Add ARM build configs for latest gcc/clang. (#468)

Co-authored-by: Michael Schellenberger Costa <[email protected]>

.github/workflows/dispatch-build-and-test.yml

@@ -22,7 +22,7 @@ jobs:
         include: ${{ fromJSON(inputs.per_cuda_compiler_matrix) }}
     with:
       cpu: ${{ matrix.cpu }}
-      test_name: ${{matrix.compiler.name}}${{matrix.compiler.version}}/C++${{matrix.std}}
+      test_name: ${{matrix.cpu}}/${{matrix.compiler.name}}${{matrix.compiler.version}}/C++${{matrix.std}}
       build_script: "./ci/build_${{ inputs.project_name }}.sh -cxx ${{matrix.compiler.exe}} -std ${{matrix.std}}"
       test_script:  "./ci/test_${{ inputs.project_name }}.sh -cxx ${{matrix.compiler.exe}} -std ${{matrix.std}}"
       container_image: rapidsai/devcontainers:${{inputs.devcontainer_version}}-cpp-${{matrix.compiler.name}}${{matrix.compiler.version}}-cuda${{matrix.cuda}}-${{matrix.os}}
@@ -37,6 +37,6 @@ jobs:
       matrix:
         include: ${{ fromJSON(inputs.per_cuda_compiler_matrix) }}
     with:
-      test_name: ${{matrix.compiler.name}}${{matrix.compiler.version}}/C++${{matrix.std}}
+      test_name: ${{matrix.cpu}}/${{matrix.compiler.name}}${{matrix.compiler.version}}/C++${{matrix.std}}
       build_script: "./ci/windows/build_${{ inputs.project_name }}.ps1 ${{matrix.std}}"
       container_image: rapidsai/devcontainers:${{inputs.devcontainer_version}}-cuda${{matrix.cuda}}-${{matrix.compiler.name}}${{matrix.compiler.version}}-${{matrix.os}}

ci/matrix.yaml

@@ -35,6 +35,7 @@ pull_request:
     - {cuda: *cuda_newest, os: 'ubuntu20.04', cpu: 'amd64', compiler: {name: 'gcc',  version: '10', exe: 'g++'},     std: [11, 14, 17, 20], jobs: ['build']}
     - {cuda: *cuda_newest, os: 'ubuntu22.04', cpu: 'amd64', compiler: {name: 'gcc',  version: '11', exe: 'g++'},     std: [11, 14, 17, 20], jobs: ['build']}
     - {cuda: *cuda_newest, os: 'ubuntu22.04', cpu: 'amd64', compiler: {name: 'gcc',  version: '12', exe: 'g++'},     std: [11, 14, 17, 20], jobs: ['build', 'test']}
+    - {cuda: *cuda_newest, os: 'ubuntu22.04', cpu: 'arm64', compiler: {name: 'gcc',  version: '12', exe: 'g++'},     std: [11, 14, 17, 20], jobs: ['build']}
     - {cuda: *cuda_oldest, os: 'ubuntu18.04', cpu: 'amd64', compiler: {name: 'llvm', version: '9',  exe: 'clang++'}, std: [11, 14, 17],     jobs: ['build']}
     - {cuda: *cuda_newest, os: 'ubuntu20.04', cpu: 'amd64', compiler: {name: 'llvm', version: '9',  exe: 'clang++'}, std: [11, 14, 17],     jobs: ['build']}
     - {cuda: *cuda_newest, os: 'ubuntu20.04', cpu: 'amd64', compiler: {name: 'llvm', version: '10', exe: 'clang++'}, std: [11, 14, 17],     jobs: ['build']}
@@ -44,6 +45,7 @@ pull_request:
     - {cuda: *cuda_newest, os: 'ubuntu20.04', cpu: 'amd64', compiler: {name: 'llvm', version: '14', exe: 'clang++'}, std: [11, 14, 17, 20], jobs: ['build']}
     - {cuda: *cuda_newest, os: 'ubuntu22.04', cpu: 'amd64', compiler: {name: 'llvm', version: '15', exe: 'clang++'}, std: [11, 14, 17, 20], jobs: ['build']}
     - {cuda: *cuda_newest, os: 'ubuntu22.04', cpu: 'amd64', compiler: {name: 'llvm', version: '16', exe: 'clang++'}, std: [11, 14, 17, 20], jobs: ['build', 'test']}
+    - {cuda: *cuda_newest, os: 'ubuntu22.04', cpu: 'arm64', compiler: {name: 'llvm', version: '16', exe: 'clang++'}, std: [11, 14, 17, 20], jobs: ['build']}
     - {cuda: *cuda_oldest, os: 'windows2022', cpu: 'amd64', compiler: {name: 'cl',   version: '14.16', exe: 'cl++'}, std: [14, 17],         jobs: ['build']}
     - {cuda: *cuda_newest, os: 'windows2022', cpu: 'amd64', compiler: {name: 'cl',   version: '14.29', exe: 'cl++'}, std: [14, 17],         jobs: ['build']}
     - {cuda: *cuda_newest, os: 'windows2022', cpu: 'amd64', compiler: {name: 'cl',   version: '14.36', exe: 'cl++'}, std: [14, 17, 20],     jobs: ['build']}

