arrayfire: 3.7.3 -> 3.9.0-pre

4 files changed, 206 insertions, 109 deletions

diff --git a/pkgs/development/libraries/arrayfire/default.nix b/pkgs/development/libraries/arrayfire/default.nix
index 2c1caf0bbb2b..5c52a43e4878 100644
--- a/pkgs/development/libraries/arrayfire/default.nix
+++ b/pkgs/development/libraries/arrayfire/default.nix
@@ -1,91 +1,183 @@
-{ lib
-, stdenv
-, fetchFromGitHub
+{ blas
+, boost
 , cmake
-, pkg-config
-, opencl-clhpp
-, ocl-icd
+, config
+, cudaPackages
+, fetchFromGitHub
 , fftw
 , fftwFloat
-, blas
+, fmt_9
+, forge
+, freeimage
+, git
+, gtest
 , lapack
-, boost
-, mesa
-, libGLU
+, lib
 , libGL
-, freeimage
+, mesa
+, ocl-icd
+, opencl-clhpp
+, pkg-config
 , python3
-, clfft
-, clblas
-, doxygen
-, buildDocs ? false
-, config
-, cudaSupport ? config.cudaSupport
-, cudatoolkit
-, darwin
+, span-lite
+, stdenv
+, doCheck ? false
+, withCPU ? true
+, withCuda ? config.cudaSupport
+, withOpenCL ? stdenv.isLinux
+, nvidiaComputeDrivers ? null
 }:
 
+assert blas.isILP64 == false;
+
 stdenv.mkDerivation rec {
   pname = "arrayfire";
-  version = "3.7.3";
+  version = "3.9.0-pre";
 
   src = fetchFromGitHub {
     owner = pname;
     repo = pname;
-    rev = "v${version}";
-    sha256 = "0gcbg6b6gs38xhks5pp0vkcqs89zl7rh9982jqlzsd0h724qddw0";
-    fetchSubmodules = true;
+    rev = "d2a66367d859cdb554f2374e29d39c88d5fff978";
+    hash = "sha256-9r1w0U9MvhduHwBpEWpqkrQPawd94EY3FAqSJghi09I=";
+  };
+
+  assets = fetchFromGitHub {
+    owner = pname;
+    repo = "assets";
+    rev = "cd08d749611b324012555ad6f23fd76c5465bd6c";
+    sha256 = "sha256-v4uhqPz1P1g1430FTmMp22xJS50bb5hZTeEX49GgMWg=";
+  };
+  clblast = fetchFromGitHub {
+    owner = "cnugteren";
+    repo = "CLBlast";
+    rev = "4500a03440e2cc54998c0edab366babf5e504d67";
+    sha256 = "sha256-I25ylQp6kHZx6Q7Ph5r3abWlQ6yeIHIDdS1eGCyArZ0=";
+  };
+  clfft = fetchFromGitHub {
+    owner = pname;
+    repo = "clfft";
+    rev = "760096b37dcc4f18ccd1aac53f3501a83b83449c";
+    sha256 = "sha256-vJo1YfC2AJIbbRj/zTfcOUmi0Oj9v64NfA9MfK8ecoY=";
+  };
+  glad = fetchFromGitHub {
+    owner = pname;
+    repo = "glad";
+    rev = "ef8c5508e72456b714820c98e034d9a55b970650";
+    sha256 = "sha256-u9Vec7XLhE3xW9vzM7uuf+b18wZsh/VMtGbB6nMVlno=";
+  };
+  threads = fetchFromGitHub {
+    owner = pname;
+    repo = "threads";
+    rev = "4d4a4f0384d1ac2f25b2c4fc1d57b9e25f4d6818";
+    sha256 = "sha256-qqsT9woJDtQvzuV323OYXm68pExygYs/+zZNmg2sN34=";
+  };
+  test-data = fetchFromGitHub {
+    owner = pname;
+    repo = "arrayfire-data";
+    rev = "a5f533d7b864a4d8f0dd7c9aaad5ff06018c4867";
+    sha256 = "sha256-AWzhsrDXyZrQN2bd0Ng/XlE8v02x7QWTiFTyaAuRXSw=";
+  };
+  cub = fetchFromGitHub {
+    owner = "NVIDIA";
+    repo = "cub";
+    rev = "1.10.0";
+    sha256 = "sha256-JyyNaTrtoSGiMP7tVUu9lFL07lyfJzRTVtx8yGy6/BI=";
+  };
+  spdlog = fetchFromGitHub {
+    owner = "gabime";
+    repo = "spdlog";
+    rev = "v1.9.2";
+    hash = "sha256-GSUdHtvV/97RyDKy8i+ticnSlQCubGGWHg4Oo+YAr8Y=";
   };
 
