[llvm] [Offload] Add framework for math conformance tests (PR #149242)

Leandro Lacerda via llvm-commits llvm-commits at lists.llvm.org
Fri Jul 18 11:00:20 PDT 2025


https://github.com/leandrolcampos updated https://github.com/llvm/llvm-project/pull/149242

>From 21b1fe1d366b60b56a858d847a245ebabcab11bc Mon Sep 17 00:00:00 2001
From: Leandro Augusto Lacerda Campos <leandrolcampos at yahoo.com.br>
Date: Thu, 17 Jul 2025 00:15:00 -0300
Subject: [PATCH 1/5] [Offload] Add framework for math conformance tests

---
 offload/unittests/CMakeLists.txt              |  11 +-
 offload/unittests/Conformance/CMakeLists.txt  |   7 +-
 .../Conformance/device_code/CMakeLists.txt    |   6 +-
 .../Conformance/device_code/LLVMLibm.c        |  30 +++
 .../unittests/Conformance/device_code/sin.c   |   4 -
 .../include/mathtest/DeviceContext.hpp        | 121 +++++++++
 .../include/mathtest/DeviceResources.hpp      | 129 ++++++++++
 .../Conformance/include/mathtest/Dim.hpp      |  42 ++++
 .../include/mathtest/ErrorHandling.hpp        |  29 +++
 .../include/mathtest/ExhaustiveGenerator.hpp  | 139 +++++++++++
 .../include/mathtest/GpuMathTest.hpp          | 159 ++++++++++++
 .../include/mathtest/HostRefChecker.hpp       |  82 ++++++
 .../include/mathtest/IndexedRange.hpp         |  91 +++++++
 .../include/mathtest/InputGenerator.hpp       |  14 ++
 .../Conformance/include/mathtest/Numerics.hpp | 235 ++++++++++++++++++
 .../include/mathtest/OffloadForward.hpp       |  22 ++
 .../Conformance/include/mathtest/Support.hpp  | 138 ++++++++++
 .../include/mathtest/TestResult.hpp           |  70 ++++++
 .../include/mathtest/TestRunner.hpp           | 118 +++++++++
 .../include/mathtest/TypeExtras.hpp           |   9 +
 .../unittests/Conformance/lib/CMakeLists.txt  |   5 +
 .../Conformance/lib/DeviceContext.cpp         | 201 +++++++++++++++
 .../Conformance/lib/DeviceResources.cpp       |  54 ++++
 .../Conformance/lib/ErrorHandling.cpp         |  37 +++
 offload/unittests/Conformance/sin.cpp         |   8 -
 .../Conformance/tests/CMakeLists.txt          |   2 +
 .../Conformance/tests/Hypotf16Test.cpp        |  51 ++++
 .../unittests/Conformance/tests/LogfTest.cpp  |  44 ++++
 28 files changed, 1834 insertions(+), 24 deletions(-)
 create mode 100644 offload/unittests/Conformance/device_code/LLVMLibm.c
 delete mode 100644 offload/unittests/Conformance/device_code/sin.c
 create mode 100644 offload/unittests/Conformance/include/mathtest/DeviceContext.hpp
 create mode 100644 offload/unittests/Conformance/include/mathtest/DeviceResources.hpp
 create mode 100644 offload/unittests/Conformance/include/mathtest/Dim.hpp
 create mode 100644 offload/unittests/Conformance/include/mathtest/ErrorHandling.hpp
 create mode 100644 offload/unittests/Conformance/include/mathtest/ExhaustiveGenerator.hpp
 create mode 100644 offload/unittests/Conformance/include/mathtest/GpuMathTest.hpp
 create mode 100644 offload/unittests/Conformance/include/mathtest/HostRefChecker.hpp
 create mode 100644 offload/unittests/Conformance/include/mathtest/IndexedRange.hpp
 create mode 100644 offload/unittests/Conformance/include/mathtest/InputGenerator.hpp
 create mode 100644 offload/unittests/Conformance/include/mathtest/Numerics.hpp
 create mode 100644 offload/unittests/Conformance/include/mathtest/OffloadForward.hpp
 create mode 100644 offload/unittests/Conformance/include/mathtest/Support.hpp
 create mode 100644 offload/unittests/Conformance/include/mathtest/TestResult.hpp
 create mode 100644 offload/unittests/Conformance/include/mathtest/TestRunner.hpp