thrust/testing/functional.cu

@@ -7,9 +7,18 @@
 
 THRUST_DISABLE_MSVC_POSSIBLE_LOSS_OF_DATA_WARNING_BEGIN
 
+// There is a unfortunate miscompilation of the gcc-12 vectorizer leading to OOB writes
+// Adding this attribute suffices that this miscompilation does not appear anymore
+#if (THRUST_HOST_COMPILER == THRUST_HOST_COMPILER_GCC) && __GNUC__ >= 12 && THRUST_CPP_DIALECT >= 2020
+#define THRUST_DISABLE_BROKEN_GCC_VECTORIZER __attribute__((optimize("no-tree-vectorize")))
+#else
+#define THRUST_DISABLE_BROKEN_GCC_VECTORIZER
+#endif
+
 const size_t NUM_SAMPLES = 10000;
 
 template <class InputVector, class OutputVector, class Operator, class ReferenceOperator>
+THRUST_DISABLE_BROKEN_GCC_VECTORIZER
 void TestUnaryFunctional(void)
 {
     typedef typename InputVector::value_type  InputType;
@@ -28,6 +37,7 @@ void TestUnaryFunctional(void)
 }
 
 template <class InputVector, class OutputVector, class Operator, class ReferenceOperator>
+THRUST_DISABLE_BROKEN_GCC_VECTORIZER
 void TestBinaryFunctional(void)
 {
     typedef typename InputVector::value_type  InputType;
@@ -51,8 +61,6 @@ void TestBinaryFunctional(void)
     ASSERT_ALMOST_EQUAL(output, std_output);
 }
 
-
-
 // XXX add bool to list
 // Instantiate a macro for all integer-like data types
 #define INSTANTIATE_INTEGER_TYPES(Macro, vector_type, operator_name)   \
@@ -175,6 +183,7 @@ DECLARE_UNARY_LOGICAL_FUNCTIONAL_UNITTEST(logical_not, LogicalNot);
 
 // Ad-hoc testing for other functionals
 template <class Vector>
+THRUST_DISABLE_BROKEN_GCC_VECTORIZER
 void TestIdentityFunctional(void)
 {
     typedef typename Vector::value_type T;
@@ -191,6 +200,7 @@ void TestIdentityFunctional(void)
 DECLARE_VECTOR_UNITTEST(TestIdentityFunctional);
 
 template <class Vector>
+THRUST_DISABLE_BROKEN_GCC_VECTORIZER
 void TestProject1stFunctional(void)
 {
     typedef typename Vector::value_type T;
@@ -211,6 +221,7 @@ void TestProject1stFunctional(void)
 DECLARE_VECTOR_UNITTEST(TestProject1stFunctional);
 
 template <class Vector>
+THRUST_DISABLE_BROKEN_GCC_VECTORIZER
 void TestProject2ndFunctional(void)
 {
     typedef typename Vector::value_type T;
@@ -231,6 +242,7 @@ void TestProject2ndFunctional(void)
 DECLARE_VECTOR_UNITTEST(TestProject2ndFunctional);
 
 template <class Vector>
+THRUST_DISABLE_BROKEN_GCC_VECTORIZER
 void TestMaximumFunctional(void)
 {
     typedef typename Vector::value_type T;
@@ -255,6 +267,7 @@ void TestMaximumFunctional(void)
 DECLARE_VECTOR_UNITTEST(TestMaximumFunctional);
 