   cmakeFlags = [
-    "-DAF_BUILD_OPENCL=OFF"
+    "-DBUILD_TESTING=ON"
     "-DAF_BUILD_EXAMPLES=OFF"
-    "-DBUILD_TESTING=OFF"
-  ] ++ lib.optional cudaSupport "-DCMAKE_LIBRARY_PATH=${cudatoolkit}/lib/stubs";
-
-  patches = [ ./no-download.patch ];
+    "-DAF_COMPUTE_LIBRARY='FFTW/LAPACK/BLAS'"
+    "-DAF_TEST_WITH_MTX_FILES=OFF"
+    "-DAF_WITH_SPDLOG_HEADER_ONLY=ON"
+    "-DAF_BUILD_FORGE=OFF"
+    (if withCPU then "-DAF_BUILD_CPU=ON" else "-DAF_BUILD_CPU=OFF")
+    (if withOpenCL then "-DAF_BUILD_OPENCL=ON" else "-DAF_BUILD_OPENCL=OFF")
+    (if withCuda then "-DAF_BUILD_CUDA=ON" else "-DAF_BUILD_CUDA=OFF")
+  ] ++ lib.optionals withCuda [
+    "-DCUDA_LIBRARIES_PATH=${cudaPackages.cudatoolkit}/lib"
+  ];
 
   postPatch = ''
-    mkdir -p ./build/third_party/clFFT/src
-    cp -R --no-preserve=mode,ownership ${clfft.src}/ ./build/third_party/clFFT/src/clFFT-ext/
-    mkdir -p ./build/third_party/clBLAS/src
-    cp -R --no-preserve=mode,ownership ${clblas.src}/ ./build/third_party/clBLAS/src/clBLAS-ext/
-    mkdir -p ./build/include/CL
-    cp -R --no-preserve=mode,ownership ${opencl-clhpp}/include/CL/cl2.hpp ./build/include/CL/cl2.hpp
-  '';
+    mkdir -p ./extern/af_glad-src
+    mkdir -p ./extern/af_threads-src
+    mkdir -p ./extern/af_assets-src
+    mkdir -p ./extern/af_test_data-src
+    mkdir -p ./extern/ocl_clfft-src
+    mkdir -p ./extern/ocl_clblast-src
+    mkdir -p ./extern/nv_cub-src
+    mkdir -p ./extern/spdlog-src
+    cp -R --no-preserve=mode,ownership ${glad}/* ./extern/af_glad-src/
+    cp -R --no-preserve=mode,ownership ${threads}/* ./extern/af_threads-src/
+    cp -R --no-preserve=mode,ownership ${assets}/* ./extern/af_assets-src/
+    cp -R --no-preserve=mode,ownership ${test-data}/* ./extern/af_test_data-src/
+    cp -R --no-preserve=mode,ownership ${clfft}/* ./extern/ocl_clfft-src/
+    cp -R --no-preserve=mode,ownership ${clblast}/* ./extern/ocl_clblast-src/
+    cp -R --no-preserve=mode,ownership ${cub}/* ./extern/nv_cub-src/
+    cp -R --no-preserve=mode,ownership ${spdlog}/* ./extern/spdlog-src/
 
-  preBuild = lib.optionalString cudaSupport ''
-    export CUDA_PATH="${cudatoolkit}"
+    substituteInPlace src/api/unified/symbol_manager.cpp \
+      --replace '"/opt/arrayfire-3/lib/",' \
+                "\"$out/lib/\", \"/opt/arrayfire-3/lib/\","
   '';
 