 create mode 100644 offload/unittests/Conformance/include/mathtest/TypeExtras.hpp
 create mode 100644 offload/unittests/Conformance/lib/CMakeLists.txt
 create mode 100644 offload/unittests/Conformance/lib/DeviceContext.cpp
 create mode 100644 offload/unittests/Conformance/lib/DeviceResources.cpp
 create mode 100644 offload/unittests/Conformance/lib/ErrorHandling.cpp
 delete mode 100644 offload/unittests/Conformance/sin.cpp
 create mode 100644 offload/unittests/Conformance/tests/CMakeLists.txt
 create mode 100644 offload/unittests/Conformance/tests/Hypotf16Test.cpp
 create mode 100644 offload/unittests/Conformance/tests/LogfTest.cpp

diff --git a/offload/unittests/CMakeLists.txt b/offload/unittests/CMakeLists.txt
index 388d15f834b1d..bea49387528b0 100644
--- a/offload/unittests/CMakeLists.txt
+++ b/offload/unittests/CMakeLists.txt
@@ -39,9 +39,9 @@ function(add_offload_test_device_code test_filename test_name)
       add_custom_command(
         OUTPUT ${output_file}
         COMMAND ${CMAKE_C_COMPILER}
-        --target=nvptx64-nvidia-cuda -march=${nvptx_arch}
-        -nogpulib --cuda-path=${CUDA_ROOT} -flto ${ARGN}
-        -c ${SRC_PATH} -o ${output_file}
+        --target=nvptx64-nvidia-cuda -march=native
+        -stdlib -nogpulib --cuda-path=${CUDA_ROOT} -flto -fno-builtin ${ARGN}
+        ${SRC_PATH} -o ${output_file}
         DEPENDS ${SRC_PATH}
       )
       add_custom_target(${test_name}.nvptx64 DEPENDS ${output_file})
@@ -62,6 +62,8 @@ function(add_offload_test_device_code test_filename test_name)
       set(output_file "${CMAKE_CURRENT_BINARY_DIR}/${test_name}.amdgpu.bin")
       add_custom_command(
         OUTPUT ${output_file}
+        # TODO(jhuber6): Add and test the '-stdlib' flag here; also consider
+        #                the '-fno-builtin' flag.
         COMMAND ${CMAKE_C_COMPILER}
         --target=amdgcn-amd-amdhsa -mcpu=${amdgpu_arch}
         -nogpulib -flto ${ARGN} -c ${SRC_PATH} -o ${output_file}
@@ -106,10 +108,9 @@ function(add_conformance_test test_name)
   endif()
 
   add_executable(${target_name} ${files})
-  add_dependencies(${target_name} ${PLUGINS_TEST_COMMON} ${test_name}.bin)
+  add_dependencies(${target_name} conformance_device_binaries)
   target_compile_definitions(${target_name} PRIVATE DEVICE_CODE_PATH="${CONFORMANCE_TEST_DEVICE_CODE_PATH}")
   target_link_libraries(${target_name} PRIVATE ${PLUGINS_TEST_COMMON} libc)
-  target_include_directories(${target_name} PRIVATE ${PLUGINS_TEST_INCLUDE})
   set_target_properties(${target_name} PROPERTIES EXCLUDE_FROM_ALL TRUE)
 
   add_custom_target(offload.conformance.${test_name}
diff --git a/offload/unittests/Conformance/CMakeLists.txt b/offload/unittests/Conformance/CMakeLists.txt
index bc3141757372a..ce0421553de05 100644
--- a/offload/unittests/Conformance/CMakeLists.txt
+++ b/offload/unittests/Conformance/CMakeLists.txt
@@ -1,8 +1,7 @@
 add_custom_target(offload.conformance)
 
-set(PLUGINS_TEST_COMMON LLVMOffload LLVMSupport)
-set(PLUGINS_TEST_INCLUDE ${LIBOMPTARGET_INCLUDE_DIR} ${CMAKE_CURRENT_SOURCE_DIR}/common)
+set(PLUGINS_TEST_COMMON MathTest)
 
 add_subdirectory(device_code)
-
-add_conformance_test(sin sin.cpp)
+add_subdirectory(lib)
+add_subdirectory(tests)
diff --git a/offload/unittests/Conformance/device_code/CMakeLists.txt b/offload/unittests/Conformance/device_code/CMakeLists.txt
index 223f04ccfb698..82c6ec9767562 100644
--- a/offload/unittests/Conformance/device_code/CMakeLists.txt
+++ b/offload/unittests/Conformance/device_code/CMakeLists.txt
@@ -1,4 +1,4 @@
-# FIXME: Currently missing dependencies to build GPU portion automatically.
