Merge remote-tracking branch 'fork/development' into simplify-niosv-sim

whitepau · whitepau · commit 1c187fa044ee · 2024-08-20T07:33:43.000-07:00
diff --git a/DirectProgramming/C++SYCL/N-BodyMethods/Nbody/src/GSimulation.cpp b/DirectProgramming/C++SYCL/N-BodyMethods/Nbody/src/GSimulation.cpp
@@ -205,7 +205,7 @@ void GSimulation::Start() {
   total_time_ = t0.Elapsed();
   total_flops_ = gflops * get_nsteps();
   av /= (double)(nf - 2);
-  dev = sqrt(dev / (double)(nf - 2) - av * av);
+  dev = std::sqrt(dev / (double)(nf - 2) - av * av);
 
   std::cout << "\n";
   std::cout << "# Total Time (s)     : " << total_time_ << "\n";
diff --git a/Libraries/oneMKL/monte_carlo_european_opt/src/montecarlo_main.cpp b/Libraries/oneMKL/monte_carlo_european_opt/src/montecarlo_main.cpp
@@ -73,7 +73,18 @@ void run()
         double total_time = 0.0;
 
         namespace mkl_rng = oneapi::mkl::rng;
-        mkl_rng::mcg59 engine(
+#if USE_PHILOX
+        using EngineTypeHost   = mkl_rng::philox4x32x10;
+        using EngineTypeDevice = mkl_rng::device::philox4x32x10<VEC_SIZE>;
+#elif USE_MRG
+        using EngineTypeHost   = mkl_rng::mrg32k3a;
+        using EngineTypeDevice = mkl_rng::device::mrg32k3a<VEC_SIZE>;
+#else
+        using EngineTypeHost   = mkl_rng::mcg59;
+        using EngineTypeDevice = mkl_rng::device::mcg59<VEC_SIZE>;
+#endif
+
+        EngineTypeHost engine(
 #if !INIT_ON_HOST
             my_queue,
 #else
@@ -86,18 +97,10 @@ void run()
         auto rng_event_3 = mkl_rng::generate(mkl_rng::uniform<DataType>(1.0, 5.0), engine, num_options, h_option_years_ptr);
 
         std::size_t n_states = global_size;
-        using EngineType =
-#if USE_PHILOX
-            mkl_rng::device::philox4x32x10<VEC_SIZE>;
-#elif USE_MRG
-            mkl_rng::device::mrg32k3a<VEC_SIZE>;
-#else
-            mkl_rng::device::mcg59<VEC_SIZE>;
-#endif
 
         // initialization needs only on first step
         auto deleter = [my_queue](auto* ptr) {sycl::free(ptr, my_queue);};
-        auto rng_states_uptr = std::unique_ptr<EngineType, decltype(deleter)>(sycl::malloc_device<EngineType>(n_states, my_queue), deleter);
+        auto rng_states_uptr = std::unique_ptr<EngineTypeDevice, decltype(deleter)>(sycl::malloc_device<EngineTypeDevice>(n_states, my_queue), deleter);
         auto* rng_states = rng_states_uptr.get();
 
         my_queue.parallel_for<k_initialize_state<DataType>>(
@@ -107,9 +110,9 @@ void run()
                 auto id = idx[0];
 #if USE_MRG
                 constexpr std::uint32_t seed = 12345u;
-                rng_states[id] = EngineType({ seed, seed, seed, seed, seed, seed }, { 0, (4096 * id) });
+                rng_states[id] = EngineTypeDevice({ seed, seed, seed, seed, seed, seed }, { 0, (4096 * id) });
 #else
-                rng_states[id] = EngineType(rand_seed, id * ITEMS_PER_WORK_ITEM * VEC_SIZE * block_n);
+                rng_states[id] = EngineTypeDevice(rand_seed, id * ITEMS_PER_WORK_ITEM * VEC_SIZE * block_n);
 #endif
         })
         .wait_and_throw();
diff --git a/Libraries/oneMKL/sparse_conjugate_gradient/sparse_cg.cpp b/Libraries/oneMKL/sparse_conjugate_gradient/sparse_cg.cpp
@@ -138,7 +138,7 @@ void run_sparse_cg_example(const sycl::device &dev)
     try {
         mkl::sparse::init_matrix_handle(&handle);
 
