swr/rast: FP consistency between POSH/RENDER pipes

- Ensure all threads have optimal floating-point control state
- Disable auto-generation of fused FP ops for VERTEX shader stage
- Disable "fast" FP ops for VERTEX shader stage

Reviewed-by: Bruce Cherniak <[email protected]>

4 files changed, 33 insertions, 11 deletions

diff --git a/src/gallium/drivers/swr/rasterizer/common/os.h b/src/gallium/drivers/swr/rasterizer/common/os.h
index 314d8184374..b00beeb36dd 100644
--- a/src/gallium/drivers/swr/rasterizer/common/os.h
+++ b/src/gallium/drivers/swr/rasterizer/common/os.h
@@ -294,4 +294,25 @@ int SWR_API
             std::string* pOptStdErr     = nullptr,   ///< (Optional Out) Standard Error text
             const std::string* pOptStdIn = nullptr); ///< (Optional In) Standard Input text
 
+
+/// Helper for setting up FP state
+/// @returns old csr state
+static INLINE uint32_t SetOptimalVectorCSR()
+{
+    uint32_t oldCSR = _mm_getcsr();
+
+    uint32_t newCSR = (oldCSR & ~(_MM_ROUND_MASK | _MM_DENORMALS_ZERO_MASK | _MM_FLUSH_ZERO_MASK));
+    newCSR |= (_MM_ROUND_NEAREST | _MM_FLUSH_ZERO_ON | _MM_DENORMALS_ZERO_ON);
+    _mm_setcsr(newCSR);
+
+    return oldCSR;
+}
+
+/// Set Vector CSR state.
+/// @param csrState - should be value returned from SetOptimalVectorCSR()
+static INLINE void RestoreVectorCSR(uint32_t csrState)
+{
+    _mm_setcsr(csrState);
+}
+
 #endif //__SWR_OS_H__
diff --git a/src/gallium/drivers/swr/rasterizer/core/api.cpp b/src/gallium/drivers/swr/rasterizer/core/api.cpp
index acbc7e077b1..203a74bdb24 100644
--- a/src/gallium/drivers/swr/rasterizer/core/api.cpp
+++ b/src/gallium/drivers/swr/rasterizer/core/api.cpp
@@ -250,9 +250,7 @@ void QueueWork(SWR_CONTEXT* pContext)
 
     if (pContext->threadInfo.SINGLE_THREADED)
     {
-        // flush denormals to 0
-        uint32_t mxcsr = _mm_getcsr();
-        _mm_setcsr(mxcsr | _MM_FLUSH_ZERO_ON | _MM_DENORMALS_ZERO_ON);
+        uint32_t mxcsr = SetOptimalVectorCSR();
 
         if (IsDraw)
         {
@@ -274,7 +272,7 @@ void QueueWork(SWR_CONTEXT* pContext)
         }
 
         // restore csr
-        _mm_setcsr(mxcsr);
+        RestoreVectorCSR(mxcsr);
     }
     else
     {
diff --git a/src/gallium/drivers/swr/rasterizer/core/frontend.cpp b/src/gallium/drivers/swr/rasterizer/core/frontend.cpp
index 8bc97c743eb..24db5275795 100644
--- a/src/gallium/drivers/swr/rasterizer/core/frontend.cpp
+++ b/src/gallium/drivers/swr/rasterizer/core/frontend.cpp
@@ -1840,10 +1840,10 @@ void ProcessDraw(SWR_CONTEXT* pContext, DRAW_CONTEXT* pDC, uint32_t workerId, vo
         {
             vIndex = _simd16_add_epi32(_simd16_set1_epi32(work.startVertexID), vScale);
 
-            fetchInfo_lo.xpIndices = 
-                pDC->pContext->pfnMakeGfxPtr(GetPrivateState(pDC), &vIndex);
-            fetchInfo_hi.xpIndices = 
-                pDC->pContext->pfnMakeGfxPtr(GetPrivateState(pDC), &vIndex + KNOB_SIMD_WIDTH * sizeof(int32_t)); // 1/2 of KNOB_SIMD16_WIDTH
+            fetchInfo_lo.xpIndices = pDC->pContext->pfnMakeGfxPtr(GetPrivateState(pDC), &vIndex);
+            fetchInfo_hi.xpIndices = pDC->pContext->pfnMakeGfxPtr(
+                GetPrivateState(pDC),
+                &vIndex + KNOB_SIMD_WIDTH * sizeof(int32_t)); // 1/2 of KNOB_SIMD16_WIDTH
         }
 
         fetchInfo_lo.CurInstance = instanceNum;
diff --git a/src/gallium/drivers/swr/rasterizer/core/threads.cpp b/src/gallium/drivers/swr/rasterizer/core/threads.cpp
index 4523616cba0..e30c1170568 100644
--- a/src/gallium/drivers/swr/rasterizer/core/threads.cpp
+++ b/src/gallium/drivers/swr/rasterizer/core/threads.cpp
@@ -421,9 +421,9 @@ INLINE void UpdateClientStats(SWR_CONTEXT* pContext, uint32_t workerId, DRAW_CON
     for (uint32_t i = 0; i < pContext->NumWorkerThreads; ++i)
     {
         stats.DepthPassCount += dynState.pStats[i].DepthPassCount;
-
         stats.PsInvocations += dynState.pStats[i].PsInvocations;
         stats.CsInvocations += dynState.pStats[i].CsInvocations;
+
     }
 
 
@@ -439,6 +439,10 @@ INLINE void ExecuteCallbacks(SWR_CONTEXT* pContext, uint32_t workerId, DRAW_CONT
         pDC->retireCallback.pfnCallbackFunc(pDC->retireCallback.userData,
                                             pDC->retireCallback.userData2,
                                             pDC->retireCallback.userData3);
+
+        // Callbacks to external code *could* change floating point control state
+        // Reset our optimal flags
+        SetOptimalVectorCSR();
     }
 }
 
@@ -870,8 +874,7 @@ DWORD workerThreadMain(LPVOID pData)
     uint32_t numaNode = pThreadData->numaId - pContext->threadInfo.BASE_NUMA_NODE;
     uint32_t numaMask = pContext->threadPool.numaMask;
 
-    // flush denormals to 0
-    _mm_setcsr(_mm_getcsr() | _MM_FLUSH_ZERO_ON | _MM_DENORMALS_ZERO_ON);
+    SetOptimalVectorCSR();
 
     // Track tiles locked by other threads. If we try to lock a macrotile and find its already
     // locked then we'll add it to this list so that we don't try and lock it again.