-add_offload_test_device_code(sin.c sin)
+add_offload_test_device_code(LLVMLibm.c LLVMLibm)
 
-set(OFFLOAD_TEST_DEVICE_CODE_PATH ${CMAKE_CURRENT_BINARY_DIR} PARENT_SCOPE)
+add_custom_target(conformance_device_binaries DEPENDS LLVMLibm.bin)
+set(CONFORMANCE_TEST_DEVICE_CODE_PATH ${CMAKE_CURRENT_BINARY_DIR} PARENT_SCOPE)
diff --git a/offload/unittests/Conformance/device_code/LLVMLibm.c b/offload/unittests/Conformance/device_code/LLVMLibm.c
new file mode 100644
index 0000000000000..5f436a235eb08
--- /dev/null
+++ b/offload/unittests/Conformance/device_code/LLVMLibm.c
@@ -0,0 +1,30 @@
+#include <gpuintrin.h>
+#include <math.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#ifdef __FLT16_MAX__
+#define HAS_FLOAT16
+typedef _Float16 float16;
+#endif
+
+#ifdef HAS_FLOAT16
+__gpu_kernel void hypotf16Kernel(const float16 *X, float16 *Y, float16 *Out,
+                                 size_t NumElements) {
+  uint32_t Index =
+      __gpu_num_threads_x() * __gpu_block_id_x() + __gpu_thread_id_x();
+
+  if (Index < NumElements) {
+    Out[Index] = hypotf16(X[Index], Y[Index]);
+  }
+}
+#endif
+
+__gpu_kernel void logfKernel(const float *X, float *Out, size_t NumElements) {
+  uint32_t Index =
+      __gpu_num_threads_x() * __gpu_block_id_x() + __gpu_thread_id_x();
+
+  if (Index < NumElements) {
+    Out[Index] = logf(X[Index]);
+  }
+}
diff --git a/offload/unittests/Conformance/device_code/sin.c b/offload/unittests/Conformance/device_code/sin.c
deleted file mode 100644
index e969e60f352a2..0000000000000
--- a/offload/unittests/Conformance/device_code/sin.c
+++ /dev/null
@@ -1,4 +0,0 @@
-#include <gpuintrin.h>
-#include <math.h>
-
-__gpu_kernel void kernel(double *out) { *out = sin(*out); }
diff --git a/offload/unittests/Conformance/include/mathtest/DeviceContext.hpp b/offload/unittests/Conformance/include/mathtest/DeviceContext.hpp
new file mode 100644
index 0000000000000..74ef83ce0c195
--- /dev/null
+++ b/offload/unittests/Conformance/include/mathtest/DeviceContext.hpp
@@ -0,0 +1,121 @@
+#pragma once
+
+#include "mathtest/DeviceResources.hpp"
+#include "mathtest/Dim.hpp"
+#include "mathtest/ErrorHandling.hpp"
+#include "mathtest/Support.hpp"
+
+#include "llvm/ADT/StringRef.h"
+
+#include <cassert>
+#include <cstddef>
+#include <memory>
+#include <string>
+#include <tuple>
+#include <type_traits>
+#include <utility>
+
+namespace mathtest {
+
+std::size_t countDevices();
+
+namespace detail {
+
+void allocManagedMemory(ol_device_handle_t DeviceHandle, std::size_t Size,
+                        void **AllocationOut) noexcept;
+} // namespace detail
+
+class DeviceContext {
+  // For simplicity, the current design of this class doesn't have support for
+  // asynchronous operations and all types of memory allocation.
+  //
+  // Other use cases could benefit from operations like enqueued kernel launch
+  // and enqueued memcpy, as well as device and host memory allocation.
+
+public:
+  // TODO: Add a constructor that also takes a 'Provider'.
