Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

[SYCL][NVPTX] Emit reqd_work_group_size attributes as NVVM annotations #14502

Merged
merged 3 commits into from
Jul 16, 2024
Merged

[SYCL][NVPTX] Emit reqd_work_group_size attributes as NVVM annotations #14502

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
67f336b
[SYCL][NVPTX] Emit reqd_work_group_size attributes as NVVM annotations
frasercrmck Jul 9, 2024
22ff4c8
fix excessively large work-group sizes
frasercrmck Jul 10, 2024
ea4e4c7
Merge remote-tracking branch 'origin/sycl' into sycl-nvptx-reqd-wg-size
frasercrmck Jul 10, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
31 changes: 31 additions & 0 deletions clang/lib/CodeGen/Targets/NVPTX.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -303,6 +303,37 @@ void NVPTXTargetCodeGenInfo::setTargetAttributes(
addNVVMMetadata(F, "maxntidz", MWGS->getXDimVal());
}

if (const auto *RWGS = FD->getAttr<SYCLReqdWorkGroupSizeAttr>()) {
llvm::SmallVector<std::optional<int64_t>, 3> Ops;
// Index-flip and pad out any missing elements. Note the misleading
// nomenclature of the methods: getXDimVal doesn't return the X dimension;
// it returns the left-most dimension (dim0). This could correspond to
// CUDA's X, Y, or Z, depending on the number of operands provided.
if (auto Dim0 = RWGS->getXDimVal())
Ops.push_back(Dim0->getExtValue());
if (auto Dim1 = RWGS->getYDimVal())
Ops.push_back(Dim1->getExtValue());
if (auto Dim2 = RWGS->getZDimVal())
Ops.push_back(Dim2->getExtValue());
std::reverse(Ops.begin(), Ops.end());
Ops.append(3 - Ops.size(), std::nullopt);

// Work-group sizes (in NVVM annotations) must be positive and less than
// INT32_MAX, whereas SYCL can allow for larger work-group sizes (see
// -fno-sycl-id-queries-fit-in-int). If any dimension is too large for
// NVPTX, don't emit any annotation at all.
if (llvm::all_of(Ops, [](std::optional<int64_t> V) {
return !V || llvm::isUInt<31>(*V);
steffenlarsen marked this conversation as resolved.
Show resolved Hide resolved
})) {
if (auto X = Ops[0])
addNVVMMetadata(F, "reqntidx", *X);
if (auto Y = Ops[1])
addNVVMMetadata(F, "reqntidy", *Y);
if (auto Z = Ops[2])
addNVVMMetadata(F, "reqntidz", *Z);
}
}

auto attrValue = [&](Expr *E) {
const auto *CE = cast<ConstantExpr>(E);
std::optional<llvm::APInt> Val = CE->getResultAsAPSInt();
Expand Down
53 changes: 51 additions & 2 deletions clang/test/CodeGenSYCL/reqd-work-group-size.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,7 +1,7 @@
// RUN: %clang_cc1 -fsycl-is-device -internal-isystem %S/Inputs -triple spir64-unknown-unknown -disable-llvm-passes -emit-llvm -o - %s | FileCheck %s
// RUN: %clang_cc1 -fsycl-is-device -internal-isystem %S/Inputs -triple amdgcn-amd-amdhsa -disable-llvm-passes -emit-llvm -o - %s | FileCheck %s
// RUN: %clang_cc1 -fsycl-is-device -internal-isystem %S/Inputs -triple nvptx-nvidia-cuda -disable-llvm-passes -emit-llvm -o - %s | FileCheck %s
// RUN: %clang_cc1 -fsycl-is-device -internal-isystem %S/Inputs -triple nvptx64-nvidia-cuda -disable-llvm-passes -emit-llvm -o - %s | FileCheck %s
// RUN: %clang_cc1 -fsycl-is-device -internal-isystem %S/Inputs -triple nvptx-nvidia-cuda -disable-llvm-passes -emit-llvm -o - %s | FileCheck %s --check-prefixes=CHECK,CHECK-NVPTX
// RUN: %clang_cc1 -fsycl-is-device -internal-isystem %S/Inputs -triple nvptx64-nvidia-cuda -disable-llvm-passes -emit-llvm -o - %s | FileCheck %s --check-prefixes=CHECK,CHECK-NVPTX

#include "sycl.hpp"

Expand Down Expand Up @@ -123,6 +123,55 @@ int main() {
// CHECK: define {{.*}} void @{{.*}}kernel_name22() #0 {{.*}} !work_group_num_dim ![[NDRWGS1D:[0-9]+]] !reqd_work_group_size ![[WGSIZE1D22:[0-9]+]]
// CHECK: define {{.*}} void @{{.*}}kernel_name24() #0 {{.*}} !work_group_num_dim ![[NDRWGS1D:[0-9]+]] !reqd_work_group_size ![[WGSIZE1D2:[0-9]+]]

