diff --git a/.github/workflows/build_and_run.yml b/.github/workflows/build_and_run.yml
index 37cd16f3..cf99d555 100644
--- a/.github/workflows/build_and_run.yml
+++ b/.github/workflows/build_and_run.yml
@@ -165,8 +165,14 @@ jobs:
           &$script_path
           echo "OCL_ICD_FILENAMES=$env:CONDA_PREFIX\Library\lib\intelocl64.dll" >> $env:GITHUB_ENV
 
-      - name: Run benchmarks
-        run: dpbench -i ${{env.WORKLOADS}} run -r2 --no-print-results || exit 1
+      - name: Run benchmarks fp64
+        run: dpbench -i ${{env.WORKLOADS}} run -r2 --no-print-results --precision=double || exit 1
 
-      - name: Generate report
+      - name: Generate report fp64
+        run: dpbench -i ${{env.WORKLOADS}} report || exit 1
+
+      - name: Run benchmarks fp32
+        run: dpbench -i ${{env.WORKLOADS}} run -r2 --no-print-results --precision=single || exit 1
+
+      - name: Generate report fp32
         run: dpbench -i ${{env.WORKLOADS}} report || exit 1
diff --git a/dpbench/benchmarks/black_scholes/black_scholes_sycl_native_ext/black_scholes_sycl/_black_scholes_kernel.hpp b/dpbench/benchmarks/black_scholes/black_scholes_sycl_native_ext/black_scholes_sycl/_black_scholes_kernel.hpp
index f1b7e2ba..ad99f5f4 100644
--- a/dpbench/benchmarks/black_scholes/black_scholes_sycl_native_ext/black_scholes_sycl/_black_scholes_kernel.hpp
+++ b/dpbench/benchmarks/black_scholes/black_scholes_sycl_native_ext/black_scholes_sycl/_black_scholes_kernel.hpp
@@ -8,8 +8,7 @@
 
 using namespace sycl;
 
-template <typename FpTy>
-class BlackScholesKernel;
+template <typename FpTy> class BlackScholesKernel;
 
 template <typename FpTy>
 void black_scholes_impl(queue Queue,
diff --git a/dpbench/benchmarks/black_scholes/black_scholes_sycl_native_ext/black_scholes_sycl/_black_scholes_sycl.cpp b/dpbench/benchmarks/black_scholes/black_scholes_sycl_native_ext/black_scholes_sycl/_black_scholes_sycl.cpp
index bf63c000..9c0202db 100644
--- a/dpbench/benchmarks/black_scholes/black_scholes_sycl_native_ext/black_scholes_sycl/_black_scholes_sycl.cpp
+++ b/dpbench/benchmarks/black_scholes/black_scholes_sycl_native_ext/black_scholes_sycl/_black_scholes_sycl.cpp
@@ -65,17 +65,20 @@ void black_scholes_sync(size_t /**/,
         throw std::runtime_error("Input arrays are not acceptable.");
 
     if (typenum == UAR_FLOAT) {
-        black_scholes_impl<float>(Queue, nopt, price.get_data<float>(),
-                        strike.get_data<float>(), t.get_data<float>(), rate,
-                        volatility, call.get_data<float>(),
-                        put.get_data<float>());
+        black_scholes_impl<float>(
+            Queue, nopt, price.get_data<float>(),
+            strike.get_data<float>(), t.get_data<float>(), rate,
+            volatility, call.get_data<float>(),
+            put.get_data<float>());
     } else if (typenum == UAR_DOUBLE) {
-        black_scholes_impl<double>(Queue, nopt, price.get_data<double>(),
-                        strike.get_data<double>(), t.get_data<double>(), rate,
-                        volatility, call.get_data<double>(),
-                        put.get_data<double>());
+        black_scholes_impl<double>(
+            Queue, nopt, price.get_data<double>(),
+            strike.get_data<double>(), t.get_data<double>(), rate,
+            volatility, call.get_data<double>(),
+            put.get_data<double>());
     } else {
-        throw std::runtime_error("Expected a double or single precision FP array.");
+        throw std::runtime_error(
+            "Expected a double or single precision FP array.");
     }
 }
 
diff --git a/dpbench/benchmarks/dbscan/dbscan_sycl_native_ext/dbscan_sycl/_dbscan_kernel.hpp b/dpbench/benchmarks/dbscan/dbscan_sycl_native_ext/dbscan_sycl/_dbscan_kernel.hpp
index a8710484..34ae8fc4 100644
--- a/dpbench/benchmarks/dbscan/dbscan_sycl_native_ext/dbscan_sycl/_dbscan_kernel.hpp
+++ b/dpbench/benchmarks/dbscan/dbscan_sycl_native_ext/dbscan_sycl/_dbscan_kernel.hpp
@@ -106,8 +106,7 @@ void getNeighborhood(size_t n,
     }
 }
 
