Add Futhark implementation (#146)

* Add Futhark.

.github/workflows/main.yaml

@@ -226,3 +226,14 @@ jobs:
       - name: Test compile hipsycl @ CMake 3.24
         if: ${{ ! cancelled() }}
         run: ./src/ci-test-compile.sh ./build hipsycl all ${{ env.CMAKE_3_24_BIN }}
+
+  test-futhark:
+    runs-on: ubuntu-22.04
+    steps:
+      - uses: actions/checkout@v4
+      - name: Prepare Futhark compiler
+        uses: diku-dk/install-futhark@HEAD
+        with:
+          version: 'latest'
+      - run: cmake -Bbuild -H. -DMODEL=futhark -DFUTHARK_BACKEND=multicore
+      - run: cmake --build build

CMakeLists.txt

@@ -4,7 +4,7 @@ if (CMAKE_VERSION VERSION_GREATER_EQUAL "3.24.0")
     cmake_policy(SET CMP0135 NEW)
 endif ()
 
-project(BabelStream VERSION 4.0 LANGUAGES CXX)
+project(BabelStream VERSION 4.0 LANGUAGES CXX C)
 
 # uncomment for debugging build issues:
 #set(CMAKE_VERBOSE_MAKEFILE ON)
@@ -162,6 +162,7 @@ register_model(acc ACC ACCStream.cpp)
 register_model(raja USE_RAJA RAJAStream.cpp)
 register_model(tbb TBB TBBStream.cpp)
 register_model(thrust THRUST ThrustStream.cu) # Thrust uses cu, even for rocThrust
+register_model(futhark FUTHARK FutharkStream.cpp)
 
 
 set(USAGE ON CACHE BOOL "Whether to print all custom flags for the selected model")

README.md

@@ -41,6 +41,7 @@ BabelStream is currently implemented in the following parallel programming model
 - SYCL and SYCL2020 (USM and accessors)
 - TBB
 - Thrust (via CUDA or HIP)
+- Futhark
 
 This project also contains implementations in alternative languages with different build systems:
 * Julia - [JuliaStream.jl](./src/julia/JuliaStream.jl)
@@ -101,7 +102,7 @@ The source for each model's implementations are located in `./src/<model>`.
 
 Currently available models are:
 ```
-omp;ocl;std-data;std-indices;std-ranges;hip;cuda;kokkos;sycl;sycl2020;acc;raja;tbb;thrust
+omp;ocl;std-data;std-indices;std-ranges;hip;cuda;kokkos;sycl;sycl2020;acc;raja;tbb;thrust;futhark
 ```
 