 template <class Vector>
+THRUST_DISABLE_BROKEN_GCC_VECTORIZER
 void TestMinimumFunctional(void)
 {
     typedef typename Vector::value_type T;
@@ -279,6 +292,7 @@ void TestMinimumFunctional(void)
 DECLARE_VECTOR_UNITTEST(TestMinimumFunctional);
 
 template <class Vector>
+THRUST_DISABLE_BROKEN_GCC_VECTORIZER
 void TestNot1(void)
 {
     typedef typename Vector::value_type T;
@@ -314,6 +328,7 @@ DECLARE_INTEGRAL_VECTOR_UNITTEST(TestNot1);
       THRUST_CPP_DIALECT == 2011)
 
 template <class Vector>
+THRUST_DISABLE_BROKEN_GCC_VECTORIZER
 void TestNot2(void)
 {
     typedef typename Vector::value_type T;

thrust/testing/transform.cu

@@ -7,8 +7,16 @@
 #include <thrust/iterator/zip_iterator.h>
 #include <thrust/iterator/retag.h>
 
+// There is a unfortunate miscompilation of the gcc-12 vectorizer leading to OOB writes
+// Adding this attribute suffices that this miscompilation does not appear anymore
+#if (THRUST_HOST_COMPILER == THRUST_HOST_COMPILER_GCC) && __GNUC__ >= 12 && THRUST_CPP_DIALECT >= 2020
+#define THRUST_DISABLE_BROKEN_GCC_VECTORIZER __attribute__((optimize("no-tree-vectorize")))
+#else
+#define THRUST_DISABLE_BROKEN_GCC_VECTORIZER
+#endif
 
 template <class Vector>
+THRUST_DISABLE_BROKEN_GCC_VECTORIZER
 void TestTransformUnarySimple(void)
 {
     typedef typename Vector::value_type T;
@@ -78,6 +86,7 @@ DECLARE_UNITTEST(TestTransformUnaryDispatchImplicit);
 
 
 template <class Vector>
+THRUST_DISABLE_BROKEN_GCC_VECTORIZER
 void TestTransformIfUnaryNoStencilSimple(void)
 {
     typedef typename Vector::value_type T;
@@ -166,6 +175,7 @@ DECLARE_UNITTEST(TestTransformIfUnaryNoStencilDispatchImplicit);
 
 
 template <class Vector>
+THRUST_DISABLE_BROKEN_GCC_VECTORIZER
 void TestTransformIfUnarySimple(void)
 {
     typedef typename Vector::value_type T;
@@ -259,6 +269,7 @@ DECLARE_UNITTEST(TestTransformIfUnaryDispatchImplicit);
 
 
 template <class Vector>
+THRUST_DISABLE_BROKEN_GCC_VECTORIZER
 void TestTransformBinarySimple(void)
 {
     typedef typename Vector::value_type T;
@@ -339,6 +350,7 @@ DECLARE_UNITTEST(TestTransformBinaryDispatchImplicit);
 
 
 template <class Vector>
+THRUST_DISABLE_BROKEN_GCC_VECTORIZER
 void TestTransformIfBinarySimple(void)
 {
     typedef typename Vector::value_type T;
@@ -447,6 +459,7 @@ DECLARE_UNITTEST(TestTransformIfBinaryDispatchImplicit);
 
 
 template <typename T>
+THRUST_DISABLE_BROKEN_GCC_VECTORIZER
 void TestTransformUnary(const size_t n)
 {
     thrust::host_vector<T>   h_input = unittest::random_integers<T>(n);
@@ -464,6 +477,7 @@ DECLARE_VARIABLE_UNITTEST(TestTransformUnary);
 
 
 template <typename T>
+THRUST_DISABLE_BROKEN_GCC_VECTORIZER
 void TestTransformUnaryToDiscardIterator(const size_t n)
 {
     thrust::host_vector<T>   h_input = unittest::random_integers<T>(n);
@@ -495,6 +509,7 @@ struct repeat2
 
 
 template<typename T>
+THRUST_DISABLE_BROKEN_GCC_VECTORIZER
 void TestTransformUnaryToDiscardIteratorZipped(const size_t n)
 {
     thrust::host_vector<T>   h_input = unittest::random_integers<T>(n);
@@ -542,6 +557,7 @@ struct is_positive
 
