[flang-commits] [flang] [flang][cuda] Lower launch_bounds values (PR #81537)
Valentin Clement バレンタイン クレメン via flang-commits
flang-commits at lists.llvm.org
Mon Feb 12 20:58:03 PST 2024
https://github.com/clementval updated https://github.com/llvm/llvm-project/pull/81537
>From 7f7050ec86acd331c796fb57f298682885bd5196 Mon Sep 17 00:00:00 2001
From: Valentin Clement <clementval at gmail.com>
Date: Mon, 12 Feb 2024 13:10:19 -0800
Subject: [PATCH 1/2] [flang][cuda] Lower launch_bounds values
---
.../flang/Optimizer/Dialect/FIRAttr.td | 12 +++++
.../flang/Optimizer/Dialect/FIROpsSupport.h | 5 +++
flang/lib/Lower/CallInterface.cpp | 45 ++++++++++++++++---
flang/lib/Optimizer/Dialect/FIRAttr.cpp | 3 +-
flang/test/Lower/CUDA/cuda-proc-attribute.cuf | 6 +++
5 files changed, 64 insertions(+), 7 deletions(-)
diff --git a/flang/include/flang/Optimizer/Dialect/FIRAttr.td b/flang/include/flang/Optimizer/Dialect/FIRAttr.td
index 00e293e2f04278..d55e93ba666e71 100644
--- a/flang/include/flang/Optimizer/Dialect/FIRAttr.td
+++ b/flang/include/flang/Optimizer/Dialect/FIRAttr.td
@@ -113,4 +113,16 @@ def fir_CUDAProcAttributeAttr :
let assemblyFormat = [{ ```<` $value `>` }];
}
+def fir_CUDALaunchBoundsAttr : fir_Attr<"CUDALaunchBounds"> {
+ let mnemonic = "launch_bounds";
+
+ let parameters = (ins
+ "mlir::IntegerAttr":$maxTBP,
+ "mlir::IntegerAttr":$minBPM,
+ OptionalParameter<"mlir::IntegerAttr">:$upperBoundClusterSize
+ );
+
+ let assemblyFormat = "`<` struct(params) `>`";
+}
+
#endif // FIR_DIALECT_FIR_ATTRS
diff --git a/flang/include/flang/Optimizer/Dialect/FIROpsSupport.h b/flang/include/flang/Optimizer/Dialect/FIROpsSupport.h
index 6ac6a3116d40b0..29fa57cd7a0d8a 100644
--- a/flang/include/flang/Optimizer/Dialect/FIROpsSupport.h
+++ b/flang/include/flang/Optimizer/Dialect/FIROpsSupport.h
@@ -75,6 +75,11 @@ static constexpr llvm::StringRef getTargetAttrName() { return "fir.target"; }
/// Attribute to mark Fortran entities with the CUDA attribute.
static constexpr llvm::StringRef getCUDAAttrName() { return "fir.cuda_attr"; }
+/// Attribute to carry CUDA launch_bounds values.
+static constexpr llvm::StringRef getCUDALaunchBoundsAttrName() {
+ return "fir.cuda_launch_bounds";
+}
+
/// Attribute to mark that a function argument is a character dummy procedure.
/// Character dummy procedure have special ABI constraints.
static constexpr llvm::StringRef getCharacterProcedureDummyAttrName() {
diff --git a/flang/lib/Lower/CallInterface.cpp b/flang/lib/Lower/CallInterface.cpp
index 41597c1b15386e..968fba7a66044a 100644
--- a/flang/lib/Lower/CallInterface.cpp
+++ b/flang/lib/Lower/CallInterface.cpp
@@ -524,6 +524,43 @@ static void addSymbolAttribute(mlir::func::FuncOp func,
mlir::StringAttr::get(&mlirContext, name));
}
+static void
+setCUDAAttributes(mlir::func::FuncOp func,
+ const Fortran::semantics::Symbol *sym,
+ std::optional<Fortran::evaluate::characteristics::Procedure>
+ characteristic) {
+ if (characteristic && characteristic->cudaSubprogramAttrs) {
+ func.getOperation()->setAttr(
+ fir::getCUDAAttrName(),
+ fir::getCUDAProcAttribute(func.getContext(),
+ *characteristic->cudaSubprogramAttrs));
+ }
+
+ if (sym) {
+ if (auto details =
+ sym->GetUltimate()
+ .detailsIf<Fortran::semantics::SubprogramDetails>()) {
+ if (!details->cudaLaunchBounds().empty()) {
+ assert(details->cudaLaunchBounds().size() >= 2 &&
+ "expect at least 2 values");
+ mlir::Type i64Ty = mlir::IntegerType::get(func.getContext(), 64);
+ auto maxTBPAttr =
+ mlir::IntegerAttr::get(i64Ty, details->cudaLaunchBounds()[0]);
+ auto minBPMAttr =
+ mlir::IntegerAttr::get(i64Ty, details->cudaLaunchBounds()[1]);
+ mlir::IntegerAttr ubAttr;
+ if (details->cudaLaunchBounds().size() > 2)
+ ubAttr =
+ mlir::IntegerAttr::get(i64Ty, details->cudaLaunchBounds()[2]);
+ func.getOperation()->setAttr(
+ fir::getCUDALaunchBoundsAttrName(),
+ fir::CUDALaunchBoundsAttr::get(func.getContext(), maxTBPAttr,
+ minBPMAttr, ubAttr));
+ }
+ }
+ }
+}
+
/// Declare drives the different actions to be performed while analyzing the
/// signature and building/finding the mlir::func::FuncOp.
