[Offload][UnitTests] Build device code as C++ #151714
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
@llvm/pr-subscribers-offload Author: Leandro Lacerda (leandrolcampos) ChangesThis commit refactors the This change enables the use of C++ features, such as templates, within device-side test kernels. This will allow for more advanced and reusable kernel wrappers, reducing boilerplate code in the conformance test suite. As part of this change:
This change affects the cc @callumfare @RossBrunton @jhuber6 Full diff: https://github.com/llvm/llvm-project/pull/151714.diff 15 Files Affected:
diff --git a/offload/unittests/CMakeLists.txt b/offload/unittests/CMakeLists.txt
index 6d165ffd4c53a..1571658c7006b 100644
--- a/offload/unittests/CMakeLists.txt
+++ b/offload/unittests/CMakeLists.txt
@@ -38,7 +38,7 @@ function(add_offload_test_device_code test_filename test_name)
set(output_file "${CMAKE_CURRENT_BINARY_DIR}/${test_name}.nvptx64.bin")
add_custom_command(
OUTPUT ${output_file}
- COMMAND ${CMAKE_C_COMPILER}
+ COMMAND ${CMAKE_CXX_COMPILER}
--target=nvptx64-nvidia-cuda -march=${nvptx_arch}
-nogpulib --cuda-path=${CUDA_ROOT} -flto ${ARGN}
${SRC_PATH} -o ${output_file}
@@ -62,7 +62,7 @@ function(add_offload_test_device_code test_filename test_name)
set(output_file "${CMAKE_CURRENT_BINARY_DIR}/${test_name}.amdgpu.bin")
add_custom_command(
OUTPUT ${output_file}
- COMMAND ${CMAKE_C_COMPILER}
+ COMMAND ${CMAKE_CXX_COMPILER}
--target=amdgcn-amd-amdhsa -mcpu=${amdgpu_arch}
-nogpulib -flto ${ARGN} ${SRC_PATH} -o ${output_file}
DEPENDS ${SRC_PATH}
diff --git a/offload/unittests/Conformance/device_code/CMakeLists.txt b/offload/unittests/Conformance/device_code/CMakeLists.txt
index 18f54b8dc5252..9cbd11096292c 100644
--- a/offload/unittests/Conformance/device_code/CMakeLists.txt
+++ b/offload/unittests/Conformance/device_code/CMakeLists.txt
@@ -1,4 +1,4 @@
-add_offload_test_device_code(LLVMLibm.c llvm-libm -stdlib -fno-builtin)
+add_offload_test_device_code(LLVMLibm.cpp llvm-libm -stdlib -fno-builtin)
add_custom_target(conformance_device_binaries DEPENDS llvm-libm.bin)
set(OFFLOAD_CONFORMANCE_DEVICE_BINARY_DIR ${CMAKE_CURRENT_BINARY_DIR} PARENT_SCOPE)
diff --git a/offload/unittests/Conformance/device_code/LLVMLibm.c b/offload/unittests/Conformance/device_code/LLVMLibm.cpp
similarity index 97%
rename from offload/unittests/Conformance/device_code/LLVMLibm.c
rename to offload/unittests/Conformance/device_code/LLVMLibm.cpp
index fe5196a539455..2c3d9bc5bf5cf 100644
--- a/offload/unittests/Conformance/device_code/LLVMLibm.c
+++ b/offload/unittests/Conformance/device_code/LLVMLibm.cpp
@@ -19,6 +19,8 @@
typedef _Float16 float16;
+extern "C" {
+
__gpu_kernel void hypotf16Kernel(const float16 *X, float16 *Y, float16 *Out,
size_t NumElements) {
uint32_t Index =
@@ -35,3 +37,4 @@ __gpu_kernel void logfKernel(const float *X, float *Out, size_t NumElements) {
if (Index < NumElements)
Out[Index] = logf(X[Index]);
}
+} // extern "C"
diff --git a/offload/unittests/OffloadAPI/device_code/CMakeLists.txt b/offload/unittests/OffloadAPI/device_code/CMakeLists.txt
index 0e4695ee9969f..50e430597e646 100644
--- a/offload/unittests/OffloadAPI/device_code/CMakeLists.txt
+++ b/offload/unittests/OffloadAPI/device_code/CMakeLists.txt
@@ -1,14 +1,14 @@
-add_offload_test_device_code(foo.c foo)
-add_offload_test_device_code(bar.c bar)
+add_offload_test_device_code(foo.cpp foo)
+add_offload_test_device_code(bar.cpp bar)