+  explicit DeviceContext(std::size_t DeviceId = 0);
+
+  template <typename T>
+  ManagedBuffer<T> createManagedBuffer(std::size_t Size) const noexcept {
+    void *UntypedAddress = nullptr;
+
+    detail::allocManagedMemory(DeviceHandle, Size * sizeof(T), &UntypedAddress);
+    T *TypedAddress = static_cast<T *>(UntypedAddress);
+
+    return ManagedBuffer<T>(TypedAddress, Size);
+  }
+
+  [[nodiscard]] std::shared_ptr<DeviceImage>
+  loadBinary(llvm::StringRef Directory, llvm::StringRef BinaryName,
+             llvm::StringRef Extension) const;
+
+  [[nodiscard]] std::shared_ptr<DeviceImage>
+  loadBinary(llvm::StringRef Directory, llvm::StringRef BinaryName) const;
+
+  template <typename KernelSignature>
+  DeviceKernel<KernelSignature>
+  getKernel(const std::shared_ptr<DeviceImage> &Image,
+            llvm::StringRef KernelName) const noexcept {
+    assert(Image && "Image provided to getKernel is null");
+
+    if (Image->DeviceHandle != this->DeviceHandle) {
+      FATAL_ERROR("Image provided to getKernel was created for a different "
+                  "device");
+    }
+
+    ol_symbol_handle_t KernelHandle = nullptr;
+    getKernelImpl(Image->Handle, KernelName, &KernelHandle);
+
+    return DeviceKernel<KernelSignature>(Image, KernelHandle);
+  }
+
+  template <typename KernelSignature, typename... ArgTypes>
+  void launchKernel(DeviceKernel<KernelSignature> Kernel, Dim NumGroups,
+                    Dim GroupSize, ArgTypes &&...Args) const noexcept {
+    using ExpectedTypes =
+        typename FunctionTypeTraits<KernelSignature>::ArgTypesTuple;
+    using ProvidedTypes = std::tuple<std::decay_t<ArgTypes>...>;
+
+    static_assert(std::is_same_v<ExpectedTypes, ProvidedTypes>,
+                  "Argument types provided to launchKernel do not match the "
+                  "kernel's signature");
+
+    if (Kernel.Image->DeviceHandle != DeviceHandle) {
+      FATAL_ERROR("Kernel provided to launchKernel was created for a different "
+                  "device");
+    }
+
+    if constexpr (sizeof...(Args) == 0) {
+      launchKernelImpl(Kernel.Handle, NumGroups, GroupSize, nullptr, 0);
+    } else {
+      auto KernelArgs = makeKernelArgsPack(std::forward<ArgTypes>(Args)...);
+
+      static_assert(
+          (std::is_trivially_copyable_v<std::decay_t<ArgTypes>> && ...),
+          "Argument types provided to launchKernel must be trivially copyable");
+
+      launchKernelImpl(Kernel.Handle, NumGroups, GroupSize, &KernelArgs,
+                       sizeof(KernelArgs));
+    }
+  }
+
+  [[nodiscard]] std::size_t getId() const noexcept { return DeviceId; }
+
+  [[nodiscard]] std::string getName() const;
+
+  [[nodiscard]] std::string getPlatform() const;
+
+private:
+  void getKernelImpl(ol_program_handle_t ProgramHandle,
+                     llvm::StringRef KernelName,
+                     ol_symbol_handle_t *KernelHandle) const noexcept;
+
+  void launchKernelImpl(ol_symbol_handle_t KernelHandle, const Dim &NumGroups,
+                        const Dim &GroupSize, const void *KernelArgs,
+                        std::size_t KernelArgsSize) const noexcept;
+
+  std::size_t DeviceId;
+  ol_device_handle_t DeviceHandle;
+};
+} // namespace mathtest
diff --git a/offload/unittests/Conformance/include/mathtest/DeviceResources.hpp b/offload/unittests/Conformance/include/mathtest/DeviceResources.hpp
new file mode 100644
index 0000000000000..51f7662ef548e
--- /dev/null
+++ b/offload/unittests/Conformance/include/mathtest/DeviceResources.hpp
@@ -0,0 +1,129 @@
+#pragma once
+
+#include "mathtest/OffloadForward.hpp"
+
+#include "llvm/ADT/ArrayRef.h"
+
+#include <cstddef>
+#include <memory>
+#include <utility>
+
+namespace mathtest {
+
+class DeviceContext;
+
+namespace detail {
+
+void freeDeviceMemory(void *Address) noexcept;
+} // namespace detail
+
+//===----------------------------------------------------------------------===//
+// ManagedBuffer
+//===----------------------------------------------------------------------===//