 #### Overriding default flags

src/futhark/FutharkStream.cpp

@@ -0,0 +1,212 @@
+// Copyright (c) 2015-16 Tom Deakin, Simon McIntosh-Smith,
+// University of Bristol HPC
+// Copyright (c) 2022 Troels Henriksen
+// University of Copenhagen
+//
+// For full license terms please see the LICENSE file distributed with this
+// source code
+
+#include <cstdlib>  // For aligned_alloc
+#include <string>
+#include "FutharkStream.h"
+
+template <class T>
+FutharkStream<T>::FutharkStream(const int ARRAY_SIZE, int device)
+{
+  this->array_size = ARRAY_SIZE;
+  this->cfg = futhark_context_config_new();
+  this->device = "#" + std::to_string(device);
+#if defined(FUTHARK_BACKEND_cuda) || defined(FUTHARK_BACKEND_opencl)
+  futhark_context_config_set_device(cfg, this->device.c_str());
+#endif
+  this->ctx = futhark_context_new(cfg);
+  this->a = NULL;
+  this->b = NULL;
+  this->c = NULL;
+}
+
+template <>
+FutharkStream<float>::~FutharkStream()
+{
+  if (this->a) {
+    futhark_free_f32_1d(this->ctx, (futhark_f32_1d*)this->a);
+  }
+  if (this->b) {
+    futhark_free_f32_1d(this->ctx, (futhark_f32_1d*)this->b);
+  }
+  if (this->c) {
+    futhark_free_f32_1d(this->ctx, (futhark_f32_1d*)this->c);
+  }
+  futhark_context_free(this->ctx);
+  futhark_context_config_free(this->cfg);
+}
+
+template <>
+FutharkStream<double>::~FutharkStream()
+{
+  if (this->a) {
+    futhark_free_f64_1d(this->ctx, (futhark_f64_1d*)this->a);
+  }
+  if (this->b) {
+    futhark_free_f64_1d(this->ctx, (futhark_f64_1d*)this->b);
+  }
+  if (this->c) {
+    futhark_free_f64_1d(this->ctx, (futhark_f64_1d*)this->c);
+  }
+  futhark_context_free(this->ctx);
+  futhark_context_config_free(this->cfg);
+}
+
+template <>
+void FutharkStream<float>::init_arrays(float initA, float initB, float initC) {
+  int array_size = this->array_size;
+  float *a = new float[array_size];
+  float *b = new float[array_size];
+  float *c = new float[array_size];
+  for (int i = 0; i < array_size; i++) {
+    a[i] = initA;
+    b[i] = initB;
+    c[i] = initC;
+  }
+  this->a = (futhark_f32_1d*)futhark_new_f32_1d(this->ctx, a, array_size);
+  this->b = (futhark_f32_1d*)futhark_new_f32_1d(this->ctx, b, array_size);
+  this->c = (futhark_f32_1d*)futhark_new_f32_1d(this->ctx, c, array_size);
+  futhark_context_sync(this->ctx);
+  delete[] a;
+  delete[] b;
+  delete[] c;
+}
+
+template <>
+void FutharkStream<double>::init_arrays(double initA, double initB, double initC) {
+  int array_size = this->array_size;
+  double *a = new double[array_size];
+  double *b = new double[array_size];
+  double *c = new double[array_size];
+  for (int i = 0; i < array_size; i++) {
+    a[i] = initA;
+    b[i] = initB;
+    c[i] = initC;
+  }
+  this->a = (futhark_f64_1d*)futhark_new_f64_1d(this->ctx, a, array_size);
+  this->b = (futhark_f64_1d*)futhark_new_f64_1d(this->ctx, b, array_size);
+  this->c = (futhark_f64_1d*)futhark_new_f64_1d(this->ctx, c, array_size);
+  futhark_context_sync(this->ctx);
+  delete[] a;
+  delete[] b;
+  delete[] c;
+}
+
+template <>
+void FutharkStream<float>::read_arrays(std::vector<float>& h_a, std::vector<float>& h_b, std::vector<float>& h_c) {
+  futhark_values_f32_1d(this->ctx, (futhark_f32_1d*)this->a, h_a.data());
+  futhark_values_f32_1d(this->ctx, (futhark_f32_1d*)this->b, h_b.data());
+  futhark_values_f32_1d(this->ctx, (futhark_f32_1d*)this->c, h_c.data());
+  futhark_context_sync(this->ctx);
+}
+
+template <>
+void FutharkStream<double>::read_arrays(std::vector<double>& h_a, std::vector<double>& h_b, std::vector<double>& h_c) {