-        mkl::sparse::set_csr_data(handle, nrows, nrows, mkl::index_base::zero,
+        mkl::sparse::set_csr_data(main_queue, handle, nrows, nrows, mkl::index_base::zero,
                                           ia_buffer, ja_buffer, a_buffer);
 
         mkl::sparse::set_matrix_property(handle, mkl::sparse::property::symmetric);
@@ -174,13 +174,14 @@ void run_sparse_cg_example(const sycl::device &dev)
 
         // Calculation B^{-1}r_0
         {
+            fp alpha = 1.0;
             mkl::sparse::trsv(main_queue, mkl::uplo::lower,
-                                      mkl::transpose::nontrans, mkl::diag::nonunit,
-                                      handle, r_buffer, t_buffer);
+                                        mkl::transpose::nontrans,
+                                        mkl::diag::nonunit, alpha, handle, r_buffer, t_buffer);
             diagonal_mv<fp, intType>(main_queue, nrows, d_buffer, t_buffer);
             mkl::sparse::trsv(main_queue, mkl::uplo::upper,
-                                      mkl::transpose::nontrans, mkl::diag::nonunit,
-                                      handle, t_buffer, w_buffer);
+                                        mkl::transpose::nontrans,
+                                        mkl::diag::nonunit, alpha, handle, t_buffer, w_buffer);
         }
 
         mkl::blas::copy(main_queue, nrows, w_buffer, 1, p_buffer, 1);
@@ -225,13 +226,14 @@ void run_sparse_cg_example(const sycl::device &dev)
 
             // Calculate w_k = B^{-1}r_k
             {
+                fp alpha = 1.0;
                 mkl::sparse::trsv(main_queue, mkl::uplo::lower,
                                           mkl::transpose::nontrans,
-                                          mkl::diag::nonunit, handle, r_buffer, t_buffer);
+                                          mkl::diag::nonunit, alpha, handle, r_buffer, t_buffer);
                 diagonal_mv<fp, intType>(main_queue, nrows, d_buffer, t_buffer);
                 mkl::sparse::trsv(main_queue, mkl::uplo::upper,
                                           mkl::transpose::nontrans,
-                                          mkl::diag::nonunit, handle, t_buffer, w_buffer);
+                                          mkl::diag::nonunit, alpha, handle, t_buffer, w_buffer);
             }
 
             // Calculate current norm of correction
@@ -271,8 +273,9 @@ void run_sparse_cg_example(const sycl::device &dev)
     catch (std::exception const &e) {
         std::cout << "\t\tCaught exception:\n" << e.what() << std::endl;
     }
-    
-    mkl::sparse::release_matrix_handle(&handle);
+
+    mkl::sparse::release_matrix_handle(main_queue, &handle);
+    main_queue.wait();
 }
 
 //
diff --git a/Libraries/oneMKL/student_t_test/t_test.cpp b/Libraries/oneMKL/student_t_test/t_test.cpp
@@ -39,7 +39,7 @@ template <typename RealType>
 std::int32_t t_test(sycl::queue &q, sycl::buffer<RealType, 1> &r,
                     std::int64_t n, RealType expected_mean) {
   std::int32_t res = -1;
-  RealType sqrt_n_observations = sycl::sqrt(static_cast<RealType>(n));
+  RealType sqrt_n_observations = std::sqrt(static_cast<RealType>(n));
 