-template<typename FpTy>
-class DBScanKernel;
+template<typename FpTy> class DBScanKernel;
 
 template <typename FpTy>
 size_t dbscan_impl(queue q,
@@ -135,8 +134,7 @@ size_t dbscan_impl(queue q,
                 size_t i2 = (i1 + 1 == n_samples ? n_samples : i1 + 1);
                 getNeighborhood<FpTy>(n_samples, n_features, data, i2 - i1,
                                       data + i1 * n_features, eps,
-                                      d_indices + i1 * n_samples,
-                                      d_sizes + i1);
+                                      d_indices + i1 * n_samples, d_sizes + i1);
             });
     });
 
diff --git a/dpbench/benchmarks/dbscan/dbscan_sycl_native_ext/dbscan_sycl/_dbscan_sycl.cpp b/dpbench/benchmarks/dbscan/dbscan_sycl_native_ext/dbscan_sycl/_dbscan_sycl.cpp
index df059b8a..8dd05aa3 100644
--- a/dpbench/benchmarks/dbscan/dbscan_sycl_native_ext/dbscan_sycl/_dbscan_sycl.cpp
+++ b/dpbench/benchmarks/dbscan/dbscan_sycl_native_ext/dbscan_sycl/_dbscan_sycl.cpp
@@ -45,10 +45,11 @@ size_t dbscan_sync(size_t n_samples,
 
     if (typenum == UAR_FLOAT) {
         return dbscan_impl<float>(queue, n_samples, n_features,
-                                data.get_data<float>(), eps, min_pts);
-    } else if (typenum == UAR_DOUBLE) {
+                                  data.get_data<float>(), eps, min_pts);
+    }
+    else if (typenum == UAR_DOUBLE) {
         return dbscan_impl<double>(queue, n_samples, n_features,
-                                data.get_data<double>(), eps, min_pts);
+                                   data.get_data<double>(), eps, min_pts);
     }
 
     throw std::runtime_error("Expected a double or single precision FP array.");
diff --git a/dpbench/benchmarks/l2_norm/l2_norm_sycl_native_ext/l2_norm_sycl/_l2_norm_kernel.hpp b/dpbench/benchmarks/l2_norm/l2_norm_sycl_native_ext/l2_norm_sycl/_l2_norm_kernel.hpp
index 2aae98ea..8f13d972 100644
--- a/dpbench/benchmarks/l2_norm/l2_norm_sycl_native_ext/l2_norm_sycl/_l2_norm_kernel.hpp
+++ b/dpbench/benchmarks/l2_norm/l2_norm_sycl_native_ext/l2_norm_sycl/_l2_norm_kernel.hpp
@@ -9,8 +9,7 @@
 
 using namespace sycl;
 
-template <typename FpTy>
-class theKernel;
+template <typename FpTy> class theKernel;
 
 template <typename FpTy>
 void l2_norm_impl(queue Queue,
diff --git a/dpbench/benchmarks/l2_norm/l2_norm_sycl_native_ext/l2_norm_sycl/_l2_norm_sycl.cpp b/dpbench/benchmarks/l2_norm/l2_norm_sycl_native_ext/l2_norm_sycl/_l2_norm_sycl.cpp
index dd4ce605..e9748a79 100644
--- a/dpbench/benchmarks/l2_norm/l2_norm_sycl_native_ext/l2_norm_sycl/_l2_norm_sycl.cpp
+++ b/dpbench/benchmarks/l2_norm/l2_norm_sycl_native_ext/l2_norm_sycl/_l2_norm_sycl.cpp
@@ -23,12 +23,13 @@ void l2_norm_sync(dpctl::tensor::usm_ndarray a, dpctl::tensor::usm_ndarray d)
 
     if (typenum == UAR_FLOAT) {
         l2_norm_impl(Queue, npoints, dims, a.get_data<float>(),
-                    d.get_data<float>());
+                     d.get_data<float>());
     } else if (typenum == UAR_DOUBLE) {
         l2_norm_impl(Queue, npoints, dims, a.get_data<double>(),
-                    d.get_data<double>());
+                     d.get_data<double>());
     } else {
-        throw std::runtime_error("Expected a double or single precision FP array.");
+        throw std::runtime_error(
+            "Expected a double or single precision FP array.");
     }
 }
 
diff --git a/dpbench/benchmarks/pairwise_distance/pairwise_distance_sycl_native_ext/pairwise_distance_sycl/_pairwise_distance_kernel.hpp b/dpbench/benchmarks/pairwise_distance/pairwise_distance_sycl_native_ext/pairwise_distance_sycl/_pairwise_distance_kernel.hpp
index 6e15959c..14d24dad 100644
--- a/dpbench/benchmarks/pairwise_distance/pairwise_distance_sycl_native_ext/pairwise_distance_sycl/_pairwise_distance_kernel.hpp
+++ b/dpbench/benchmarks/pairwise_distance/pairwise_distance_sycl_native_ext/pairwise_distance_sycl/_pairwise_distance_kernel.hpp
@@ -12,6 +12,8 @@ using namespace sycl;
 #define SQRT(x) sqrt(x)
 #endif
 