 
 template <typename T>
+THRUST_DISABLE_BROKEN_GCC_VECTORIZER
 void TestTransformIfUnaryNoStencil(const size_t n)
 {
     thrust::host_vector<T>   h_input   = unittest::random_integers<T>(n);
@@ -564,6 +580,7 @@ DECLARE_VARIABLE_UNITTEST(TestTransformIfUnaryNoStencil);
 
 
 template <typename T>
+THRUST_DISABLE_BROKEN_GCC_VECTORIZER
 void TestTransformIfUnary(const size_t n)
 {
     thrust::host_vector<T>   h_input   = unittest::random_integers<T>(n);
@@ -590,6 +607,7 @@ DECLARE_VARIABLE_UNITTEST(TestTransformIfUnary);
 
 
 template <typename T>
+THRUST_DISABLE_BROKEN_GCC_VECTORIZER
 void TestTransformIfUnaryToDiscardIterator(const size_t n)
 {
     thrust::host_vector<T>   h_input   = unittest::random_integers<T>(n);
@@ -619,6 +637,7 @@ DECLARE_VARIABLE_UNITTEST(TestTransformIfUnaryToDiscardIterator);
 
 
 template <typename T>
+THRUST_DISABLE_BROKEN_GCC_VECTORIZER
 void TestTransformBinary(const size_t n)
 {
     thrust::host_vector<T>   h_input1 = unittest::random_integers<T>(n);
@@ -643,6 +662,7 @@ DECLARE_VARIABLE_UNITTEST(TestTransformBinary);
 
 
 template <typename T>
+THRUST_DISABLE_BROKEN_GCC_VECTORIZER
 void TestTransformBinaryToDiscardIterator(const size_t n)
 {
     thrust::host_vector<T>   h_input1 = unittest::random_integers<T>(n);
@@ -664,6 +684,7 @@ DECLARE_VARIABLE_UNITTEST(TestTransformBinaryToDiscardIterator);
 
 
 template <typename T>
+THRUST_DISABLE_BROKEN_GCC_VECTORIZER
 void TestTransformIfBinary(const size_t n)
 {
     thrust::host_vector<T>   h_input1  = unittest::random_integers<T>(n);
@@ -711,6 +732,7 @@ DECLARE_VARIABLE_UNITTEST(TestTransformIfBinary);
 
 
 template <typename T>
+THRUST_DISABLE_BROKEN_GCC_VECTORIZER
 void TestTransformIfBinaryToDiscardIterator(const size_t n)
 {
     thrust::host_vector<T>   h_input1  = unittest::random_integers<T>(n);
@@ -745,6 +767,7 @@ DECLARE_VARIABLE_UNITTEST(TestTransformIfBinaryToDiscardIterator);
 
 #if ((__GNUC__ * 10000 + __GNUC_MINOR__ * 100) == 40400) || defined(__INTEL_COMPILER)
 template <typename T>
+THRUST_DISABLE_BROKEN_GCC_VECTORIZER
 void TestTransformUnaryCountingIterator()
 {
     // G++ 4.4.x has a known failure with auto-vectorization (due to -O3 or
@@ -758,6 +781,7 @@ void TestTransformUnaryCountingIterator()
 }
 #else
 template <typename T>
+THRUST_DISABLE_BROKEN_GCC_VECTORIZER
 void TestTransformUnaryCountingIterator()
 {
     size_t const n = 15 * sizeof(T);
@@ -780,6 +804,7 @@ DECLARE_GENERIC_UNITTEST(TestTransformUnaryCountingIterator);
 
 #if (__GNUC__ * 10000 + __GNUC_MINOR__ * 100) == 40400
 template <typename T>
+THRUST_DISABLE_BROKEN_GCC_VECTORIZER
 void TestTransformBinaryCountingIterator()
 {
     // GCC 4.4.x has a known failure with auto-vectorization (due to -O3 or -ftree-vectorize) of this test
@@ -789,6 +814,7 @@ void TestTransformBinaryCountingIterator()
 }
 #else
 template <typename T>
+THRUST_DISABLE_BROKEN_GCC_VECTORIZER
 void TestTransformBinaryCountingIterator()
 {
     size_t const n = 15 * sizeof(T);
@@ -825,6 +851,7 @@ struct plus_mod3
 };
 
 template <typename Vector>
+THRUST_DISABLE_BROKEN_GCC_VECTORIZER
 void TestTransformWithIndirection(void)
 {
     // add numbers modulo 3 with external lookup table