655,6 +661,11 @@ void SettingsLoader::RereadSettings()
                            &m_settings.useFp16GenMips,
                            InternalSettingScope::PrivatePalKey);
 
+    static_cast<Pal::Device*>(m_pDevice)->ReadSetting(pkeepGPUMemPoolMappedForPipelineCreationStr,
+                           Util::ValueType::Boolean,
+                           &m_settings.keepGPUMemPoolMappedForPipelineCreation,
+                           InternalSettingScope::PrivatePalKey);
+
     static_cast<Pal::Device*>(m_pDevice)->ReadSetting(pUseDccStr,
                            Util::ValueType::Uint,
                            &m_settings.useDcc,
@@ -1100,6 +1111,11 @@ void SettingsLoader::InitSettingsInfo()
     info.valueSize = sizeof(m_settings.useFp16GenMips);
     m_settingsInfoMap.Insert(192229910, info);
 
+    info.type      = SettingType::Boolean;
+    info.pValuePtr = &m_settings.keepGPUMemPoolMappedForPipelineCreation;
+    info.valueSize = sizeof(m_settings.keepGPUMemPoolMappedForPipelineCreation);
+    m_settingsInfoMap.Insert(3985380410, info);
+
     info.type      = SettingType::Boolean;
     info.pValuePtr = &m_settings.tmzEnabled;
     info.valueSize = sizeof(m_settings.tmzEnabled);

diff --git a/src/core/g_palSettings.h b/src/core/g_palSettings.h
@@ -279,6 +279,7 @@ struct PalSettings : public Pal::DriverSettings
     bool                                        overlayReportMes;
     bool                                        mipGenUseFastPath;
     bool                                        useFp16GenMips;
+    bool                                        keepGPUMemPoolMappedForPipelineCreation;
     bool                                        tmzEnabled;
 #if PAL_DEVELOPER_BUILD
     uint64                                      dbgHelperBits;
@@ -378,6 +379,7 @@ static const char* pDebugForceResourceAdditionalPaddingStr = "#3601080919";
 static const char* pOverlayReportMesStr = "#1685803860";
 static const char* pMipGenUseFastPathStr = "#3353227045";
 static const char* pUseFp16GenMipsStr = "#192229910";
+static const char* pkeepGPUMemPoolMappedForPipelineCreationStr = "#3985380410";
 static const char* pTmzEnabledStr = "#2606194033";
 #if PAL_DEVELOPER_BUILD
 static const char* pDbgHelperBitsStr = "#3894710420";

diff --git a/src/core/hw/gfxip/pipeline.cpp b/src/core/hw/gfxip/pipeline.cpp
@@ -881,7 +881,16 @@ Result PipelineUploader::UploadUsingCpu(
     const SectionAddressCalculator& addressCalc,
     void**                          ppMappedPtr)
 {
-    Result result = m_pGpuMemory->Map(&m_pMappedPtr);
+    Result result;
+    if (m_pDevice->Settings().keepGPUMemPoolMappedForPipelineCreation == false)
+    {
+        result = m_pGpuMemory->Map(&m_pMappedPtr);
+    }
+    else
+    {
+        result = m_pDevice->MemMgr()->Map(m_pGpuMemory, &m_pMappedPtr);
+    }
+
     if (result == Result::Success)
     {
         m_pMappedPtr = VoidPtrInc(m_pMappedPtr, static_cast<size_t>(m_baseOffset));
@@ -1104,7 +1113,10 @@ Result PipelineUploader::End(
         else
         {
             PAL_ASSERT(m_pMappedPtr != nullptr);
-            result = m_pGpuMemory->Unmap();
+            if (m_pDevice->Settings().keepGPUMemPoolMappedForPipelineCreation == false)
+            {
+                m_pGpuMemory->Unmap();
+            }
         }
 
         m_pMappedPtr = nullptr;

diff --git a/src/core/internalMemMgr.cpp b/src/core/internalMemMgr.cpp
@@ -138,6 +138,18 @@ InternalMemMgr::InternalMemMgr(
 // Explicitly frees all GPU memory allocations.
 void InternalMemMgr::FreeAllocations()
 {
+
+    for (auto it = m_poolList.Begin(); it.Get() != nullptr; it.Next())
+    {
+        PAL_ASSERT((it.Get() != nullptr) && (it.Get()->pBuddyAllocator != nullptr));
+
+        if (it.Get()->mapped && it.Get()->pGpuMemory != nullptr)
+        {
+            it.Get()->pGpuMemory->Unmap();
+            it.Get()->mapped = false;
+        }
+    }
+
     // Delete the GPU memory objects using the references list
     while (m_references.NumElements() != 0)
     {
@@ -563,4 +575,55 @@ uint32 InternalMemMgr::GetReferencesCount()
     return static_cast<uint32>(m_references.NumElements());
 }
 