# Compile with optimizations to eliminate AMDGPU implicit arguments.
-add_offload_test_device_code(noargs.c noargs -O3)
-add_offload_test_device_code(localmem.c localmem)
-add_offload_test_device_code(localmem_reduction.c localmem_reduction)
-add_offload_test_device_code(localmem_static.c localmem_static)
-add_offload_test_device_code(global.c global)
-add_offload_test_device_code(global_ctor.c global_ctor)
-add_offload_test_device_code(global_dtor.c global_dtor)
-add_offload_test_device_code(sequence.c sequence)
+add_offload_test_device_code(noargs.cpp noargs -O3)
+add_offload_test_device_code(localmem.cpp localmem)
+add_offload_test_device_code(localmem_reduction.cpp localmem_reduction)
+add_offload_test_device_code(localmem_static.cpp localmem_static)
+add_offload_test_device_code(global.cpp global)
+add_offload_test_device_code(global_ctor.cpp global_ctor)
+add_offload_test_device_code(global_dtor.cpp global_dtor)
+add_offload_test_device_code(sequence.cpp sequence)
add_custom_target(offload_device_binaries DEPENDS
foo.bin
diff --git a/offload/unittests/OffloadAPI/device_code/bar.c b/offload/unittests/OffloadAPI/device_code/bar.cpp
similarity index 63%
rename from offload/unittests/OffloadAPI/device_code/bar.c
rename to offload/unittests/OffloadAPI/device_code/bar.cpp
index 786aa2f5d61e7..b5191671f293f 100644
--- a/offload/unittests/OffloadAPI/device_code/bar.c
+++ b/offload/unittests/OffloadAPI/device_code/bar.cpp
@@ -1,5 +1,5 @@
#include <gpuintrin.h>
-__gpu_kernel void foo(int *out) {
+extern "C" __gpu_kernel void foo(int *out) {
out[__gpu_thread_id(0)] = __gpu_thread_id(0) + 1;
}
diff --git a/offload/unittests/OffloadAPI/device_code/foo.c b/offload/unittests/OffloadAPI/device_code/foo.cpp
similarity index 65%
rename from offload/unittests/OffloadAPI/device_code/foo.c
rename to offload/unittests/OffloadAPI/device_code/foo.cpp
index 83cdc53cddd8d..cdc20015fc3e2 100644
--- a/offload/unittests/OffloadAPI/device_code/foo.c
+++ b/offload/unittests/OffloadAPI/device_code/foo.cpp
@@ -1,6 +1,6 @@
#include <gpuintrin.h>
#include <stdint.h>
-__gpu_kernel void foo(uint32_t *out) {
+extern "C" __gpu_kernel void foo(uint32_t *out) {
out[__gpu_thread_id(0)] = __gpu_thread_id(0);
}
diff --git a/offload/unittests/OffloadAPI/device_code/global.c b/offload/unittests/OffloadAPI/device_code/global.cpp
similarity index 92%
rename from offload/unittests/OffloadAPI/device_code/global.c
rename to offload/unittests/OffloadAPI/device_code/global.cpp
index 9f27f9424324f..dada16c87766c 100644
--- a/offload/unittests/OffloadAPI/device_code/global.c
+++ b/offload/unittests/OffloadAPI/device_code/global.cpp
@@ -1,6 +1,8 @@
#include <gpuintrin.h>
#include <stdint.h>
+extern "C" {
+
[[gnu::visibility("default")]]
uint32_t global[64];
@@ -13,3 +15,4 @@ __gpu_kernel void read(uint32_t *out) {
out[__gpu_thread_id(0) + (__gpu_num_threads(0) * __gpu_block_id(0))] =
global[__gpu_thread_id(0)];
}
+} // extern "C"
diff --git a/offload/unittests/OffloadAPI/device_code/global_ctor.c b/offload/unittests/OffloadAPI/device_code/global_ctor.cpp
similarity index 95%
rename from offload/unittests/OffloadAPI/device_code/global_ctor.c
rename to offload/unittests/OffloadAPI/device_code/global_ctor.cpp
index 27e2d71d7566e..a14f1d59bf950 100644
--- a/offload/unittests/OffloadAPI/device_code/global_ctor.c
