Fix typos

unsupported/Eigen/CXX11/src/Tensor/README.md

@@ -898,7 +898,7 @@ containing the real part of the complex values of the original tensor.
 
 ### (Operation) imag()
 
-Returns a tensor with the same dimensions as the orginal tensor
+Returns a tensor with the same dimensions as the original tensor
 containing the imaginary part of the complex values of the original
 tensor.
 
@@ -910,7 +910,7 @@ exponent.
 
 The type of the exponent, Scalar, is always the same as the type of the
 tensor coefficients.  For example, only integer exponents can be used in
-conjuntion with tensors of integer values.
+conjunction with tensors of integer values.
 
 You can use cast() to lift this restriction.  For example this computes
 cubic roots of an int Tensor:

unsupported/Eigen/CXX11/src/Tensor/TensorContractionSycl.h

@@ -104,10 +104,10 @@ struct TTPanelSize {
   static EIGEN_CONSTEXPR StorageIndex TileSizeDimM = LocalThreadSizeM * WorkLoadPerThreadM;
   // TileSizeDimN: determines the tile size for the n dimension
   static EIGEN_CONSTEXPR StorageIndex TileSizeDimN = LocalThreadSizeN * WorkLoadPerThreadN;
-  // LoadPerThreadLhs: determines workload per thread for loading Lhs Tensor. This must be divisable by packetsize
+  // LoadPerThreadLhs: determines workload per thread for loading Lhs Tensor. This must be divisible by packetsize
   static EIGEN_CONSTEXPR StorageIndex LoadPerThreadLhs =
       ((TileSizeDimK * WorkLoadPerThreadM * WorkLoadPerThreadN) / (TileSizeDimN));
-  // LoadPerThreadRhs: determines workload per thread for loading Rhs Tensor. This must be divisable by packetsize
+  // LoadPerThreadRhs: determines workload per thread for loading Rhs Tensor. This must be divisible by packetsize
   static EIGEN_CONSTEXPR StorageIndex LoadPerThreadRhs =
       ((TileSizeDimK * WorkLoadPerThreadM * WorkLoadPerThreadN) / (TileSizeDimM));
   // BC : determines if supporting bank conflict is required
@@ -674,7 +674,7 @@ class TensorContractionKernel {
     for (StorageIndex wLPTN = 0; wLPTN < Properties::WorkLoadPerThreadN / PrivateNStride; wLPTN++) {
       // output leading dimension
       StorageIndex outputLD = 0;
-      // When local memory is used the PrivateNstride is always 1 because the coalesed access on N is loaded into Local
+      // When local memory is used the PrivateNstride is always 1 because the coalesced access on N is loaded into Local
       // memory and extracting from local to global is the same as no transposed version. However, when local memory is
       // not used and RHS is transposed we packetize the load for RHS.
       EIGEN_UNROLL_LOOP
@@ -898,7 +898,7 @@ class TensorContractionKernel {
     EIGEN_CONSTEXPR StorageIndex LSD = InputBlockProperties::is_rhs ? LSDR : LSDL;
     static_assert(((LocalOffset % (TileSizeDimNC / InputBlockProperties::nc_stride) == 0) &&
                    (LocalOffset % (Properties::TileSizeDimK / InputBlockProperties::c_stride) == 0)),
-                  " LocalOffset must be divisable by stride");
+                  " LocalOffset must be divisible by stride");
     const StorageIndex &NC = InputBlockProperties::is_rhs ? triple_dim.N : triple_dim.M;
     StorageIndex localThreadNC = local_index.first;
     StorageIndex localThreadC = local_index.second;

unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h

@@ -36,7 +36,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T* constCast(const T* data) {
 // used for referring to a Pointer on TensorEvaluator class. While the TensorExpression
 // is a device-agnostic type and need MakePointer class for type conversion,
 // the TensorEvaluator class can be specialized for a device, hence it is possible
-// to construct different types of temproray storage memory in TensorEvaluator
+// to construct different types of temporary storage memory in TensorEvaluator
 // for different devices by specializing the following StorageMemory class.
 template <typename T, typename device>
 struct StorageMemory : MakePointer<T> {};

unsupported/Eigen/CXX11/src/Tensor/TensorRandom.h

@@ -136,7 +136,7 @@ class UniformRandomGenerator {
     // thread but for SYCL ((CLOCK * 6364136223846793005ULL) + 0xda3e39cb94b95bdbULL) is passed to each thread and each
     // thread adds the  (global_thread_id* 6364136223846793005ULL) for itself only once, in order to complete the
     // construction similar to CUDA Therefore, the thread Id injection is not available at this stage.
-    // However when the operator() is called the thread ID will be available. So inside the opeator,
+    // However when the operator() is called the thread ID will be available. So inside the operator,
     // we add the thrreadID, BlockId,... (which is equivalent of i)
     // to the seed and construct the unique m_state per thead similar to cuda.
     m_exec_once = false;

unsupported/Eigen/CXX11/src/Tensor/TensorReductionSycl.h

@@ -433,7 +433,7 @@ struct PartialReducerLauncher {
     EIGEN_CONSTEXPR Index localRange = PannelParameters::LocalThreadSizeP * PannelParameters::LocalThreadSizeR;
     // In this step, we force the code not to be more than 2-step reduction:
     // Our empirical research shows that if each thread reduces at least 64
-    // elemnts individually, we get better performance. However, this can change
+    // elements individually, we get better performance. However, this can change
     // on different platforms. In this step we force the code not to be
     // morthan step reduction: Our empirical research shows that for inner_most
     // dim reducer, it is better to have 8 group in a reduce dimension for sizes
@@ -495,7 +495,7 @@ struct FullReducer<Self, Op, Eigen::SyclDevice, Vectorizable> {
     typename Self::Index inputSize = self.impl().dimensions().TotalSize();
     // In this step we force the code not to be more than 2-step reduction:
     // Our empirical research shows that if each thread reduces at least 512
-    // elemnts individually, we get better performance.
+    // elements individually, we get better performance.
     const Index reductionPerThread = 2048;
     // const Index num_work_group =
     Index reductionGroup = dev.getPowerOfTwo(