// CHECK-NVPTX: = !{ptr @{{.*}}kernel_name1, !"reqntidx", i32 16}
// CHECK-NVPTX: = !{ptr @{{.*}}kernel_name1, !"reqntidy", i32 16}
// CHECK-NVPTX: = !{ptr @{{.*}}kernel_name1, !"reqntidz", i32 32}
// CHECK-NVPTX: = !{ptr @{{.*}}kernel_name3, !"reqntidx", i32 8}
// CHECK-NVPTX: = !{ptr @{{.*}}kernel_name3, !"reqntidy", i32 8}
// CHECK-NVPTX: = !{ptr @{{.*}}kernel_name3, !"reqntidz", i32 8}
// CHECK-NVPTX: = !{ptr @{{.*}}kernel_name4, !"reqntidx", i32 2}
// CHECK-NVPTX: = !{ptr @{{.*}}kernel_name4, !"reqntidy", i32 2}
// CHECK-NVPTX: = !{ptr @{{.*}}kernel_name4, !"reqntidz", i32 2}
// CHECK-NVPTX: = !{ptr @{{.*}}kernel_name6, !"reqntidx", i32 2}
// CHECK-NVPTX: = !{ptr @{{.*}}kernel_name6, !"reqntidy", i32 8}
// CHECK-NVPTX: = !{ptr @{{.*}}kernel_name6, !"reqntidz", i32 1}
// CHECK-NVPTX: = !{ptr @{{.*}}kernel_name7, !"reqntidx", i32 16}
// CHECK-NVPTX: = !{ptr @{{.*}}kernel_name7, !"reqntidy", i32 16}
// CHECK-NVPTX: = !{ptr @{{.*}}kernel_name7, !"reqntidz", i32 32}
// CHECK-NVPTX: = !{ptr @{{.*}}kernel_name9, !"reqntidx", i32 8}
// CHECK-NVPTX: = !{ptr @{{.*}}kernel_name9, !"reqntidy", i32 8}
// CHECK-NVPTX: = !{ptr @{{.*}}kernel_name9, !"reqntidz", i32 8}
// CHECK-NVPTX: = !{ptr @{{.*}}kernel_name10, !"reqntidx", i32 2}
// CHECK-NVPTX: = !{ptr @{{.*}}kernel_name10, !"reqntidy", i32 2}
// CHECK-NVPTX: = !{ptr @{{.*}}kernel_name10, !"reqntidz", i32 2}
// CHECK-NVPTX: = !{ptr @{{.*}}kernel_name12, !"reqntidx", i32 2}
// CHECK-NVPTX: = !{ptr @{{.*}}kernel_name12, !"reqntidy", i32 8}
// CHECK-NVPTX: = !{ptr @{{.*}}kernel_name12, !"reqntidz", i32 1}
// CHECK-NVPTX: = !{ptr @{{.*}}kernel_name13, !"reqntidx", i32 16}
// CHECK-NVPTX: = !{ptr @{{.*}}kernel_name13, !"reqntidy", i32 32}
// CHECK-NVPTX-NOT: = !{ptr @{{.*}}kernel_name13, !"reqntidz"
// CHECK-NVPTX: = !{ptr @{{.*}}kernel_name15, !"reqntidx", i32 8}
// CHECK-NVPTX: = !{ptr @{{.*}}kernel_name15, !"reqntidy", i32 8}
// CHECK-NVPTX-NOT: = !{ptr @{{.*}}kernel_name15, !"reqntidz"
// CHECK-NVPTX: = !{ptr @{{.*}}kernel_name16, !"reqntidx", i32 2}
// CHECK-NVPTX: = !{ptr @{{.*}}kernel_name16, !"reqntidy", i32 2}
// CHECK-NVPTX-NOT: = !{ptr @{{.*}}kernel_name16, !"reqntidz"
// CHECK-NVPTX: = !{ptr @{{.*}}kernel_name18, !"reqntidx", i32 8}
// CHECK-NVPTX: = !{ptr @{{.*}}kernel_name18, !"reqntidy", i32 1}
// CHECK-NVPTX-NOT: = !{ptr @{{.*}}kernel_name18, !"reqntidz"
// CHECK-NVPTX: = !{ptr @{{.*}}kernel_name19, !"reqntidx", i32 32}
// CHECK-NVPTX-NOT: = !{ptr @{{.*}}kernel_name19, !"reqntidy",
// CHECK-NVPTX-NOT: = !{ptr @{{.*}}kernel_name19, !"reqntidz",
// CHECK-NVPTX: = !{ptr @{{.*}}kernel_name21, !"reqntidx", i32 8}
// CHECK-NVPTX-NOT: = !{ptr @{{.*}}kernel_name21, !"reqntidy",
// CHECK-NVPTX-NOT: = !{ptr @{{.*}}kernel_name21, !"reqntidz",
// CHECK-NVPTX: = !{ptr @{{.*}}kernel_name22, !"reqntidx", i32 2}
// CHECK-NVPTX-NOT: = !{ptr @{{.*}}kernel_name22, !"reqntidy",
// CHECK-NVPTX-NOT: = !{ptr @{{.*}}kernel_name22, !"reqntidz",
// CHECK-NVPTX: = !{ptr @{{.*}}kernel_name24, !"reqntidx", i32 1}
// CHECK-NVPTX-NOT: = !{ptr @{{.*}}kernel_name24, !"reqntidy",
// CHECK-NVPTX-NOT: = !{ptr @{{.*}}kernel_name24, !"reqntidz",

// CHECK: ![[NDRWGS3D]] = !{i32 3}
// CHECK: ![[WGSIZE3D32]] = !{i32 16, i32 16, i32 32}
// CHECK: ![[WGSIZE3D88]] = !{i32 8, i32 8, i32 8}
Expand Down
Loading