+Result InternalMemMgr::Map(
+        GpuMemory* pGpuMemory,
+        void** ppData)
+{
+    Util::MutexAuto allocatorLock(&m_allocatorLock); // Ensure thread-safety using the lock
+    PAL_ASSERT(pGpuMemory != nullptr);
+
+    Result result = Result::ErrorInvalidValue;
+    if (pGpuMemory->WasBuddyAllocated())
+    {
+        // Try to find the allocation in the pool list
+        for (auto it = m_poolList.Begin(); it.Get() != nullptr; it.Next())
+        {
+            GpuMemoryPool* pPool = it.Get();
+
+            PAL_ASSERT((pPool->pGpuMemory != nullptr) && (pPool->pBuddyAllocator != nullptr));
+
+            if (pPool->pGpuMemory == pGpuMemory)
+            {
+                if (!pPool->mapped)
+                {
+                    result = pPool->pGpuMemory->Map(&pPool->pData);
+                    if (result != Result::Success) {
+                        break;
+                    }
+                    pPool->mapped = true;
+                }
+                *ppData = pPool->pData;
+                result = Result::Success;
+                break;
+            }
+        }
+
+        // If we didn't find the allocation in the pool list then something went wrong with the allocation scheme
+        PAL_ASSERT(result == Result::Success);
+    }
+    else
+    {
+        result = pGpuMemory->Map(ppData);
+    }
+
+    return result;
+}
+
+Result InternalMemMgr::Unmap(
+        GpuMemory* pGpuMemory)
+{
+    return Result::Success;
+}
+
+
 } // Pal
diff --git a/src/core/internalMemMgr.h b/src/core/internalMemMgr.h
@@ -56,6 +56,8 @@ struct GpuMemoryPool
     uint64                          pagingFenceVal;         // Paging fence value
 
     Util::BuddyAllocator<Platform>* pBuddyAllocator;        // Buddy allocator used for the suballocation
+    bool                            mapped;                 // flag to check if this pool is already mapped
+    void*                           pData;                  // address of the already existing mapping
 };
 
 // =====================================================================================================================
@@ -115,6 +117,14 @@ class InternalMemMgr
     // Number of all allocations in the reference list. Note that this function takes the reference list lock.
     uint32 GetReferencesCount();
 
+    Result Map(
+        GpuMemory* pGpuMemory,
+        void** ppData);
+
+    Result Unmap(
+        GpuMemory* pGpuMemory);
+
+
 private:
     Result AllocateBaseGpuMem(
         const GpuMemoryCreateInfo&          createInfo,

diff --git a/src/core/settings_core.json b/src/core/settings_core.json
@@ -1881,6 +1881,23 @@
       "VariableName": "useFp16GenMips",
       "Description": "If mipGenUseFastPath == true and this is true - use the fp16 single-pass GenMips compute pass."
     },
+    {
+      "Name": "keepGPUMemPoolMappedForPipelineCreation",
+      "Tags": [
+        "Resource Settings",
+        "Performance"
+      ],
+      "Defaults": {
+        "Default": false
+      },
+      "Flags": {
+        "RereadSetting": true
+      },
+      "Scope": "PrivatePalKey",
+      "Type": "bool",
+      "VariableName": "keepGPUMemPoolMappedForPipelineCreation",
+      "Description": "If keepGPUMemPoolMappedForPipelineCreation == true the mapped gpu memory for pipeline creation will not be unmapped. This mapped memory can be reused later by another pipeline creation."
+    },
     {
       "Name": "TmzEnabled",
       "Tags": [
@@ -2019,4 +2036,4 @@
       "Description": "Maximum string length for a miscellaneous string setting"
     }
   ]
-}
+}