From: Tobias Grosser Date: Thu, 4 Aug 2016 12:39:03 +0000 (+0000) Subject: GPGPU: Add private memory support X-Git-Tag: llvmorg-4.0.0-rc1~13334 X-Git-Url: http://review.tizen.org/git/?a=commitdiff_plain;h=130ca30f92d818f234416191c6d6be55ab8b6f5b;p=platform%2Fupstream%2Fllvm.git GPGPU: Add private memory support llvm-svn: 277722 --- diff --git a/polly/lib/CodeGen/PPCGCodeGeneration.cpp b/polly/lib/CodeGen/PPCGCodeGeneration.cpp index 751925a..5bb75a6 100644 --- a/polly/lib/CodeGen/PPCGCodeGeneration.cpp +++ b/polly/lib/CodeGen/PPCGCodeGeneration.cpp @@ -80,6 +80,10 @@ static cl::opt SharedMemory("polly-acc-use-shared", cl::desc("Use shared memory"), cl::Hidden, cl::init(false), cl::ZeroOrMore, cl::cat(PollyCategory)); +static cl::opt PrivateMemory("polly-acc-use-private", + cl::desc("Use private memory"), cl::Hidden, + cl::init(false), cl::ZeroOrMore, + cl::cat(PollyCategory)); static cl::opt CudaVersion("polly-acc-cuda-version", @@ -1263,20 +1267,27 @@ void GPUNodeBuilder::createKernelVariables(ppcg_kernel *Kernel, Function *FN) { ArrayTy = ArrayType::get(ArrayTy, Bound); } - assert(Var.type == ppcg_access_shared && "Only shared memory supported"); - - GlobalVariable *SharedVar = new GlobalVariable( - *M, ArrayTy, false, GlobalValue::InternalLinkage, 0, Var.name, nullptr, - GlobalValue::ThreadLocalMode::NotThreadLocal, 3); - SharedVar->setAlignment(EleTy->getPrimitiveSizeInBits() / 8); - ConstantAggregateZero *Zero = ConstantAggregateZero::get(ArrayTy); - SharedVar->setInitializer(Zero); - + const ScopArrayInfo *SAI; + Value *Allocation; + if (Var.type == ppcg_access_shared) { + auto GlobalVar = new GlobalVariable( + *M, ArrayTy, false, GlobalValue::InternalLinkage, 0, Var.name, + nullptr, GlobalValue::ThreadLocalMode::NotThreadLocal, 3); + GlobalVar->setAlignment(EleTy->getPrimitiveSizeInBits() / 8); + ConstantAggregateZero *Zero = ConstantAggregateZero::get(ArrayTy); + GlobalVar->setInitializer(Zero); + Allocation = GlobalVar; + } else if (Var.type == ppcg_access_private) { + Allocation = Builder.CreateAlloca(ArrayTy, 0, "private_array"); + } else { + llvm_unreachable("unknown variable type"); + } + Builder.GetInsertBlock()->dump(); + SAI = S.getOrCreateScopArrayInfo(Allocation, EleTy, Sizes, + ScopArrayInfo::MK_Array); Id = isl_id_alloc(S.getIslCtx(), Var.name, nullptr); - IDToValue[Id] = SharedVar; - const ScopArrayInfo *SAI = S.getOrCreateScopArrayInfo( - SharedVar, EleTy, Sizes, ScopArrayInfo::MK_Array); - LocalArrays.push_back(SharedVar); + IDToValue[Id] = Allocation; + LocalArrays.push_back(Allocation); KernelIds.push_back(Id); IDToSAI[Id] = SAI; } @@ -1411,7 +1422,7 @@ public: Options->tile_size = 32; - Options->use_private_memory = false; + Options->use_private_memory = PrivateMemory; Options->use_shared_memory = SharedMemory; Options->max_shared_memory = 48 * 1024; diff --git a/polly/test/GPGPU/private-memory.ll