+  futhark_values_f64_1d(this->ctx, (futhark_f64_1d*)this->a, h_a.data());
+  futhark_values_f64_1d(this->ctx, (futhark_f64_1d*)this->b, h_b.data());
+  futhark_values_f64_1d(this->ctx, (futhark_f64_1d*)this->c, h_c.data());
+  futhark_context_sync(this->ctx);
+}
+
+template <>
+void FutharkStream<float>::copy() {
+  futhark_free_f32_1d(this->ctx, (futhark_f32_1d*)this->c);
+  futhark_entry_f32_copy(this->ctx, (futhark_f32_1d**)&this->c, (futhark_f32_1d*)this->a);
+  futhark_context_sync(this->ctx);
+}
+
+template <>
+void FutharkStream<double>::copy() {
+  futhark_free_f64_1d(this->ctx, (futhark_f64_1d*)this->c);
+  futhark_entry_f64_copy(this->ctx, (futhark_f64_1d**)&this->c, (futhark_f64_1d*)this->a);
+  futhark_context_sync(this->ctx);
+}
+
+template <>
+void FutharkStream<float>::mul() {
+  futhark_free_f32_1d(this->ctx, (futhark_f32_1d*)this->b);
+  futhark_entry_f32_mul(this->ctx, (futhark_f32_1d**)&this->b, (futhark_f32_1d*)this->c);
+  futhark_context_sync(this->ctx);
+}
+
+template <>
+void FutharkStream<double>::mul() {
+  futhark_free_f64_1d(this->ctx, (futhark_f64_1d*)this->b);
+  futhark_entry_f64_mul(this->ctx, (futhark_f64_1d**)&this->b, (futhark_f64_1d*)this->c);
+  futhark_context_sync(this->ctx);
+}
+
+template <>
+void FutharkStream<float>::add() {
+  futhark_free_f32_1d(this->ctx, (futhark_f32_1d*)this->c);
+  futhark_entry_f32_add(this->ctx, (futhark_f32_1d**)&this->c, (futhark_f32_1d*)this->a, (futhark_f32_1d*)this->b);
+  futhark_context_sync(this->ctx);
+}
+
+template <>
+void FutharkStream<double>::add() {
+  futhark_free_f64_1d(this->ctx, (futhark_f64_1d*)this->c);
+  futhark_entry_f64_add(this->ctx, (futhark_f64_1d**)&this->c, (futhark_f64_1d*)this->a, (futhark_f64_1d*)this->b);
+  futhark_context_sync(this->ctx);
+}
+
+template <>
+void FutharkStream<float>::triad() {
+  futhark_free_f32_1d(this->ctx, (futhark_f32_1d*)this->c);
+  futhark_entry_f32_triad(this->ctx, (futhark_f32_1d**)&this->c, (futhark_f32_1d*)this->a, (futhark_f32_1d*)this->b);
+  futhark_context_sync(this->ctx);
+}
+
+template <>
+void FutharkStream<double>::triad() {
+  futhark_free_f64_1d(this->ctx, (futhark_f64_1d*)this->a);
+  futhark_entry_f64_triad(this->ctx, (futhark_f64_1d**)&this->a, (futhark_f64_1d*)this->b, (futhark_f64_1d*)this->c);
+  futhark_context_sync(this->ctx);
+}
+
+template <>
+void FutharkStream<float>::nstream() {
+  futhark_f32_1d* d;
+  futhark_entry_f32_triad(this->ctx, &d, (futhark_f32_1d*)this->a, (futhark_f32_1d*)this->b);
+  futhark_free_f32_1d(this->ctx, (futhark_f32_1d*)this->c);
+  this->c = d;
+  futhark_context_sync(this->ctx);
+}
+
+template <>
+void FutharkStream<double>::nstream() {
+  futhark_f64_1d* d;
+  futhark_entry_f64_triad(this->ctx, &d, (futhark_f64_1d*)this->a, (futhark_f64_1d*)this->b);
+  futhark_free_f64_1d(this->ctx, (futhark_f64_1d*)this->c);
+  this->c = d;
+  futhark_context_sync(this->ctx);
+}
+
+template <>
+float FutharkStream<float>::dot() {
+  float res;
+  futhark_entry_f32_dot(this->ctx, &res, (futhark_f32_1d*)this->a, (futhark_f32_1d*)this->b);
+  futhark_context_sync(this->ctx);
+  return res;
+}
+
+template <>
+double FutharkStream<double>::dot() {
+  double res;
+  futhark_entry_f64_dot(this->ctx, &res, (futhark_f64_1d*)this->a, (futhark_f64_1d*)this->b);
+  futhark_context_sync(this->ctx);
+  return res;
+}
+
+void listDevices(void)
+{
+  std::cout << "Device selection not supported." << std::endl;
+}
+
+template class FutharkStream<float>;
+template class FutharkStream<double>;