template <typename T>
@@ -559,12 +596,8 @@ void Fortran::lower::CallInterface<T>::declare() {
if (!placeHolder.value().attributes.empty())
func.setArgAttrs(placeHolder.index(), placeHolder.value().attributes);
side().setFuncAttrs(func);
- }
- if (characteristic && characteristic->cudaSubprogramAttrs) {
- func.getOperation()->setAttr(
- fir::getCUDAAttrName(),
- fir::getCUDAProcAttribute(func.getContext(),
- *characteristic->cudaSubprogramAttrs));
+
+ setCUDAAttributes(func, side().getProcedureSymbol(), characteristic);
}
}
}
diff --git a/flang/lib/Optimizer/Dialect/FIRAttr.cpp b/flang/lib/Optimizer/Dialect/FIRAttr.cpp
index 8df7a6c5cfc5d5..8d780e03dcbe73 100644
--- a/flang/lib/Optimizer/Dialect/FIRAttr.cpp
+++ b/flang/lib/Optimizer/Dialect/FIRAttr.cpp
@@ -298,5 +298,6 @@ void fir::printFirAttribute(FIROpsDialect *dialect, mlir::Attribute attr,
void FIROpsDialect::registerAttributes() {
addAttributes<ClosedIntervalAttr, ExactTypeAttr, FortranVariableFlagsAttr,
LowerBoundAttr, PointIntervalAttr, RealAttr, SubclassAttr,
- UpperBoundAttr, CUDADataAttributeAttr, CUDAProcAttributeAttr>();
+ UpperBoundAttr, CUDADataAttributeAttr, CUDAProcAttributeAttr,
+ CUDALaunchBoundsAttr>();
}
diff --git a/flang/test/Lower/CUDA/cuda-proc-attribute.cuf b/flang/test/Lower/CUDA/cuda-proc-attribute.cuf
index 050731086d8525..ade38d4b7f0917 100644
--- a/flang/test/Lower/CUDA/cuda-proc-attribute.cuf
+++ b/flang/test/Lower/CUDA/cuda-proc-attribute.cuf
@@ -32,3 +32,9 @@ attributes(host) attributes(device) integer function fct_host_device; end
attributes(device) attributes(host) integer function fct_device_host; end
! CHECK: func.func @_QPfct_device_host() -> i32 attributes {fir.cuda_attr = #fir.cuda_proc<host_device>}
+
+attributes(global) launch_bounds(1, 2) subroutine sub_lbounds1(); end
+! CHECK: func.func @_QPsub_lbounds1() attributes {fir.cuda_attr = #fir.cuda_proc<global>, fir.cuda_launch_bounds = #fir.launch_bounds<maxTBP = 1 : i64, minBPM = 2 : i64>}
+
+attributes(global) launch_bounds(1, 2, 3) subroutine sub_lbounds2(); end
+! CHECK: func.func @_QPsub_lbounds2() attributes {fir.cuda_attr = #fir.cuda_proc<global>, fir.cuda_launch_bounds = #fir.launch_bounds<maxTBP = 1 : i64, minBPM = 2 : i64, upperBoundClusterSize = 3 : i64>}
>From dd9a29a870f50b38d1261c520a2376b748cd13f2 Mon Sep 17 00:00:00 2001
From: Valentin Clement <clementval at gmail.com>
Date: Mon, 12 Feb 2024 20:57:51 -0800
Subject: [PATCH 2/2] Fix type with maxTPB
---
flang/include/flang/Optimizer/Dialect/FIRAttr.td | 2 +-
flang/lib/Lower/CallInterface.cpp | 4 ++--
flang/test/Lower/CUDA/cuda-proc-attribute.cuf | 4 ++--
3 files changed, 5 insertions(+), 5 deletions(-)
diff --git a/flang/include/flang/Optimizer/Dialect/FIRAttr.td b/flang/include/flang/Optimizer/Dialect/FIRAttr.td