+
+template <typename T> class [[nodiscard]] ManagedBuffer {
+public:
+  ~ManagedBuffer() noexcept {
+    if (Address) {
+      detail::freeDeviceMemory(Address);
+    }
+  }
+
+  ManagedBuffer(const ManagedBuffer &) = delete;
+  ManagedBuffer &operator=(const ManagedBuffer &) = delete;
+
+  ManagedBuffer(ManagedBuffer &&Other) noexcept
+      : Address(Other.Address), Size(Other.Size) {
+    Other.Address = nullptr;
+    Other.Size = 0;
+  }
+
+  ManagedBuffer &operator=(ManagedBuffer &&Other) noexcept {
+    if (this == &Other)
+      return *this;
+
+    if (Address) {
+      detail::freeDeviceMemory(Address);
+    }
+
+    Address = Other.Address;
+    Size = Other.Size;
+
+    Other.Address = nullptr;
+    Other.Size = 0;
+
+    return *this;
+  }
+
+  [[nodiscard]] T *data() noexcept { return Address; }
+
+  [[nodiscard]] const T *data() const noexcept { return Address; }
+
+  [[nodiscard]] std::size_t getSize() const noexcept { return Size; }
+
+  [[nodiscard]] operator llvm::MutableArrayRef<T>() noexcept {
+    return llvm::MutableArrayRef<T>(data(), getSize());
+  }
+
+  [[nodiscard]] operator llvm::ArrayRef<T>() const noexcept {
+    return llvm::ArrayRef<T>(data(), getSize());
+  }
+
+private:
+  friend class DeviceContext;
+
+  explicit ManagedBuffer(T *Address, std::size_t Size) noexcept
+      : Address(Address), Size(Size) {}
+
+  T *Address = nullptr;
+  std::size_t Size = 0;
+};
+
+//===----------------------------------------------------------------------===//
+// DeviceImage
+//===----------------------------------------------------------------------===//
+
+class [[nodiscard]] DeviceImage {
+public:
+  ~DeviceImage() noexcept;
+  DeviceImage &operator=(DeviceImage &&Other) noexcept;
+
+  DeviceImage(const DeviceImage &) = delete;
+  DeviceImage &operator=(const DeviceImage &) = delete;
+
+  DeviceImage(DeviceImage &&Other) noexcept;
+
+private:
+  friend class DeviceContext;
+
+  explicit DeviceImage(ol_device_handle_t DeviceHandle,
+                       ol_program_handle_t Handle) noexcept;
+
+  ol_device_handle_t DeviceHandle = nullptr;
+  ol_program_handle_t Handle = nullptr;
+};
+
+//===----------------------------------------------------------------------===//
+// DeviceKernel
+//===----------------------------------------------------------------------===//
+
+template <typename KernelSignature> class [[nodiscard]] DeviceKernel {
+public:
+  DeviceKernel() = delete;
+
+  DeviceKernel(const DeviceKernel &) = default;
+  DeviceKernel &operator=(const DeviceKernel &) = default;
+  DeviceKernel(DeviceKernel &&) noexcept = default;
+  DeviceKernel &operator=(DeviceKernel &&) noexcept = default;
+
+private:
+  friend class DeviceContext;
+
+  explicit DeviceKernel(std::shared_ptr<DeviceImage> Image,
+                        ol_symbol_handle_t Kernel)
+      : Image(std::move(Image)), Handle(Kernel) {}
+
+  std::shared_ptr<DeviceImage> Image;
+  ol_symbol_handle_t Handle = nullptr;
+};
+} // namespace mathtest
diff --git a/offload/unittests/Conformance/include/mathtest/Dim.hpp b/offload/unittests/Conformance/include/mathtest/Dim.hpp
new file mode 100644
index 0000000000000..948c10e94cbb3
--- /dev/null
+++ b/offload/unittests/Conformance/include/mathtest/Dim.hpp
@@ -0,0 +1,42 @@
+#pragma once
+
+#include <cassert>
+#include <cstddef>
+#include <cstdint>
+#include <initializer_list>
+
+namespace mathtest {
+
+class Dim {
+public:
+  Dim() = delete;
+
+  constexpr Dim(uint32_t X, uint32_t Y = 1, uint32_t Z = 1) noexcept
+      : Data{X, Y, Z} {
+    assert(X > 0 && Y > 0 && Z > 0 && "Dimensions must be positive");
+  }
+
+  constexpr Dim(std::initializer_list<uint32_t> Dimensions) noexcept
+      : Data{1, 1, 1} {
+    assert(Dimensions.size() <= 3 &&
+           "The number of dimensions must be less than or equal to 3");
+
+    std::size_t Index = 0;
+    for (uint32_t DimValue : Dimensions) {
+      Data[Index++] = DimValue;
+    }
+
+    assert(Data[0] > 0 && Data[1] > 0 && Data[2] > 0 &&