src/futhark/FutharkStream.h

@@ -0,0 +1,60 @@
+// Copyright (c) 2015-16 Tom Deakin, Simon McIntosh-Smith,
+// University of Bristol HPC
+// Copyright (c) 2022 Troels Henriksen
+// University of Copenhagen
+//
+// For full license terms please see the LICENSE file distributed with this
+// source code
+
+#pragma once
+
+#include <iostream>
+#include <stdexcept>
+
+#include "Stream.h"
+#include "babelstream.h"
+
+#if defined(FUTHARK_BACKEND_c)
+#define IMPLEMENTATION_STRING "Futhark (sequential)"
+#elif defined(FUTHARK_BACKEND_multicore)
+#define IMPLEMENTATION_STRING "Futhark (parallel CPU)"
+#elif defined(FUTHARK_BACKEND_opencl)
+#define IMPLEMENTATION_STRING "Futhark (OpencL)"
+#elif defined(FUTHARK_BACKEND_cuda)
+#define IMPLEMENTATION_STRING "Futhark (CUDA)"
+#else
+#define IMPLEMENTATION_STRING "Futhark (unknown backend)"
+#endif
+
+template <class T>
+class FutharkStream : public Stream<T>
+{
+protected:
+  // Size of arrays
+  int array_size;
+  // For device selection.
+  std::string device;
+
+  // Futhark stuff
+  struct futhark_context_config *cfg;
+  struct futhark_context *ctx;
+
+  // Device side arrays
+  void* a;
+  void* b;
+  void* c;
+
+public:
+  FutharkStream(const int, int);
+  ~FutharkStream();
+
+  virtual void copy() override;
+  virtual void add() override;
+  virtual void mul() override;
+  virtual void triad() override;
+  virtual void nstream() override;
+  virtual T dot() override;
+
+  virtual void init_arrays(T initA, T initB, T initC) override;
+  virtual void read_arrays(std::vector<T>& a, std::vector<T>& b, std::vector<T>& c) override;
+};

src/futhark/babelstream.fut

@@ -0,0 +1,62 @@
+module type kernels = {
+  type t
+  val copy [n] : [n]t -> *[n]t
+  val mul [n] : t -> [n]t -> [n]t
+  val add [n] : [n]t -> [n]t -> [n]t
+  val triad [n] : t -> [n]t -> [n]t -> [n]t
+  val dot [n] : [n]t -> [n]t -> t
+  -- Uniqueness allows nstream to mutate the 'a' array.
+  val nstream [n] : t -> *[n]t -> [n]t -> [n]t -> [n]t
+}
+
+module kernels (P: real) : kernels with t = P.t = {
+  type t = P.t
+  def copy = copy
+  def mul scalar c = map (P.*scalar) c
+  def add = map2 (P.+)
+  def triad scalar b c = map2 (P.+) b (map (P.* scalar) c)
+  def dot a b = reduce (P.+) (P.i32 0) (map2 (P.*) a b)
+  def nstream scalar a b c = map2 (P.+) a (map2 (P.+) b (map (P.*scalar) c))
+}
+
+module f32_kernels = kernels f32
+def f32_start_scalar : f32 = 0.4
+entry f32_copy = f32_kernels.copy
+entry f32_mul = f32_kernels.mul f32_start_scalar
+entry f32_add = f32_kernels.add
+entry f32_triad = f32_kernels.triad f32_start_scalar
+entry f32_nstream = f32_kernels.nstream f32_start_scalar
+entry f32_dot = f32_kernels.dot
+
+module f64_kernels = kernels f64
+def f64_start_scalar : f64 = 0.4
+entry f64_copy = f64_kernels.copy
+entry f64_mul = f64_kernels.mul f64_start_scalar
+entry f64_add = f64_kernels.add
+entry f64_triad = f64_kernels.triad f64_start_scalar
+entry f64_nstream = f64_kernels.nstream f64_start_scalar
+entry f64_dot = f64_kernels.dot
+
+-- ==
+-- entry: f32_copy f32_mul
+-- random input { [33554432]f32 }
+
+-- ==
+-- entry: f32_add f32_dot f32_triad
+-- random input { [33554432]f32 [33554432]f32 }
+
+-- ==
+-- entry: f32_nstream
+-- random input { [33554432]f32 [33554432]f32 [33554432]f32 }
+
+-- ==
+-- entry: f64_copy f64_mul
+-- random input { [33554432]f64 }
+
+-- ==
+-- entry: f64_add f64_dot f64_triad
+-- random input { [33554432]f64 [33554432]f64 }
+
+-- ==
+-- entry: f64_nstream
+-- random input { [33554432]f64 [33554432]f64 [33554432]f64 }