b/polly/test/GPGPU/private-memory.ll new file mode 100644 index 0000000..882f420 --- /dev/null +++ b/polly/test/GPGPU/private-memory.ll @@ -0,0 +1,82 @@ +; RUN: opt %loadPolly -polly-codegen-ppcg -polly-acc-dump-code \ +; RUN: -polly-acc-use-private \ +; RUN: -disable-output < %s | \ +; RUN: FileCheck -check-prefix=CODE %s + +; RUN: opt %loadPolly -polly-codegen-ppcg \ +; RUN: -polly-acc-use-private \ +; RUN: -disable-output -polly-acc-dump-kernel-ir < %s | \ +; RUN: FileCheck -check-prefix=KERNEL %s + +; REQUIRES: pollyacc + +; void add(float *A) { +; for (long i = 0; i < 32; i++) +; for (long j = 0; j < 10; j++) +; A[i] += 1; +; } + +; CODE: # kernel0 +; CODE: { +; CODE: read(t0); +; CODE: for (int c3 = 0; c3 <= 9; c3 += 1) +; CODE: Stmt_bb5(t0, c3); +; CODE: write(t0); +; CODE: } + +; KERNEL: %private_array = alloca [1 x float] + +; KERNEL: %polly.access.cast.private_array = bitcast [1 x float]* %private_array to float* +; KERNEL-NEXT: %polly.access.private_array = getelementptr float, float* %polly.access.cast.private_array, i64 0 +; KERNEL-NEXT: %polly.access.cast.MemRef_A = bitcast i8* %MemRef_A to float* +; KERNEL-NEXT: %polly.access.MemRef_A = getelementptr float, float* %polly.access.cast.MemRef_A, i64 %t0 +; KERNEL-NEXT: %shared.read = load float, float* %polly.access.MemRef_A +; KERNEL-NEXT: store float %shared.read, float* %polly.access.private_array + +; KERNEL: %polly.access.cast.private_array5 = bitcast [1 x float]* %private_array to float* +; KERNEL-NEXT: %polly.access.private_array6 = getelementptr float, float* %polly.access.cast.private_array5, i64 0 +; KERNEL-NEXT: %polly.access.cast.MemRef_A7 = bitcast i8* %MemRef_A to float* +; KERNEL-NEXT: %polly.access.MemRef_A8 = getelementptr float, float* %polly.access.cast.MemRef_A7, i64 %t0 +; KERNEL-NEXT: %shared.write = load float, float* %polly.access.private_array6 +; KERNEL-NEXT: store float %shared.write, float* %polly.access.MemRef_A8 + +target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128" + +define void @add(float* %A) { +bb: + br label %bb2 + +bb2: ; preds = %bb11, %bb + %i.0 = phi i64 [ 0, %bb ], [ %tmp12, %bb11 ] + %exitcond1 = icmp ne i64 %i.0, 32 + br i1 %exitcond1, label %bb3, label %bb13 + +bb3: ; preds = %bb2 + br label %bb4 + +bb4: ; preds = %bb8, %bb3 + %j.0 = phi i64 [ 0, %bb3 ], [ %tmp9, %bb8 ] + %exitcond = icmp ne i64 %j.0, 10 + br i1 %exitcond, label %bb5, label %bb10 + +bb5: ; preds = %bb4 + %tmp = getelementptr inbounds float, float* %A, i64 %i.0 + %tmp6 = load float, float* %tmp, align 4 + %tmp7 = fadd float %tmp6, 1.000000e+00 + store float %tmp7, float* %tmp, align 4 + br label %bb8 + +bb8: ; preds = %bb5 + %tmp9 = add nuw nsw i64 %j.0, 1 + br label %bb4 + +bb10: ; preds = %bb4 + br label %bb11 + +bb11: ; preds = %bb10 + %tmp12 = add nuw nsw i64 %i.0, 1 + br label %bb2 + +bb13: ; preds = %bb2 + ret void +}