+           "Dimensions must be positive");
+  }
+
+  [[nodiscard]] constexpr uint32_t
+  operator[](std::size_t Index) const noexcept {
+    assert(Index < 3 && "Index is out of range");
+    return Data[Index];
+  }
+
+private:
+  uint32_t Data[3];
+};
+} // namespace mathtest
diff --git a/offload/unittests/Conformance/include/mathtest/ErrorHandling.hpp b/offload/unittests/Conformance/include/mathtest/ErrorHandling.hpp
new file mode 100644
index 0000000000000..932aa79e4c902
--- /dev/null
+++ b/offload/unittests/Conformance/include/mathtest/ErrorHandling.hpp
@@ -0,0 +1,29 @@
+#pragma once
+
+#include "mathtest/OffloadForward.hpp"
+
+#include "llvm/ADT/Twine.h"
+
+#define FATAL_ERROR(Message)                                                   \
+  mathtest::detail::reportFatalError(Message, __FILE__, __LINE__, __func__)
+
+#define OL_CHECK(ResultExpr)                                                   \
+  do {                                                                         \
+    ol_result_t Result = (ResultExpr);                                         \
+    if (Result != OL_SUCCESS) {                                                \
+      mathtest::detail::reportOffloadError(#ResultExpr, Result, __FILE__,      \
+                                           __LINE__, __func__);                \
+    }                                                                          \
+  } while (false)
+
+namespace mathtest {
+namespace detail {
+
+[[noreturn]] void reportFatalError(const llvm::Twine &Message, const char *File,
+                                   int Line, const char *FuncName);
+
+[[noreturn]] void reportOffloadError(const char *ResultExpr, ol_result_t Result,
+                                     const char *File, int Line,
+                                     const char *FuncName);
+} // namespace detail
+} // namespace mathtest
diff --git a/offload/unittests/Conformance/include/mathtest/ExhaustiveGenerator.hpp b/offload/unittests/Conformance/include/mathtest/ExhaustiveGenerator.hpp
new file mode 100644
index 0000000000000..1725a5b35f358
--- /dev/null
+++ b/offload/unittests/Conformance/include/mathtest/ExhaustiveGenerator.hpp
@@ -0,0 +1,139 @@
+#pragma once
+
+#include "mathtest/IndexedRange.hpp"
+#include "mathtest/InputGenerator.hpp"
+
+#include "llvm/ADT/ArrayRef.h"
+#include "llvm/Support/Parallel.h"
+
+#include <algorithm>
+#include <array>
+#include <atomic>
+#include <cassert>
+#include <cstddef>
+#include <cstdint>
+#include <tuple>
+
+namespace mathtest {
+
+template <typename... InTypes>
+class [[nodiscard]] ExhaustiveGenerator final
+    : public InputGenerator<InTypes...> {
+  static constexpr std::size_t NumInputs = sizeof...(InTypes);
+  static_assert(NumInputs > 0, "The number of inputs must be at least 1");
+
+public:
+  explicit constexpr ExhaustiveGenerator(
+      const IndexedRange<InTypes> &...Ranges) noexcept
+      : RangesTuple(Ranges...) {
+    bool Overflowed = getSizeWithOverflow(Ranges..., Size);
+
+    assert(!Overflowed && "The input space size is too large");
+    assert((Size > 0) && "The input space size must be at least 1");
+
+    IndexArrayType DimSizes = {};
+    {
+      std::size_t Index = 0;
+      ((DimSizes[Index++] = Ranges.getSize()), ...);
+    }
+
+    Strides[NumInputs - 1] = 1;
+    if constexpr (NumInputs > 1) {
+      for (int Index = static_cast<int>(NumInputs) - 2; Index >= 0; --Index) {
+        Strides[Index] = Strides[Index + 1] * DimSizes[Index + 1];
+      }
+    }
+  }
+
+  [[nodiscard]] std::size_t
+  fill(llvm::MutableArrayRef<InTypes>... Buffers) noexcept override {
+    const std::array<std::size_t, NumInputs> BufferSizes = {Buffers.size()...};
+    const std::size_t BufferSize = BufferSizes[0];
+    assert((BufferSize != 0) && "Buffer size cannot be zero");
+    assert(std::all_of(BufferSizes.begin(), BufferSizes.end(),
+                       [&](std::size_t Size) { return Size == BufferSize; }) &&
+           "All input buffers must have the same size");
+