-  nativeBuildInputs = [
-    cmake
-    pkg-config
-    python3
-  ];
-
-  strictDeps = true;
+  inherit doCheck;
+  checkPhase =
+    let
+      LD_LIBRARY_PATH = builtins.concatStringsSep ":" (
+        [ "${forge}/lib" "${freeimage}/lib" ]
+        ++ lib.optional withCuda "${cudaPackages.cudatoolkit}/lib64"
+        ++ lib.optional (nvidiaComputeDrivers != null) "${nvidiaComputeDrivers}/lib"
+      );
+      ctestFlags = builtins.concatStringsSep " " (
+        [ "--output-on-errors" "-j1" ]
+        # See https://github.com/arrayfire/arrayfire/issues/3484
+        ++ lib.optional withOpenCL "-E '(inverse_dense|cholesky_dense)'"
+      );
+    in
+    ''
+      export LD_LIBRARY_PATH=${LD_LIBRARY_PATH}
+    '' + lib.optionalString (withOpenCL && nvidiaComputeDrivers != null) ''
+      export OCL_ICD_VENDORS=${nvidiaComputeDrivers}/etc/OpenCL/vendors
+    '' + ''
+      AF_TRACE=all AF_PRINT_ERRORS=1 ctest ${ctestFlags}
+    '';
 
   buildInputs = [
-    opencl-clhpp
+    blas
+    boost.dev
+    boost.out
     fftw
     fftwFloat
-    blas
+    fmt_9
+    forge
+    freeimage
+    gtest
     lapack
-    libGLU
     libGL
-    mesa
-    freeimage
-    boost.out
-    boost.dev
-  ] ++ lib.optionals stdenv.isLinux [
     ocl-icd
-  ] ++ lib.optionals cudaSupport [
-    cudatoolkit
-  ] ++ lib.optionals buildDocs [
-    doxygen
-  ] ++ lib.optionals stdenv.isDarwin [
-    darwin.apple_sdk_11_0.frameworks.Accelerate
+    opencl-clhpp
+    span-lite
+  ]
+  ++ lib.optionals withCuda [
+    cudaPackages.cudatoolkit
+    cudaPackages.cudnn
+  ]
+  ++ lib.optionals withOpenCL [
+    mesa
+  ];
+
+  nativeBuildInputs = [
+    cmake
+    git
+    pkg-config
+    python3
   ];
 
   meta = with lib; {
@@ -96,6 +188,6 @@ stdenv.mkDerivation rec {
     license = licenses.bsd3;
     homepage = "https://arrayfire.com/";
     platforms = platforms.linux ++ platforms.darwin;
-    maintainers = with maintainers; [ chessai ];
+    maintainers = with maintainers; [ chessai twesterhout ];
   };
 }
diff --git a/pkgs/development/libraries/arrayfire/no-download.patch b/pkgs/development/libraries/arrayfire/no-download.patch
deleted file mode 100644
index 2b3ac492a54d..000000000000
--- a/pkgs/development/libraries/arrayfire/no-download.patch
+++ /dev/null
@@ -1,28 +0,0 @@
-diff --git a/CMakeModules/build_clBLAS.cmake b/CMakeModules/build_clBLAS.cmake
-index 8de529e8..6361b613 100644
---- a/CMakeModules/build_clBLAS.cmake
-+++ b/CMakeModules/build_clBLAS.cmake
-@@ -14,8 +14,7 @@ find_package(OpenCL)
- 
- ExternalProject_Add(
-     clBLAS-ext
--    GIT_REPOSITORY https://github.com/arrayfire/clBLAS.git
--    GIT_TAG arrayfire-release
-+    DOWNLOAD_COMMAND true
-     BUILD_BYPRODUCTS ${clBLAS_location}
-     PREFIX "${prefix}"
-     INSTALL_DIR "${prefix}"
-diff --git a/CMakeModules/build_clFFT.cmake b/CMakeModules/build_clFFT.cmake
-index 28be38a3..85e3915e 100644
---- a/CMakeModules/build_clFFT.cmake
-+++ b/CMakeModules/build_clFFT.cmake
-@@ -20,8 +20,7 @@ ENDIF()
- 
- ExternalProject_Add(
-     clFFT-ext
--    GIT_REPOSITORY https://github.com/arrayfire/clFFT.git
--    GIT_TAG arrayfire-release
-+    DOWNLOAD_COMMAND true
-     PREFIX "${prefix}"
-     INSTALL_DIR "${prefix}"
-     UPDATE_COMMAND ""
diff --git a/pkgs/development/libraries/forge/default.nix b/pkgs/development/libraries/forge/default.nix
index ba85cea73d69..5497aec839ed 100644
--- a/pkgs/development/libraries/forge/default.nix
+++ b/pkgs/development/libraries/forge/default.nix
@@ -1,38 +1,72 @@
-{ lib, stdenv, fetchFromGitHub, cmake, pkg-config
-, arrayfire, expat, fontconfig, freeimage, freetype, boost
-, mesa, libGLU, libGL, glfw3, SDL2, cudatoolkit
+{ boost
+, cmake
+, expat
+, fetchFromGitHub
+, fontconfig
+, freeimage
+, freetype
+, git
+, glfw3
+, glm
+, lib
+, libGLU
+, libGL
+, mesa
+, opencl-clhpp
+, pkg-config
+, stdenv
+, SDL2
 }:
 
