Merge pull request #44 from mhaseeb123/fft-gpu-new

GPU-based FFT + CMake cleanup + linalg::matrix_product

CMakeLists.txt

@@ -171,8 +171,14 @@ endif()
 
 # need to add appropriate flags for stdexec
 set(CMAKE_CXX_FLAGS
-    "${CMAKE_CXX_FLAGS} -stdpar=${STDPAR} -mp=${OMP} --gcc-toolchain=/opt/cray/pe/gcc/12.2.0/bin/ -pthread"
-)
+    "${CMAKE_CXX_FLAGS} -stdpar=${STDPAR} -mp=${OMP}")
+
+# add -cudalib=cublas if -stdpar=gpu
+if (STDPAR STREQUAL "gpu")
+  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DUSE_GPU")
+else()
+  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -UUSE_GPU")
+endif()
 
 # ##############################################################################
 # Add sub-directories

apps/1d_stencil/CMakeLists.txt

@@ -13,7 +13,6 @@ target_include_directories(
   PRIVATE ${CMAKE_BINARY_DIR} ${CMAKE_CURRENT_LIST_DIR}/../../include
           ${ARGPARSE_INCLUDE_DIR} ${MDSPAN_INCLUDE_DIR})
 
-if("${STDPAR}" STREQUAL "gpu")
   # TODO, fix cmake
   add_executable(stencil_stdexec stencil_stdexec.cpp)
   target_link_libraries(stencil_stdexec stdexec)
@@ -22,6 +21,7 @@ if("${STDPAR}" STREQUAL "gpu")
     PRIVATE ${CMAKE_BINARY_DIR} ${CMAKE_CURRENT_LIST_DIR}/../../include
             ${ARGPARSE_INCLUDE_DIR} ${MDSPAN_INCLUDE_DIR})
 
+if("${STDPAR}" STREQUAL "gpu")
   add_executable(stencil_cuda stencil_cuda.cpp)
   target_include_directories(
     stencil_cuda

apps/1d_stencil/stencil_stdexec.cpp

@@ -1,7 +1,7 @@
 /*
  * MIT License
  *
- * Copyright (c) 2023 Weile Wei 
+ * Copyright (c) 2023 Weile Wei
  * Copyright (c) 2023 The Regents of the University of California,
  * through Lawrence Berkeley National Laboratory (subject to receipt of any
  * required approvals from the U.S. Dept. of Energy).All rights reserved.
@@ -27,8 +27,10 @@
 //
 // This example provides a stdexec implementation for the 1D stencil code.
 #include <exec/static_thread_pool.hpp>
+#if defined(USE_GPU)
 #include <nvexec/multi_gpu_context.cuh>
 #include <nvexec/stream_context.cuh>
+#endif
 #include <stdexec/execution.hpp>
 
 #include "argparse/argparse.hpp"
@@ -45,7 +47,12 @@ struct args_params_t : public argparse::Args {
     bool& no_header = kwarg("no-header", "Do not print csv header row (default: false)").set_default(false);
     bool& help = flag("h, help", "print help");
     bool& time = kwarg("t, time", "print time").set_default(true);
-    std::string& sch = kwarg("sch", "stdexec scheduler: [options: cpu, gpu, multigpu]").set_default("cpu");
+    std::string& sch = kwarg("sch", "stdexec scheduler: [options: cpu"
+  #if defined (USE_GPU)
+                            ", gpu, multigpu"
+  #endif //USE_GPU
+                            "]").set_default("cpu");
+
     int& nthreads = kwarg("nthreads", "number of threads").set_default(std::thread::hardware_concurrency());
 };
 
@@ -121,12 +128,14 @@ int benchmark(args_params_t const& args) {
             case sch_t::CPU:
                 solution = step.do_work(exec::static_thread_pool(nthreads).get_scheduler(), size, nt);
                 break;
+#if defined(USE_GPU)
             case sch_t::GPU:
                 solution = step.do_work(nvexec::stream_context().get_scheduler(), size, nt);
                 break;
             case sch_t::MULTIGPU:
                 solution = step.do_work(nvexec::multi_gpu_stream_context().get_scheduler(), size, nt);
                 break;
+#endif // USE_GPU
             default:
                 std::cerr << "Unknown scheduler type encountered." << std::endl;
                 break;

apps/comm-study/comm-study-no-senders.cpp

@@ -74,7 +74,7 @@ auto work(P& A, P& B, P& Y, int N) {
 
   // get sum(Y) - one last memcpy (not USM) D2H
   sum +=
-      std::transform_reduce(std::execution::par_unseq, &Y[0], &Y[N], 0.0, std::plus<T>(), [](T &val){return val * val;});
+      std::reduce(std::execution::par_unseq, &Y[0], &Y[N], 0.0, std::plus<T>());
 
   return sum / N;
 }

apps/fft/CMakeLists.txt

@@ -2,6 +2,11 @@ project(fft LANGUAGES CXX)
 
 file(GLOB CPP_SOURCES "*.cpp")
 