+    uint64_t StartFlatIndex, BatchSize;
+    while (true) {
+      uint64_t CurrentFlatIndex =
+          FlatIndexGenerator.load(std::memory_order_relaxed);
+      if (CurrentFlatIndex >= Size)
+        return 0;
+
+      BatchSize = std::min<uint64_t>(BufferSize, Size - CurrentFlatIndex);
+      uint64_t NextFlatIndex = CurrentFlatIndex + BatchSize;
+
+      if (FlatIndexGenerator.compare_exchange_weak(
+              CurrentFlatIndex, NextFlatIndex,
+              std::memory_order_acq_rel, // Success
+              std::memory_order_acquire  // Failure
+              )) {
+        StartFlatIndex = CurrentFlatIndex;
+        break;
+      }
+    }
+
+    auto BufferPtrsTuple = std::make_tuple(Buffers.data()...);
+
+    llvm::parallelFor(0, BatchSize, [&](std::size_t Offset) {
+      writeInputs(StartFlatIndex, Offset, BufferPtrsTuple);
+    });
+
+    return static_cast<std::size_t>(BatchSize);
+  }
+
+private:
+  using RangesTupleType = std::tuple<IndexedRange<InTypes>...>;
+  using IndexArrayType = std::array<uint64_t, NumInputs>;
+
+  static bool getSizeWithOverflow(const IndexedRange<InTypes> &...Ranges,
+                                  uint64_t &Size) noexcept {
+    Size = 1;
+    bool Overflowed = false;
+
+    auto Multiplier = [&](const uint64_t RangeSize) {
+      if (!Overflowed) {
+        Overflowed = __builtin_mul_overflow(Size, RangeSize, &Size);
+      }
+    };
+
+    (Multiplier(Ranges.getSize()), ...);
+
+    return Overflowed;
+  }
+
+  template <typename BufferPtrsTupleType>
+  void writeInputs(uint64_t StartFlatIndex, uint64_t Offset,
+                   BufferPtrsTupleType BufferPtrsTuple) const noexcept {
+    auto NDIndex = getNDIndex(StartFlatIndex + Offset);
+    writeInputsImpl<0>(NDIndex, Offset, BufferPtrsTuple);
+  }
+
+  constexpr IndexArrayType getNDIndex(uint64_t FlatIndex) const noexcept {
+    IndexArrayType NDIndex;
+
+    for (std::size_t Index = 0; Index < NumInputs; ++Index) {
+      NDIndex[Index] = FlatIndex / Strides[Index];
+      FlatIndex -= NDIndex[Index] * Strides[Index];
+    }
+
+    return NDIndex;
+  }
+
+  template <std::size_t Index, typename BufferPtrsTupleType>
+  void writeInputsImpl(IndexArrayType NDIndex, uint64_t Offset,
+                       BufferPtrsTupleType BufferPtrsTuple) const noexcept {
+    if constexpr (Index < NumInputs) {
+      const auto &Range = std::get<Index>(RangesTuple);
+      std::get<Index>(BufferPtrsTuple)[Offset] = Range[NDIndex[Index]];
+      writeInputsImpl<Index + 1>(NDIndex, Offset, BufferPtrsTuple);
+    }
+  }
+
+  uint64_t Size = 1;
+  RangesTupleType RangesTuple;
+  IndexArrayType Strides = {};
+  std::atomic<uint64_t> FlatIndexGenerator = 0;
+};
+} // namespace mathtest
diff --git a/offload/unittests/Conformance/include/mathtest/GpuMathTest.hpp b/offload/unittests/Conformance/include/mathtest/GpuMathTest.hpp
new file mode 100644
index 0000000000000..e5d1c6c77f634
--- /dev/null
+++ b/offload/unittests/Conformance/include/mathtest/GpuMathTest.hpp
@@ -0,0 +1,159 @@
+#pragma once
+
+#include "mathtest/DeviceContext.hpp"
+#include "mathtest/DeviceResources.hpp"
+#include "mathtest/HostRefChecker.hpp"
+#include "mathtest/InputGenerator.hpp"
+#include "mathtest/Support.hpp"
+#include "mathtest/TestResult.hpp"
+
+#include "llvm/ADT/ArrayRef.h"
+#include "llvm/ADT/StringRef.h"
+#include "llvm/ADT/StringSwitch.h"
+
+#include <cassert>
+#include <cstddef>
+#include <cstdint>
+#include <memory>
+#include <tuple>
+#include <utility>
+
+namespace mathtest {
+
+template <auto Func, typename Checker = HostRefChecker<Func>>
+class [[nodiscard]] GpuMathTest final {
+  using FunctionTraits = FunctionTraits<Func>;
+  using OutType = typename FunctionTraits::ReturnType;
+  using InTypesTuple = typename FunctionTraits::ArgTypesTuple;
+
+  template <typename... Ts>
+  using PartialResultType = TestResult<OutType, Ts...>;
+  using KernelSignature = KernelSignatureOf_t<Func>;
+
+  template <typename... Ts>
+  using TypeIdentitiesTuple = std::tuple<TypeIdentityOf<Ts>...>;