   // Create buffers to be passed inside oneMKL stats functions
   sycl::buffer<RealType, 1> mean_buf(sycl::range{1});
@@ -55,8 +55,8 @@ std::int32_t t_test(sycl::queue &q, sycl::buffer<RealType, 1> &r,
   // Create Host accessors and check the condition
   sycl::host_accessor mean_acc(mean_buf);
   sycl::host_accessor variance_acc(variance_buf);
-  if ((sycl::abs(mean_acc[0] - expected_mean) * sqrt_n_observations /
-       sycl::sqrt(variance_acc[0])) < static_cast<RealType>(threshold)) {
+  if ((std::abs(mean_acc[0] - expected_mean) * sqrt_n_observations /
+       std::sqrt(variance_acc[0])) < static_cast<RealType>(threshold)) {
     res = 1;
   } else {
     res = 0;
@@ -100,19 +100,19 @@ std::int32_t t_test(sycl::queue &q, sycl::buffer<RealType, 1> &r1,
   bool almost_equal = (variance1_acc[0] < 2 * variance2_acc[0]) ||
                       (variance2_acc[0] < 2 * variance1_acc[0]);
   if (almost_equal) {
-    if ((sycl::abs(mean1_acc[0] - mean2_acc[0]) /
-         sycl::sqrt((static_cast<RealType>(1.0) / static_cast<RealType>(n1) +
-                     static_cast<RealType>(1.0) / static_cast<RealType>(n2)) *
-                    ((n1 - 1) * (n1 - 1) * variance1_acc[0] +
-                     (n2 - 1) * (n2 - 1) * variance2_acc[0]) /
-                    (n1 + n2 - 2))) < static_cast<RealType>(threshold)) {
+    if ((std::abs(mean1_acc[0] - mean2_acc[0]) /
+         std::sqrt((static_cast<RealType>(1.0) / static_cast<RealType>(n1) +
+                    static_cast<RealType>(1.0) / static_cast<RealType>(n2)) *
+                  ((n1 - 1) * (n1 - 1) * variance1_acc[0] +
+                   (n2 - 1) * (n2 - 1) * variance2_acc[0]) /
+                   (n1 + n2 - 2))) < static_cast<RealType>(threshold)) {
       res = 1;
     } else {
       res = 0;
     }
   } else {
-    if ((sycl::abs(mean1_acc[0] - mean2_acc[0]) /
-         sycl::sqrt((variance1_acc[0] + variance2_acc[0]))) <
+    if ((std::abs(mean1_acc[0] - mean2_acc[0]) /
+         std::sqrt((variance1_acc[0] + variance2_acc[0]))) <
         static_cast<RealType>(threshold)) {
       res = 1;
     } else {
diff --git a/Libraries/oneMKL/student_t_test/t_test_usm.cpp b/Libraries/oneMKL/student_t_test/t_test_usm.cpp
@@ -39,7 +39,7 @@ template <typename RealType>
 std::int32_t t_test(sycl::queue& q, RealType* r, std::int64_t n,
                     RealType expected_mean) {
   std::int32_t res = -1;
-  RealType sqrt_n_observations = sycl::sqrt(static_cast<RealType>(n));
+  RealType sqrt_n_observations = std::sqrt(static_cast<RealType>(n));
 
   // Allocate memory to be passed inside oneMKL stats functions
   RealType* mean = sycl::malloc_shared<RealType>(1, q);
@@ -53,8 +53,8 @@ std::int32_t t_test(sycl::queue& q, RealType* r, std::int64_t n,
   oneapi::mkl::stats::central_moment(q, mean, dataset, variance);
   q.wait_and_throw();
   // Check the condition
-  if ((sycl::abs(mean[0] - expected_mean) * sqrt_n_observations /
-       sycl::sqrt(variance[0])) < static_cast<RealType>(threshold)) {
+  if ((std::abs(mean[0] - expected_mean) * sqrt_n_observations /
+       std::sqrt(variance[0])) < static_cast<RealType>(threshold)) {
     res = 1;
   } else {
     res = 0;
@@ -95,8 +95,8 @@ std::int32_t t_test(sycl::queue& q, RealType* r1, std::int64_t n1,
   bool almost_equal =
       (variance1[0] < 2 * variance2[0]) || (variance2[0] < 2 * variance1[0]);
   if (almost_equal) {
-    if ((sycl::abs(mean1[0] - mean2[0]) /
-         sycl::sqrt((static_cast<RealType>(1.0) / static_cast<RealType>(n1) +
+    if ((std::abs(mean1[0] - mean2[0]) /
+         std::sqrt((static_cast<RealType>(1.0) / static_cast<RealType>(n1) +
                      static_cast<RealType>(1.0) / static_cast<RealType>(n2)) *
                     ((n1 - 1) * (n1 - 1) * variance1[0] +
                      (n2 - 1) * (n2 - 1) * variance2[0]) /
@@ -106,8 +106,8 @@ std::int32_t t_test(sycl::queue& q, RealType* r1, std::int64_t n1,
       res = 0;
     }
   } else {
-    if ((sycl::abs(mean1[0] - mean2[0]) /
-         sycl::sqrt((variance1[0] + variance2[0]))) <
+    if ((std::abs(mean1[0] - mean2[0]) /
+         std::sqrt((variance1[0] + variance2[0]))) <
         static_cast<RealType>(threshold)) {
       res = 1;
     } else {
@@ -201,4 +201,4 @@ int main(int argc, char** argv) {
   std::cout << "T-test result with two input arrays: " << res1 << "\n\n";
 
   return 0;
-}
+}
