Basic 2D LDC using boltzgen for kernel generation

Using cell lists as parameters for multiple non-branching kernels seems
to reduce performance by ~50 MLUPS (for single precision D2Q9).
This might be alleviated by padding the cell lists to enable thread
layout control or by improved kernel dispatching.

On the upside this OpenCL program runs not only on GPUs but is also vectorized on Intel
CPUs yielding about 180 MLUPS (single precision) and - anticlimactically - 85 MLUPS for
double precision on a i7-4790K.
However both these values compare well to the performance of established CPU LBM codes.

1 files changed, 43 insertions, 0 deletions

diff --git a/shell.nix b/shell.nix
new file mode 100644
index 0000000..99d794b
--- /dev/null
+++ b/shell.nix
@@ -0,0 +1,43 @@
+{ pkgs ? import <nixpkgs> { }, ... }:
+
+pkgs.stdenvNoCC.mkDerivation rec {
+  name = "pycl-env";
+  env = pkgs.buildEnv { name = name; paths = buildInputs; };
+
+  buildInputs = let
+    boltzgen = pkgs.python3.pkgs.buildPythonPackage rec {
+      pname = "boltzgen";
+      version = "0.1";
+
+      src = pkgs.fetchFromGitHub {
+        owner  = "KnairdA";
+        repo   = "boltzgen";
+        rev    = "v0.1";
+        sha256 = "072kx4jrzd0g9rn63hjb0yic7qhbga47lp2vbz7rq3gvkqv1hz4d";
+      };
+
+      propagatedBuildInputs = with pkgs.python37Packages; [
+        sympy
+        numpy
+        Mako
+      ];
+    };
+
+    local-python = pkgs.python3.withPackages (python-packages: with python-packages; [
+      boltzgen
+      numpy
+      pyopencl setuptools
+      matplotlib
+    ]);
+
+  in [
+    local-python
+    pkgs.opencl-info
+  ];
+
+  shellHook = ''
+    export NIX_SHELL_NAME="${name}"
+    export PYOPENCL_COMPILER_OUTPUT=1
+    export PYTHONPATH="$PWD:$PYTHONPATH"
+  '';
+}