swr: [rasterizer core] Affinitize thread scratch space to numa node of worker

Acked-by: Brian Paul <[email protected]>

3 files changed, 16 insertions, 4 deletions

diff --git a/src/gallium/drivers/swr/rasterizer/core/api.cpp b/src/gallium/drivers/swr/rasterizer/core/api.cpp
index f0f7956b590..442cdd420f4 100644
--- a/src/gallium/drivers/swr/rasterizer/core/api.cpp
+++ b/src/gallium/drivers/swr/rasterizer/core/api.cpp
@@ -93,8 +93,16 @@ HANDLE SwrCreateContext(
     ///@note We could lazily allocate this but its rather small amount of memory.
     for (uint32_t i = 0; i < pContext->NumWorkerThreads; ++i)
     {
-        ///@todo Use numa API for allocations using numa information from thread data (if exists).
-        pContext->pScratch[i] = (uint8_t*)_aligned_malloc((32 * 1024), KNOB_SIMD_WIDTH * 4);
+#if defined(_WIN32)
+        uint32_t numaNode = pContext->threadPool.pThreadData ?
+            pContext->threadPool.pThreadData[i].numaId : 0;
+        pContext->pScratch[i] = (uint8_t*)VirtualAllocExNuma(
+            GetCurrentProcess(), nullptr, 32 * sizeof(KILOBYTE),
+            MEM_RESERVE | MEM_COMMIT, PAGE_READWRITE,
+            numaNode);
+#else
+        pContext->pScratch[i] = (uint8_t*)_aligned_malloc(32 * sizeof(KILOBYTE), KNOB_SIMD_WIDTH * 4);
+#endif
     }
 
     // State setup AFTER context is fully initialized
@@ -138,7 +146,11 @@ void SwrDestroyContext(HANDLE hContext)
     // Free scratch space.
     for (uint32_t i = 0; i < pContext->NumWorkerThreads; ++i)
     {
+#if defined(_WIN32)
+        VirtualFree(pContext->pScratch[i], 0, MEM_RELEASE);
+#else
         _aligned_free(pContext->pScratch[i]);
+#endif
     }
 
     delete(pContext->pHotTileMgr);
diff --git a/src/gallium/drivers/swr/rasterizer/core/arena.h b/src/gallium/drivers/swr/rasterizer/core/arena.h
index 67d81a44347..0241f5b900d 100644
--- a/src/gallium/drivers/swr/rasterizer/core/arena.h
+++ b/src/gallium/drivers/swr/rasterizer/core/arena.h
@@ -209,7 +209,7 @@ struct CachingAllocatorT : DefaultAllocator
 };
 typedef CachingAllocatorT<> CachingAllocator;
 
-template<typename T = DefaultAllocator, size_t BlockSizeT = (128 * 1024)>
+template<typename T = DefaultAllocator, size_t BlockSizeT = 128 * sizeof(KILOBYTE)>
 class TArena
 {
 public:
diff --git a/src/gallium/drivers/swr/rasterizer/core/backend.cpp b/src/gallium/drivers/swr/rasterizer/core/backend.cpp
index 7fb83edf169..ad0a5a07032 100644
--- a/src/gallium/drivers/swr/rasterizer/core/backend.cpp
+++ b/src/gallium/drivers/swr/rasterizer/core/backend.cpp
@@ -83,7 +83,7 @@ void ProcessComputeBE(DRAW_CONTEXT* pDC, uint32_t workerId, uint32_t threadGroup
     if (pDC->pSpillFill[workerId] == nullptr)
     {
         ///@todo Add state which indicates the spill fill size.
-        pDC->pSpillFill[workerId] = (uint8_t*)pDC->pArena->AllocAlignedSync(4096 * 1024, sizeof(float) * 8);
+        pDC->pSpillFill[workerId] = (uint8_t*)pDC->pArena->AllocAlignedSync(4 * sizeof(MEGABYTE), sizeof(float) * 8);
     }
 
     const API_STATE& state = GetApiState(pDC);