+++ b/offload/unittests/OffloadAPI/device_code/global_ctor.cpp
@@ -1,6 +1,8 @@
#include <gpuintrin.h>
#include <stdint.h>
+extern "C" {
+
uint32_t global[64];
[[gnu::constructor(202)]] void ctorc() {
@@ -23,3 +25,4 @@ __gpu_kernel void global_ctor(uint32_t *out) {
out[__gpu_thread_id(0) + (__gpu_num_threads(0) * __gpu_block_id(0))] =
global[__gpu_thread_id(0)];
}
+} // extern "C"
diff --git a/offload/unittests/OffloadAPI/device_code/global_dtor.c b/offload/unittests/OffloadAPI/device_code/global_dtor.cpp
similarity index 87%
rename from offload/unittests/OffloadAPI/device_code/global_dtor.c
rename to offload/unittests/OffloadAPI/device_code/global_dtor.cpp
index cadcc19cc296b..6b1f941342b40 100644
--- a/offload/unittests/OffloadAPI/device_code/global_dtor.c
+++ b/offload/unittests/OffloadAPI/device_code/global_dtor.cpp
@@ -1,6 +1,8 @@
#include <gpuintrin.h>
#include <stdint.h>
+extern "C" {
+
uint32_t global[64];
[[gnu::destructor]] void dtor() {
@@ -11,3 +13,4 @@ uint32_t global[64];
__gpu_kernel void global_dtor() {
// no-op
}
+} // extern "C"
diff --git a/offload/unittests/OffloadAPI/device_code/localmem.c b/offload/unittests/OffloadAPI/device_code/localmem.cpp
similarity index 84%
rename from offload/unittests/OffloadAPI/device_code/localmem.c
rename to offload/unittests/OffloadAPI/device_code/localmem.cpp
index d70847900bc43..9542e2cb1d648 100644
--- a/offload/unittests/OffloadAPI/device_code/localmem.c
+++ b/offload/unittests/OffloadAPI/device_code/localmem.cpp
@@ -3,7 +3,7 @@
extern __gpu_local uint32_t shared_mem[];
-__gpu_kernel void localmem(uint32_t *out) {
+extern "C" __gpu_kernel void localmem(uint32_t *out) {
shared_mem[__gpu_thread_id(0)] = __gpu_thread_id(0);
shared_mem[__gpu_thread_id(0)] *= 2;
out[__gpu_thread_id(0) + (__gpu_num_threads(0) * __gpu_block_id(0))] =
diff --git a/offload/unittests/OffloadAPI/device_code/localmem_reduction.c b/offload/unittests/OffloadAPI/device_code/localmem_reduction.cpp
similarity index 83%
rename from offload/unittests/OffloadAPI/device_code/localmem_reduction.c
rename to offload/unittests/OffloadAPI/device_code/localmem_reduction.cpp
index 8a9a46cfb6a11..2c0a3e80b16e7 100644
--- a/offload/unittests/OffloadAPI/device_code/localmem_reduction.c
+++ b/offload/unittests/OffloadAPI/device_code/localmem_reduction.cpp
@@ -3,7 +3,7 @@
extern __gpu_local uint32_t shared_mem[];
-__gpu_kernel void localmem_reduction(uint32_t *out) {
+extern "C" __gpu_kernel void localmem_reduction(uint32_t *out) {
shared_mem[__gpu_thread_id(0)] = 2;
__gpu_sync_threads();
diff --git a/offload/unittests/OffloadAPI/device_code/localmem_static.c b/offload/unittests/OffloadAPI/device_code/localmem_static.cpp
similarity index 85%
rename from offload/unittests/OffloadAPI/device_code/localmem_static.c
rename to offload/unittests/OffloadAPI/device_code/localmem_static.cpp
index 928b48422a0d6..a8dd95473742c 100644
--- a/offload/unittests/OffloadAPI/device_code/localmem_static.c
+++ b/offload/unittests/OffloadAPI/device_code/localmem_static.cpp
@@ -4,7 +4,7 @@
[[clang::loader_uninitialized]]
__gpu_local uint32_t shared_mem[64];
-__gpu_kernel void localmem_static(uint32_t *out) {
+extern "C" __gpu_kernel void localmem_static(uint32_t *out) {
shared_mem[__gpu_thread_id(0)] = 2;