src/futhark/model.cmake

@@ -0,0 +1,55 @@
+# Use
+#
+#   cmake -Bbuild -H. -DMODEL=futhark -DFUTHARK_BACKEND=foo -DFUTHARK_COMPILER=foo/bar/bin/futhark
+#
+# to use the Futhark backend, where 'foo' must be one of 'multicore',
+# 'c', 'opencl', or 'cuda'.  Defaults to 'multicore'.
+#
+# Use -DFUTHARK_COMPILER to set the path to the Futhark compiler
+# binary.  Defaults to 'futhark' on the PATH.
+
+register_flag_optional(FUTHARK_BACKEND
+  "Use a specific Futhark backend, possible options are:
+         - c
+         - multicore
+         - opencl
+         - cuda"
+  "multicore")
+
+register_flag_optional(FUTHARK_COMPILER
+  "Absolute path to the Futhark compiler, defaults to the futhark compiler on PATH"
+  "futhark")
+
+macro(setup)
+  add_custom_command(
+    OUTPUT
+    ${CMAKE_CURRENT_BINARY_DIR}/babelstream.c
+    ${CMAKE_CURRENT_BINARY_DIR}/babelstream.h
+    COMMAND ${FUTHARK_COMPILER} ${FUTHARK_BACKEND}
+    --library src/futhark/babelstream.fut
+    -o ${CMAKE_CURRENT_BINARY_DIR}/babelstream
+    DEPENDS src/futhark/babelstream.fut
+    WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
+    VERBATIM
+  )
+  if (${FUTHARK_BACKEND} STREQUAL "c")
+    # Nothing to do.
+  elseif (${FUTHARK_BACKEND} STREQUAL "multicore")
+    set(THREADS_PREFER_PTHREAD_FLAG ON)
+    find_package(Threads REQUIRED)
+    register_link_library(Threads::Threads)
+  elseif (${FUTHARK_BACKEND} STREQUAL "opencl")
+    find_package(OpenCL REQUIRED)
+    register_link_library(OpenCL::OpenCL)
+  elseif (${FUTHARK_BACKEND} STREQUAL "cuda")
+    find_package(CUDA REQUIRED)
+    register_link_library("nvrtc" "cuda" "cudart")
+  else ()
+    message(FATAL_ERROR "Unsupported Futhark backend: ${FUTHARK_BACKEND}")
+  endif()
+endmacro()
+
+macro(setup_target)
+  target_sources(${EXE_NAME} PUBLIC "${CMAKE_CURRENT_BINARY_DIR}/babelstream.c")
+  include_directories("${CMAKE_CURRENT_BINARY_DIR}")
+endmacro()

src/main.cpp

@@ -49,6 +49,8 @@
 #include "SYCLStream2020.h"
 #elif defined(OMP)
 #include "OMPStream.h"
+#elif defined(FUTHARK)
+#include "FutharkStream.h"
 #endif
 
 // Default size of 2^25
@@ -298,6 +300,10 @@ void run()
   // Use the OpenMP implementation
   stream = new OMPStream<T>(ARRAY_SIZE, deviceIndex);
 
+#elif defined(FUTHARK)
+  // Use the Futhark implementation
+  stream = new FutharkStream<T>(ARRAY_SIZE, deviceIndex);
+
 #endif
 
   stream->init_arrays(startA, startB, startC);