[flang-commits] [flang] [flang][cuda] Avoid to issue data transfer in device context (PR #90247)
Valentin Clement バレンタイン クレメン via flang-commits
flang-commits at lists.llvm.org
Fri Apr 26 13:02:14 PDT 2024
https://github.com/clementval updated https://github.com/llvm/llvm-project/pull/90247
>From c82e690702c04d337eee249ba9c088f71e0bcb03 Mon Sep 17 00:00:00 2001
From: Valentin Clement <clementval at gmail.com>
Date: Fri, 26 Apr 2024 11:29:52 -0700
Subject: [PATCH 1/2] [flang][cuda] Avoid to issue data transfer in device
context
Data transfer should not be issued in device function.
---
flang/lib/Lower/Bridge.cpp | 24 ++++++++++++++++----
flang/test/Lower/CUDA/cuda-data-transfer.cuf | 10 ++++++++
2 files changed, 29 insertions(+), 5 deletions(-)
diff --git a/flang/lib/Lower/Bridge.cpp b/flang/lib/Lower/Bridge.cpp
index 19c00884bd1b7e..e3679ef2afbbc4 100644
--- a/flang/lib/Lower/Bridge.cpp
+++ b/flang/lib/Lower/Bridge.cpp
@@ -3807,16 +3807,30 @@ class FirConverter : public Fortran::lower::AbstractConverter {
return temps;
}
+ static bool isDeviceContext(fir::FirOpBuilder &builder) {
+ if (builder.getRegion().getParentOfType<fir::CUDAKernelOp>())
+ return true;
+ if (auto funcOp =
+ builder.getRegion().getParentOfType<mlir::func::FuncOp>()) {
+ if (auto cudaProcAttr =
+ funcOp.getOperation()->getAttrOfType<fir::CUDAProcAttributeAttr>(
+ fir::getCUDAAttrName())) {
+ return cudaProcAttr.getValue() != fir::CUDAProcAttribute::Host;
+ }
+ }
+ return false;
+ }
+
void genDataAssignment(
const Fortran::evaluate::Assignment &assign,
const Fortran::evaluate::ProcedureRef *userDefinedAssignment) {
mlir::Location loc = getCurrentLocation();
fir::FirOpBuilder &builder = getFirOpBuilder();
- bool isInDeviceContext =
- builder.getRegion().getParentOfType<fir::CUDAKernelOp>();
- bool isCUDATransfer = Fortran::evaluate::HasCUDAAttrs(assign.lhs) ||
- Fortran::evaluate::HasCUDAAttrs(assign.rhs);
+ bool isInDeviceContext = isDeviceContext(builder);
+ bool isCUDATransfer = (Fortran::evaluate::HasCUDAAttrs(assign.lhs) ||
+ Fortran::evaluate::HasCUDAAttrs(assign.rhs)) &&
+ !isInDeviceContext;
bool hasCUDAImplicitTransfer =
Fortran::evaluate::HasCUDAImplicitTransfer(assign.rhs);
llvm::SmallVector<mlir::Value> implicitTemps;
@@ -3879,7 +3893,7 @@ class FirConverter : public Fortran::lower::AbstractConverter {
Fortran::lower::StatementContext localStmtCtx;
hlfir::Entity rhs = evaluateRhs(localStmtCtx);
hlfir::Entity lhs = evaluateLhs(localStmtCtx);
- if (isCUDATransfer && !hasCUDAImplicitTransfer && !isInDeviceContext)
+ if (isCUDATransfer && !hasCUDAImplicitTransfer)
genCUDADataTransfer(builder, loc, assign, lhs, rhs);
else
builder.create<hlfir::AssignOp>(loc, rhs, lhs,
diff --git a/flang/test/Lower/CUDA/cuda-data-transfer.cuf b/flang/test/Lower/CUDA/cuda-data-transfer.cuf
index 70483685d20019..add1052e576c15 100644
--- a/flang/test/Lower/CUDA/cuda-data-transfer.cuf
+++ b/flang/test/Lower/CUDA/cuda-data-transfer.cuf
@@ -141,3 +141,13 @@ end subroutine
! CHECK: fir.cuda_kernel<<<*, *>>>
! CHECK-NOT: fir.cuda_data_transfer
! CHECK: hlfir.assign
+
+attributes(global) subroutine sub5(a)
+ integer, device :: a
+ integer :: i
+ i = threadIdx%x + (blockIdx%x - 1) * blockDim%x
+ a = i
+end subroutine
+
+! CHECK-LABEL: func.func @_QPsub5
+! CHECK-NOT: fir.cuda_data_transfer
>From 788b9f9d744dc20b750afc13826303e5ebbbeb1f Mon Sep 17 00:00:00 2001
From: Valentin Clement <clementval at gmail.com>
Date: Fri, 26 Apr 2024 13:02:02 -0700
Subject: [PATCH 2/2] Treat host,device as host context
---
flang/lib/Lower/Bridge.cpp | 6 +++++-
flang/test/Lower/CUDA/cuda-data-transfer.cuf | 10 +++++++++-
2 files changed, 14 insertions(+), 2 deletions(-)
diff --git a/flang/lib/Lower/Bridge.cpp b/flang/lib/Lower/Bridge.cpp
index e3679ef2afbbc4..eb1660a706a88d 100644
--- a/flang/lib/Lower/Bridge.cpp
+++ b/flang/lib/Lower/Bridge.cpp
@@ -3807,6 +3807,9 @@ class FirConverter : public Fortran::lower::AbstractConverter {
return temps;
}
+ // Check if the insertion point is currently in a device context. HostDevice
+ // subprogram are not considered fully device context so it will return false
+ // for it.
static bool isDeviceContext(fir::FirOpBuilder &builder) {
if (builder.getRegion().getParentOfType<fir::CUDAKernelOp>())
return true;
@@ -3815,7 +3818,8 @@ class FirConverter : public Fortran::lower::AbstractConverter {
if (auto cudaProcAttr =
funcOp.getOperation()->getAttrOfType<fir::CUDAProcAttributeAttr>(
fir::getCUDAAttrName())) {
- return cudaProcAttr.getValue() != fir::CUDAProcAttribute::Host;
+ return cudaProcAttr.getValue() != fir::CUDAProcAttribute::Host
+ && cudaProcAttr.getValue() != fir::CUDAProcAttribute::HostDevice;
}
}
return false;
diff --git a/flang/test/Lower/CUDA/cuda-data-transfer.cuf b/flang/test/Lower/CUDA/cuda-data-transfer.cuf
index add1052e576c15..0a2608639bce7c 100644
--- a/flang/test/Lower/CUDA/cuda-data-transfer.cuf
+++ b/flang/test/Lower/CUDA/cuda-data-transfer.cuf
@@ -145,9 +145,17 @@ end subroutine
attributes(global) subroutine sub5(a)
integer, device :: a
integer :: i
- i = threadIdx%x + (blockIdx%x - 1) * blockDim%x
a = i
end subroutine
! CHECK-LABEL: func.func @_QPsub5
! CHECK-NOT: fir.cuda_data_transfer
+
+attributes(host,device) subroutine sub6(a)
+ integer, device :: a
+ integer :: i
+ a = i
+end subroutine
+
+! CHECK-LABEL: func.func @_QPsub6
+! CHECK: fir.cuda_data_transfer
More information about the flang-commits
mailing list