+# add -cudalib=cublas if -stdpar=gpu
+if (STDPAR STREQUAL "gpu")
+  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -cudalib=cublas")
+endif()
+
 foreach(source_file ${CPP_SOURCES})
   if(NOT STDPAR STREQUAL "gpu")
     if("${source_file}" MATCHES ".*stdpar.*gpu.*" OR "${source_file}"
@@ -18,16 +23,19 @@ foreach(source_file ${CPP_SOURCES})
   add_executable(${exec_name} ${_EXCLUDE} ${source_file})
 
   # add dependency on argparse
-  add_dependencies(${exec_name} argparse magic_enum)
+  add_dependencies(${exec_name} argparse)
 
   set_source_files_properties(${source_file} PROPERTIES LANGUAGE CXX
                                                         LINKER_LANGUAGE CXX)
   target_include_directories(
     ${exec_name}
     PRIVATE ${CMAKE_BINARY_DIR} ${CMAKE_CURRENT_LIST_DIR}/../../include
-            ${ARGPARSE_INCLUDE_DIR} ${MAGICENUM_INCLUDE_DIR} ${MDSPAN_INCLUDE_DIR})
+            ${ARGPARSE_INCLUDE_DIR} ${MDSPAN_INCLUDE_DIR})
+
+  # uncomment only if using nvc++/23.1 - no need if nvc++/23.7
+  # target_link_directories(${exec_name} PRIVATE /opt/nvidia/hpc_sdk/Linux_x86_64/23.1/math_libs/lib64)
 
-  target_link_libraries(${exec_name} PUBLIC ${MPI_LIBS} stdexec)
+  target_link_libraries(${exec_name} PUBLIC ${MPI_LIBS} stdexec blas)
 
   set_target_properties(
     ${exec_name}

apps/fft/fft-serial.cpp

@@ -30,13 +30,74 @@
 
 #include "fft.hpp"
 
+//
+// serial fft function
+//
+[[nodiscard]] std::vector<data_t> fft_serial(const data_t *x, const int N, bool debug = false)
+{
+    std::vector<data_t> x_r(N);
+    std::vector<uint32_t> id(N);
+
+    // bit shift
+    int shift = 32 - ilog2(N);
+
+    // twiddle data in x[n]
+    for (int k = 0; k < N; k++)
+    {
+        id[k] = reverse_bits32(k) >> shift;
+        x_r[k] = x[id[k]];
+    }
+
+    // niterations
+    int niters = ilog2(N);
+    // local merge partition size
+    int lN = 2;
+
+    // set cout precision
+    std::cout << std::fixed << std::setprecision(1);
+
+    std::cout << "FFT progress: ";
+
+    for (int k = 0; k < niters; k++, lN*=2)
+    {
+        std::cout << (100.0 * k)/niters << "%.." << std::flush;
+
+        static Timer dtimer;
+
+        // number of partitions
+        int nparts = N/lN;
+        int tpp = lN/2;
+
+        if (debug)
+            dtimer.start();
+
+        // merge
+        for (int k = 0; k < N/2; k++)
+        {
+            // compute indices
+            int  e   = (k/tpp)*lN + (k % tpp);
+            auto o   = e + tpp;
+            auto i   = (k % tpp);
+            auto tmp = x_r[e] + x_r[o] * WNk(N, i * nparts);
+            x_r[o]     = x_r[e] - x_r[o] * WNk(N, i * nparts);
+            x_r[e]     = tmp;
+        }
+
+        if (debug)
+        std::cout << "This iter time: " << dtimer.stop() << " ms" << std::endl;
+    }
+
+    std::cout << "100%" << std::endl;
+    return x_r;
+}
+
 //
 // simulation
 //
 int main(int argc, char* argv[])
 {
     // parse params
-    fft_params_t args = argparse::parse<fft_params_t>(argc, argv);
+    const fft_params_t args = argparse::parse<fft_params_t>(argc, argv);
 
     // see if help wanted
     if (args.help)
@@ -64,9 +125,6 @@ int main(int argc, char* argv[])
         x_n.resize(N);
     }
 
-    // y[n] = fft(x[n]);
-    sig_t y_n(x_n);
-
     if (print_sig)
     {
         std::cout << std::endl << "x[n] = ";
@@ -81,7 +139,8 @@ int main(int argc, char* argv[])
     Timer timer;
 
     // fft radix-2 algorithm
-    fft_serial(y_n.data(), N, N);
+    // y[n] = fft(x[n]);
+    sig_t y_n(std::move(fft_serial(x_n.data(), N, args.debug)));
 
     // stop timer
     auto elapsed = timer.stop();
@@ -101,7 +160,7 @@ int main(int argc, char* argv[])
     // validate the recursively computed fft
     if (validate)
     {
-        if (x_n.isFFT(y_n))
+        if (x_n.isFFT(y_n, exec::static_thread_pool(std::thread::hardware_concurrency()).get_scheduler()))
             std::cout << "SUCCESS: y[n] == fft(x[n])" << std::endl;
         else
             std::cout << "FAILED: y[n] != fft(x[n])" << std::endl;