__gpu_sync_threads();
diff --git a/offload/unittests/OffloadAPI/device_code/noargs.c b/offload/unittests/OffloadAPI/device_code/noargs.c
deleted file mode 100644
index 36e609aa26a09..0000000000000
--- a/offload/unittests/OffloadAPI/device_code/noargs.c
+++ /dev/null
@@ -1,3 +0,0 @@
-#include <gpuintrin.h>
-
-__gpu_kernel void noargs() { (void)0; }
diff --git a/offload/unittests/OffloadAPI/device_code/noargs.cpp b/offload/unittests/OffloadAPI/device_code/noargs.cpp
new file mode 100644
index 0000000000000..58f989c714fed
--- /dev/null
+++ b/offload/unittests/OffloadAPI/device_code/noargs.cpp
@@ -0,0 +1,3 @@
+#include <gpuintrin.h>
+
+extern "C" __gpu_kernel void noargs() { (void)0; }
diff --git a/offload/unittests/OffloadAPI/device_code/sequence.c b/offload/unittests/OffloadAPI/device_code/sequence.cpp
similarity index 71%
rename from offload/unittests/OffloadAPI/device_code/sequence.c
rename to offload/unittests/OffloadAPI/device_code/sequence.cpp
index 7662f2d817496..07f92944346f5 100644
--- a/offload/unittests/OffloadAPI/device_code/sequence.c
+++ b/offload/unittests/OffloadAPI/device_code/sequence.cpp
@@ -1,7 +1,7 @@
#include <gpuintrin.h>
#include <stdint.h>
-__gpu_kernel void sequence(uint32_t idx, uint32_t *inout) {
+extern "C" __gpu_kernel void sequence(uint32_t idx, uint32_t *inout) {
if (idx == 0)
inout[idx] = 0;
else if (idx == 1)
|
jhuber6
requested review from
RossBrunton,
jhuber6 and
callumfare
and removed request for
jhuber6
jhuber6
approved these changes
Aug 1, 2025
leandrolcampos
force-pushed
the
CppDeviceCode
branch
from
cd484f0 to
86401a4
Compare
|
What C++ features do you want to use here? I'm weary of complexity creep in what should be very simple source files. |
|
Hi @RossBrunton, The primary C++ feature I need is templates, specifically for the new conformance tests in Without C++, the main alternative would be to use C preprocessor macros to generate this boilerplate. By compiling as C++, I can instead use a single, type-safe variadic template helper. The intent is not to add complex C++ logic to the kernels themselves, but only to use templates as a safer and more maintainable alternative to macros. To be clear, this use of templates will be confined to the new conformance tests, that share the same CMake function |
|
My preference in that case would still be macros, but I've not looked into this so if you have a template solution that is cleaner I'm happy with that. |
RossBrunton
approved these changes
Aug 4, 2025
C/C++ doesn't make a huge difference, just requires the C ABI on the symbols you look up. Usually a template function is easier to read than a bunch of macros. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This commit refactors the
add_offload_test_device_codeCMake function to compile device code using the C++ compiler (CMAKE_CXX_COMPILER) instead of the C compiler.This change enables the use of C++ features, such as templates, within device-side test kernels. This will allow for more advanced and reusable kernel wrappers, reducing boilerplate code in the conformance test suite.
As part of this change:
.cfiles for device code inunittests/have been renamed to.cpp.extern "C"to ensure C linkage and prevent name mangling.This change affects the
OffloadAPIandConformancetest suites.cc @callumfare @RossBrunton @jhuber6