From f261d24e34421b4977605d2f4e495358d6ae272c Mon Sep 17 00:00:00 2001
From: Sergei Shudler <sergei.shudler@intel.com>
Date: Tue, 19 Jul 2022 16:13:48 +0300
Subject: [PATCH] Optimized sample generation in Sobol sequence code

---
 src/luxrays/accelerators/embreeaccel.cpp  |  6 ++++++
 src/slg/engines/pathcpu/pathcputhread.cpp |  6 ++++++
 src/slg/samplers/sobolsequence.cpp        | 14 ++++++++++----
 3 files changed, 22 insertions(+), 4 deletions(-)

diff --git a/src/luxrays/accelerators/embreeaccel.cpp b/src/luxrays/accelerators/embreeaccel.cpp
index 7e6bf39e1..34cc9f37f 100644
--- a/src/luxrays/accelerators/embreeaccel.cpp
+++ b/src/luxrays/accelerators/embreeaccel.cpp
@@ -34,6 +34,12 @@ namespace luxrays {
 EmbreeAccel::EmbreeAccel(const Context *context) : ctx(context),
 		uniqueRTCSceneByMesh(MeshPtrCompare), uniqueGeomByMesh(MeshPtrCompare),
 		uniqueInstMatrixByMesh(MeshPtrCompare) {
+#if defined (_MSC_VER) || defined (__INTEL_COMPILER)
+	// Embree recommends settings these flags to improve performance
+	_MM_SET_FLUSH_ZERO_MODE(_MM_FLUSH_ZERO_ON);
+	_MM_SET_DENORMALS_ZERO_MODE(_MM_DENORMALS_ZERO_ON);
+#endif
+
 	embreeDevice = rtcNewDevice(NULL);
 	embreeScene = NULL;
 }
diff --git a/src/slg/engines/pathcpu/pathcputhread.cpp b/src/slg/engines/pathcpu/pathcputhread.cpp
index 814741046..89682bac7 100644
--- a/src/slg/engines/pathcpu/pathcputhread.cpp
+++ b/src/slg/engines/pathcpu/pathcputhread.cpp
@@ -37,6 +37,12 @@ PathCPURenderThread::PathCPURenderThread(PathCPURenderEngine *engine,
 }
 
 void PathCPURenderThread::RenderFunc() {
+#if defined (_MSC_VER) || defined (__INTEL_COMPILER)
+	// Embree recommends settings these flags in each thread to improve performance
+	_MM_SET_FLUSH_ZERO_MODE(_MM_FLUSH_ZERO_ON);
+	_MM_SET_DENORMALS_ZERO_MODE(_MM_DENORMALS_ZERO_ON);
+#endif
+
 	//SLG_LOG("[PathCPURenderEngine::" << threadIndex << "] Rendering thread started");
 
 	//--------------------------------------------------------------------------
diff --git a/src/slg/samplers/sobolsequence.cpp b/src/slg/samplers/sobolsequence.cpp
index f1f28a798..cb6e3626e 100644
--- a/src/slg/samplers/sobolsequence.cpp
+++ b/src/slg/samplers/sobolsequence.cpp
@@ -44,13 +44,19 @@ void SobolSequence::RequestSamples(const u_int size) {
 }
 
 u_int SobolSequence::SobolDimension(const u_int index, const u_int dimension) const {
-	const u_int offset = dimension * SOBOL_BITS;
+	u_int offset = dimension * SOBOL_BITS;
 	u_int result = 0;
 	u_int i = index;
 
-	for (u_int j = 0; i; i >>= 1, j++) {
-		if (i & 1)
-			result ^= directions[offset + j];
+	while (i) {
+		result ^= ((i & 1) * directions[offset]);
+		i >>= 1; offset++;
+		result ^= ((i & 1) * directions[offset]);
+		i >>= 1; offset++;
+		result ^= ((i & 1) * directions[offset]);
+		i >>= 1; offset++;
+		result ^= ((i & 1) * directions[offset]);
+		i >>= 1; offset++;
 	}
 
 	return result;