コード例 #1
0
ファイル: JitCompiler.cs プロジェクト: xeno-by/conflux
        // todo. cache jitted kernels
        public static JittedKernel DoCompile(CudaConfig cfg, Type t_kernel)
        {
            t_kernel = t_kernel.Hierarchy().AssertFirst(t => !t.Assembly.HasAttr<CompilerGeneratedAttribute>());
            using (var ctx = new JitContext(cfg, t_kernel))
            {
                var inliner_ctx = new ExpansionContext(Kernel);
                Hir = Hir.Expand(inliner_ctx);
                Log.EnsureBlankLine();
                Log.WriteLine("After inlining:");
                Log.WriteLine(Hir.DumpAsText());

                MemoryAllocator.InferAllocationScheme();
                Log.EnsureBlankLine();
                Log.WriteLine("Non-standard allocations:");
                var nonstandard_allocs = 0;
                Allocs.Fields.Where(kvp => kvp.Value != MemoryTier.Global).ForEach(kvp => { Log.WriteLine(kvp.Key.GetCSharpRef(ToCSharpOptions.Informative)); nonstandard_allocs++; });
                Allocs.Symbols.Where(kvp => kvp.Value != MemoryTier.Private).ForEach(kvp => { Log.WriteLine(kvp.Key); nonstandard_allocs++; });
                Log.WriteLine((nonstandard_allocs == 0 ? "None" : "") + Environment.NewLine);

                // todo. also implement the following:
                // 1) downgrade to SSA
                // 2) perform SCC from "Constant Propagation with Conditional Branches"
                //    when performing SCC don't forget to funcletize stuff e.g. Impl::Cfg and Impl::Device
                // 3) eliminate dead code

                Generator.Traverse(Hir);
                Log.EnsureBlankLine();
                Log.WriteLine("Generated PTX:");
                Log.WriteLine();
                Log.WriteLine(Ptx);

                return new JittedKernel(ctx);
            }
        }
コード例 #2
0
ファイル: JittedKernel.cs プロジェクト: xeno-by/conflux
        public JittedKernel(JitContext ctx)
        {
            _cfg = ctx.Cfg.Clone();
            _ptx = ctx.Ptx;
            _hir = ctx.Hir.Fluent(blk => blk.FreezeForever());

            var flds = ctx.Allocator.Fields;
            _memcpyHostToDevice = kernel => 
            {
                var args = new KernelArguments();
                flds.Keys.ForEach(fld =>
                {
                    var value = fld.GetValue(kernel);
                    if (flds[fld] is SlotLayout)
                    {
                        args.Add(value.In());
                    }
                    else if (flds[fld] is ArrayLayout)
                    {
                        var arr = value.AssertCast<Array>();
                        args.Add(arr.InOut());

                        var rank = arr.GetType().GetArrayRank();
                        0.UpTo(rank - 1).ForEach(i => args.Add(arr.GetLength(i).In()));
                    }
                    else
                    {
                        throw AssertionHelper.Fail();
                    }
                });

                return args;
            };

            _memcpyDeviceToHost = (result, kernel) =>
            {
                var idx = 0;
                flds.Keys.ForEach(fld =>
                {
                    Object value;
                    if (flds[fld] is SlotLayout)
                    {
                        value = result[idx];
                        idx += 1;
                    }
                    else if (flds[fld] is ArrayLayout)
                    {
                        value = result[idx];
                        idx += 3;
                    }
                    else
                    {
                        throw AssertionHelper.Fail();
                    }

                    fld.SetValue(kernel, value);
                });
            };
        }
コード例 #3
0
ファイル: WaveKernelTests.cs プロジェクト: xeno-by/conflux
 public void WaveCuda()
 {
     var cfg = new CudaConfig();
     cfg.Codebase.OptIn(t => t.Assembly.GetName().Name == "Conflux.Playground");
     var kernel = cfg.Configure<WaveKernel>();
     var empty = new Matrix<Cell>(202, 202);
     var after2s = kernel.Execute(empty);
     // todo. validate the result
 }
コード例 #4
0
ファイル: JitContext.cs プロジェクト: xeno-by/conflux
        public JitContext(CudaConfig cfg, Type t_kernel)
        {
            Current.AssertNull();
            Current = this;

            Cfg = cfg;
            Kernel = t_kernel.GetMethod("RunKernel", BF.All);
            Kernel.AssertNotNull();
            Kernel.IsInstance().AssertTrue();
            Kernel.Params().AssertEmpty();
            (Kernel.Ret() == typeof(void)).AssertTrue();

            Hir = Kernel.Decompile().Body;
            AllocationScheme = null;
            Allocator = new Allocator();
            Generator = new Generator();
            Module = new PtxModule(Cfg.Target, Cfg.Version);
            Entry = Module.AddEntry(Kernel.Name);
            Params = Entry.Params;
        }
コード例 #5
0
ファイル: Tests.Boilerplate.cs プロジェクト: xeno-by/conflux
        protected void TestKernelCrosscompilation(Type t_kernel)
        {
            var cfg = new CudaConfig{BlockDim = new dim3(16, 16, 1)};
            cfg.Codebase.OptIn(t => t.Assembly.GetName().Name == "Conflux.Playground");

            using (Runtimes.Activate(new CudaRuntime(cfg, t_kernel)))
            {
                var result = JitCompiler.DoCompile(cfg, t_kernel);
                var s_ptx_actual = result.Ptx;
                var s_hir_actual = result.Hir.DumpAsText();

                var asm = MethodInfo.GetCurrentMethod().DeclaringType.Assembly;
                var @namespace = MethodInfo.GetCurrentMethod().DeclaringType.Namespace;
                @namespace += ".Reference.";
                var ptx_fileName = asm.GetManifestResourceNames().SingleOrDefault2(
                    n => String.Compare(n, @namespace + t_kernel.Name + ".ptx", true) == 0);
                var hir_fileName = asm.GetManifestResourceNames().SingleOrDefault2(
                    n => String.Compare(n, @namespace + t_kernel.Name + ".hir", true) == 0);

                Verify(s_ptx_actual, ptx_fileName, "crosscompiled PTX", t_kernel);
                Verify(s_hir_actual, hir_fileName, "crosscompiled HIR", t_kernel);
            }
        }
コード例 #6
0
 protected CudaConfig(CudaConfig proto) : base(proto)
 {
 }
コード例 #7
0
 protected CudaConfig(CudaConfig proto) : base(proto) {}
コード例 #8
0
 public void BigTest_Cuda()
 {
     var cfg = new CudaConfig();
     cfg.Codebase.OptIn(t => t.Assembly.GetName().Name == "Conflux.Playground");
     1.TimesDo(() => BigTest(cfg));
 }