swr: [rasterizer jitter] vpermps support

[mesa.git] / src / gallium / drivers / swr / rasterizer / jitter / builder_misc.cpp

diff --git a/src/gallium/drivers/swr/rasterizer/jitter/builder_misc.cpp b/src/gallium/drivers/swr/rasterizer/jitter/builder_misc.cpp
index 7ebaca05151f186da0fd2533dcd501eb472f0173..b55752c102556390fc6fd1955725caaca586d9c2 100644 (file)
--- a/src/gallium/drivers/swr/rasterizer/jitter/builder_misc.cpp
+++ b/src/gallium/drivers/swr/rasterizer/jitter/builder_misc.cpp
@@ -28,6 +28,8 @@
 * 
 ******************************************************************************/
 #include "builder.h"
+#include "common/rdtsc_buckets.h"
+
 #include "llvm/Support/DynamicLibrary.h"
 
 void __cdecl CallPrint(const char* fmt, ...);
@@ -774,8 +776,57 @@ Value *Builder::PERMD(Value* a, Value* idx)
     }
     else
     {
-        res = VSHUFFLE(a, a, idx);
+        if (isa<Constant>(idx))
+        {
+            res = VSHUFFLE(a, a, idx);
+        }
+        else
+        {
+            res = VUNDEF_I();
+            for (uint32_t l = 0; l < JM()->mVWidth; ++l)
+            {
+                Value* pIndex = VEXTRACT(idx, C(l));
+                Value* pVal = VEXTRACT(a, pIndex);
+                res = VINSERT(res, pVal, C(l));
+            }
+        }
+    }
+    return res;
+}
+
+//////////////////////////////////////////////////////////////////////////
+/// @brief Generate a VPERMPS operation (shuffle 32 bit float values 
+/// across 128 bit lanes) in LLVM IR.  If not supported on the underlying 
+/// platform, emulate it
+/// @param a - 256bit SIMD lane(8x32bit) of float values.
+/// @param idx - 256bit SIMD lane(8x32bit) of 3 bit lane index values
+Value *Builder::PERMPS(Value* a, Value* idx)
+{
+    Value* res;
+    // use avx2 permute instruction if available
+    if (JM()->mArch.AVX2())
+    {
+        // llvm 3.6.0 swapped the order of the args to vpermd
+        res = VPERMPS(idx, a);
+    }
+    else
+    {
+        if (isa<Constant>(idx))
+        {
+            res = VSHUFFLE(a, a, idx);
+        }
+        else
+        {
+            res = VUNDEF_F();
+            for (uint32_t l = 0; l < JM()->mVWidth; ++l)
+            {
+                Value* pIndex = VEXTRACT(idx, C(l));
+                Value* pVal = VEXTRACT(a, pIndex);
+                res = VINSERT(res, pVal, C(l));
+            }
+        }
     }
+
     return res;
 }
 
@@ -1447,3 +1498,39 @@ Value *Builder::VINSERTI128(Value* a, Value* b, Constant* imm8)
     return VSHUFFLE(a, inter, ConstantVector::get(idx2));
 #endif
 }
+
+// rdtsc buckets macros
+void Builder::RDTSC_START(Value* pBucketMgr, Value* pId)
+{
+    std::vector<Type*> args{
+        PointerType::get(mInt32Ty, 0),   // pBucketMgr
+        mInt32Ty                        // id
+    };
+
+    FunctionType* pFuncTy = FunctionType::get(Type::getVoidTy(JM()->mContext), args, false);
+    Function* pFunc = cast<Function>(JM()->mpCurrentModule->getOrInsertFunction("BucketManager_StartBucket", pFuncTy));
+    if (sys::DynamicLibrary::SearchForAddressOfSymbol("BucketManager_StartBucket") == nullptr)
+    {
+        sys::DynamicLibrary::AddSymbol("BucketManager_StartBucket", (void*)&BucketManager_StartBucket);
+    }
+
+    CALL(pFunc, { pBucketMgr, pId });
+}
+
+void Builder::RDTSC_STOP(Value* pBucketMgr, Value* pId)
+{
+    std::vector<Type*> args{
+        PointerType::get(mInt32Ty, 0),   // pBucketMgr
+        mInt32Ty                        // id
+    };
+
+    FunctionType* pFuncTy = FunctionType::get(Type::getVoidTy(JM()->mContext), args, false);
+    Function* pFunc = cast<Function>(JM()->mpCurrentModule->getOrInsertFunction("BucketManager_StopBucket", pFuncTy));
+    if (sys::DynamicLibrary::SearchForAddressOfSymbol("BucketManager_StopBucket") == nullptr)
+    {
+        sys::DynamicLibrary::AddSymbol("BucketManager_StopBucket", (void*)&BucketManager_StopBucket);
+    }
+
+    CALL(pFunc, { pBucketMgr, pId });
+}
+