 stdenv.mkDerivation rec {
   pname = "forge";
-  version = "1.0.4";
+  version = "1.0.8";
 
   src = fetchFromGitHub {
     owner = "arrayfire";
-    repo = "forge";
-    rev = "v${version}";
-    sha256 = "00pmky6kccd7pwi8sma79qpmzr2f9pbn6gym3gyqm64yckw6m484";
-    fetchSubmodules = true;
+    repo = pname;
+    rev = "v1.0.8";
+    sha256 = "sha256-lSZAwcqAHiuZkpYcVfwvZCfNmEF3xGN9S/HuZQrGeKU=";
   };
+  glad = fetchFromGitHub {
+    owner = "arrayfire";
+    repo = "glad";
+    rev = "b94680aee5b8ce01ae1644c5f2661769366c765a";
+    sha256 = "sha256-CrZy76gOGMpy9f1NuMK4tokZ57U//zYeNH5ZYY0SC2U=";
+  };
+
+  postPatch = ''
+    mkdir -p ./extern/fg_glad-src
+    cp -R --no-preserve=mode,ownership ${glad}/* ./extern/fg_glad-src/
+    ln -s ${opencl-clhpp} ./extern/cl2hpp
+
+    substituteInPlace CMakeModules/ForgeConfigureDepsVars.cmake \
+      --replace 'set(BUILD_OFFLINE OFF)' 'set(BUILD_OFFLINE ON)'
+  '';
+
+  cmakeFlags = [ "-DFETCHCONTENT_FULLY_DISCONNECTED=ON" ];
 
   nativeBuildInputs = [
     cmake
+    git
     pkg-config
   ];
 
   buildInputs = [
-    expat
-    fontconfig
-    freetype
     boost.out
     boost.dev
+    expat
+    fontconfig
     freeimage
-    mesa
-    libGLU libGL
+    freetype
     glfw3
+    glm
+    libGL
+    libGLU
+    opencl-clhpp
     SDL2
-    cudatoolkit
-    arrayfire
+  ] ++ lib.optionals (!stdenv.isDarwin) [
+    mesa
   ];
 
   meta = with lib; {
@@ -44,8 +78,7 @@ stdenv.mkDerivation rec {
     '';
     license = licenses.bsd3;
     homepage = "https://arrayfire.com/";
-    platforms = platforms.linux;
-    maintainers = with maintainers; [ chessai ];
+    platforms = platforms.linux ++ platforms.darwin;
+    maintainers = with maintainers; [ chessai twesterhout ];
   };
-
 }
diff --git a/pkgs/top-level/all-packages.nix b/pkgs/top-level/all-packages.nix
index 44e8876d1977..4df715fd3029 100644
--- a/pkgs/top-level/all-packages.nix
+++ b/pkgs/top-level/all-packages.nix
@@ -20548,7 +20548,9 @@ with pkgs;
 
   armadillo = callPackage ../development/libraries/armadillo { };
 
-  arrayfire = darwin.apple_sdk_11_0.callPackage ../development/libraries/arrayfire { };
+  arrayfire = darwin.apple_sdk_11_0.callPackage ../development/libraries/arrayfire {
+    cudaPackages = cudaPackages_12;
+  };
 
   arrow-cpp = callPackage ../development/libraries/arrow-cpp { };
 
@@ -21190,9 +21192,7 @@ with pkgs;
 
   fflas-ffpack = callPackage ../development/libraries/fflas-ffpack { };
 
-  forge = callPackage ../development/libraries/forge {
-    cudatoolkit = buildPackages.cudatoolkit_11;
-  };
+  forge = callPackage ../development/libraries/forge { };
 
   linbox = callPackage ../development/libraries/linbox { };