index d55e93ba666e71..3602c67de1412a 100644
--- a/flang/include/flang/Optimizer/Dialect/FIRAttr.td
+++ b/flang/include/flang/Optimizer/Dialect/FIRAttr.td
@@ -117,7 +117,7 @@ def fir_CUDALaunchBoundsAttr : fir_Attr<"CUDALaunchBounds"> {
let mnemonic = "launch_bounds";
let parameters = (ins
- "mlir::IntegerAttr":$maxTBP,
+ "mlir::IntegerAttr":$maxTPB,
"mlir::IntegerAttr":$minBPM,
OptionalParameter<"mlir::IntegerAttr">:$upperBoundClusterSize
);
diff --git a/flang/lib/Lower/CallInterface.cpp b/flang/lib/Lower/CallInterface.cpp
index 968fba7a66044a..f990e0b7ce4dcf 100644
--- a/flang/lib/Lower/CallInterface.cpp
+++ b/flang/lib/Lower/CallInterface.cpp
@@ -544,7 +544,7 @@ setCUDAAttributes(mlir::func::FuncOp func,
assert(details->cudaLaunchBounds().size() >= 2 &&
"expect at least 2 values");
mlir::Type i64Ty = mlir::IntegerType::get(func.getContext(), 64);
- auto maxTBPAttr =
+ auto maxTPBAttr =
mlir::IntegerAttr::get(i64Ty, details->cudaLaunchBounds()[0]);
auto minBPMAttr =
mlir::IntegerAttr::get(i64Ty, details->cudaLaunchBounds()[1]);
@@ -554,7 +554,7 @@ setCUDAAttributes(mlir::func::FuncOp func,
mlir::IntegerAttr::get(i64Ty, details->cudaLaunchBounds()[2]);
func.getOperation()->setAttr(
fir::getCUDALaunchBoundsAttrName(),
- fir::CUDALaunchBoundsAttr::get(func.getContext(), maxTBPAttr,
+ fir::CUDALaunchBoundsAttr::get(func.getContext(), maxTPBAttr,
minBPMAttr, ubAttr));
}
}
diff --git a/flang/test/Lower/CUDA/cuda-proc-attribute.cuf b/flang/test/Lower/CUDA/cuda-proc-attribute.cuf
index ade38d4b7f0917..9eb2b85aaf0b83 100644
--- a/flang/test/Lower/CUDA/cuda-proc-attribute.cuf
+++ b/flang/test/Lower/CUDA/cuda-proc-attribute.cuf
@@ -34,7 +34,7 @@ attributes(device) attributes(host) integer function fct_device_host; end
! CHECK: func.func @_QPfct_device_host() -> i32 attributes {fir.cuda_attr = #fir.cuda_proc<host_device>}
attributes(global) launch_bounds(1, 2) subroutine sub_lbounds1(); end
-! CHECK: func.func @_QPsub_lbounds1() attributes {fir.cuda_attr = #fir.cuda_proc<global>, fir.cuda_launch_bounds = #fir.launch_bounds<maxTBP = 1 : i64, minBPM = 2 : i64>}
+! CHECK: func.func @_QPsub_lbounds1() attributes {fir.cuda_attr = #fir.cuda_proc<global>, fir.cuda_launch_bounds = #fir.launch_bounds<maxTPB = 1 : i64, minBPM = 2 : i64>}
attributes(global) launch_bounds(1, 2, 3) subroutine sub_lbounds2(); end
-! CHECK: func.func @_QPsub_lbounds2() attributes {fir.cuda_attr = #fir.cuda_proc<global>, fir.cuda_launch_bounds = #fir.launch_bounds<maxTBP = 1 : i64, minBPM = 2 : i64, upperBoundClusterSize = 3 : i64>}
+! CHECK: func.func @_QPsub_lbounds2() attributes {fir.cuda_attr = #fir.cuda_proc<global>, fir.cuda_launch_bounds = #fir.launch_bounds<maxTPB = 1 : i64, minBPM = 2 : i64, upperBoundClusterSize = 3 : i64>}
More information about the flang-commits
mailing list