+template <typename FpTy> class PairwiseDistanceKernel;
+
 template <typename FpTy>
 void pairwise_distance_impl(queue Queue,
                             size_t npoints,
@@ -21,7 +23,7 @@ void pairwise_distance_impl(queue Queue,
                             FpTy *distance_op)
 {
     Queue.submit([&](handler &h) {
-        h.parallel_for<class PairwiseDistanceKernel>(
+        h.parallel_for<PairwiseDistanceKernel<FpTy>>(
             range<1>{npoints}, [=](id<1> myID) {
                 size_t i = myID[0];
                 for (size_t j = 0; j < npoints; j++) {
diff --git a/dpbench/benchmarks/pairwise_distance/pairwise_distance_sycl_native_ext/pairwise_distance_sycl/_pairwise_distance_sycl.cpp b/dpbench/benchmarks/pairwise_distance/pairwise_distance_sycl_native_ext/pairwise_distance_sycl/_pairwise_distance_sycl.cpp
index 3a5dc725..bb584206 100644
--- a/dpbench/benchmarks/pairwise_distance/pairwise_distance_sycl_native_ext/pairwise_distance_sycl/_pairwise_distance_sycl.cpp
+++ b/dpbench/benchmarks/pairwise_distance/pairwise_distance_sycl_native_ext/pairwise_distance_sycl/_pairwise_distance_sycl.cpp
@@ -51,16 +51,23 @@ void pairwise_distance_sync(dpctl::tensor::usm_ndarray X1,
     auto Queue = X1.get_queue();
     auto ndims = 3;
     auto npoints = X1.get_size() / ndims;
+    auto typenum = X1.get_typenum();
 
     if (!ensure_compatibility(X1, X2, D))
         throw std::runtime_error("Input arrays are not acceptable.");
 
-    if (X1.get_typenum() != UAR_DOUBLE || X2.get_typenum() != UAR_DOUBLE) {
-        throw std::runtime_error("Expected a double precision FP array.");
+    if (typenum == UAR_FLOAT) {
+        pairwise_distance_impl(Queue, npoints, ndims, X1.get_data<float>(),
+                               X2.get_data<float>(), D.get_data<float>());
+    }
+    else if (typenum == UAR_DOUBLE) {
+        pairwise_distance_impl(Queue, npoints, ndims, X1.get_data<double>(),
+                               X2.get_data<double>(), D.get_data<double>());
+    }
+    else {
+        throw std::runtime_error(
+            "Expected a double precision FP array.");
     }
-
-    pairwise_distance_impl(Queue, npoints, ndims, X1.get_data<double>(),
-                           X2.get_data<double>(), D.get_data<double>());
 }
 
 PYBIND11_MODULE(_pairwise_distance_sycl, m)
diff --git a/dpbench/benchmarks/rambo/rambo_sycl_native_ext/rambo_sycl/_rambo_kernel.hpp b/dpbench/benchmarks/rambo/rambo_sycl_native_ext/rambo_sycl/_rambo_kernel.hpp
index 857db1ee..c090a72c 100644
--- a/dpbench/benchmarks/rambo/rambo_sycl_native_ext/rambo_sycl/_rambo_kernel.hpp
+++ b/dpbench/benchmarks/rambo/rambo_sycl_native_ext/rambo_sycl/_rambo_kernel.hpp
@@ -12,8 +12,7 @@
 
 using namespace sycl;
 
-template <typename FpTy>
-class RamboKernel;
+template <typename FpTy> class RamboKernel;
 
 template <typename FpTy>
 event rambo_impl(queue Queue,
diff --git a/dpbench/benchmarks/rambo/rambo_sycl_native_ext/rambo_sycl/_rambo_sycl.cpp b/dpbench/benchmarks/rambo/rambo_sycl_native_ext/rambo_sycl/_rambo_sycl.cpp
index be60bb9c..c6a8e2ff 100644
--- a/dpbench/benchmarks/rambo/rambo_sycl_native_ext/rambo_sycl/_rambo_sycl.cpp
+++ b/dpbench/benchmarks/rambo/rambo_sycl_native_ext/rambo_sycl/_rambo_sycl.cpp
@@ -65,13 +65,16 @@ void rambo_sync(size_t nevts,
                             F1.get_data<float>(), Q1.get_data<float>(),
                             output.get_data<float>());
         e.wait();
-    } else if (typenum == UAR_DOUBLE) {
+    }
+    else if (typenum == UAR_DOUBLE) {
         auto e = rambo_impl(Queue, nevts, nout, C1.get_data<double>(),
                             F1.get_data<double>(), Q1.get_data<double>(),
                             output.get_data<double>());
         e.wait();
-    } else {
-        throw std::runtime_error("Expected a double or single precision FP array.");
+    }
+    else {
+        throw std::runtime_error(
+            "Expected a double or single precision FP array.");
     }
 }