+  using InTypeIdentitiesTuple =
+      ApplyTupleTypes_t<InTypesTuple, TypeIdentitiesTuple>;
+
+  static constexpr std::size_t DefaultBufferSize =
+      DefaultBufferSizeFor_v<OutType, InTypesTuple>;
+  static constexpr uint32_t DefaultGroupSize = 512;
+
+public:
+  using FunctionConfig = FunctionConfig<Func>;
+  using ResultType = ApplyTupleTypes_t<InTypesTuple, PartialResultType>;
+  using GeneratorType = ApplyTupleTypes_t<InTypesTuple, InputGenerator>;
+
+  explicit GpuMathTest(std::shared_ptr<DeviceContext> Context,
+                       llvm::StringRef Provider,
+                       llvm::StringRef DeviceBinsDirectory)
+      : Context(std::move(Context)),
+        Kernel(getKernel(this->Context, Provider, DeviceBinsDirectory)) {
+    assert(this->Context && "Context must not be null");
+  }
+
+  ResultType run(GeneratorType &Generator,
+                 std::size_t BufferSize = DefaultBufferSize,
+                 uint32_t GroupSize = DefaultGroupSize) const noexcept {
+    assert(BufferSize > 0 && "Buffer size must be a positive value");
+    assert(GroupSize > 0 && "Group size must be a positive value");
+
+    auto [InBuffersTuple, OutBuffer] = createBuffers(BufferSize);
+    ResultType FinalResult;
+
+    while (true) {
+      const std::size_t BatchSize = std::apply(
+          [&](auto &...Buffers) { return Generator.fill(Buffers...); },
+          InBuffersTuple);
+
+      if (BatchSize == 0)
+        break;
+
+      const auto BatchResult =
+          processBatch(InBuffersTuple, OutBuffer, BatchSize, GroupSize);
+
+      FinalResult.accumulate(BatchResult);
+    }
+
+    return FinalResult;
+  }
+
+  [[nodiscard]] const DeviceContext &getContext() const noexcept {
+    assert(Context && "Context must not be null");
+    return *Context;
+  }
+
+private:
+  static DeviceKernel<KernelSignature>
+  getKernel(const std::shared_ptr<DeviceContext> &Context,
+            llvm::StringRef Provider,
+            llvm::StringRef DeviceBinsDirectory) noexcept {
+    llvm::StringRef BinaryName = llvm::StringSwitch<llvm::StringRef>(Provider)
+                                     .Case("llvm-libm", "LLVMLibm")
+                                     .Default("");
+
+    if (BinaryName.empty()) {
+      FATAL_ERROR(llvm::Twine("Unsupported provider: '") + Provider + "'");
+    }
+
+    const auto Image = Context->loadBinary(DeviceBinsDirectory, BinaryName);
+
+    return Context->getKernel<KernelSignature>(Image,
+                                               FunctionConfig::KernelName);
+  }
+
+  [[nodiscard]] auto createBuffers(std::size_t BufferSize) const {
+    auto InBuffersTuple = std::apply(
+        [&](auto... InTypeIdentities) {
+          return std::make_tuple(
+              Context->createManagedBuffer<
+                  typename decltype(InTypeIdentities)::type>(BufferSize)...);
+        },
+        InTypeIdentitiesTuple{});
+    auto OutBuffer = Context->createManagedBuffer<OutType>(BufferSize);
+
+    return std::make_pair(std::move(InBuffersTuple), std::move(OutBuffer));
+  }
+
+  template <typename InBuffersTupleType>
+  [[nodiscard]] ResultType
+  processBatch(const InBuffersTupleType &InBuffersTuple,
+               ManagedBuffer<OutType> &OutBuffer, std::size_t BatchSize,
+               uint32_t GroupSize) const noexcept {
+    const uint32_t NumGroups = (BatchSize + GroupSize - 1) / GroupSize;
+    const auto KernelArgsTuple = std::apply(
+        [&](const auto &...InBuffers) {
+          return std::make_tuple(InBuffers.data()..., OutBuffer.data(),
+                                 BatchSize);
+        },
+        InBuffersTuple);
+
+    std::apply(
+        [&](const auto &...KernelArgs) {
+          Context->launchKernel(Kernel, NumGroups, GroupSize, KernelArgs...);
+        },
+        KernelArgsTuple);
+
+    return check(InBuffersTuple, OutBuffer, BatchSize);
+  }
+
+  template <typename InBuffersTupleType>
+  [[nodiscard]] static ResultType
+  check(const InBuffersTupleType &InBuffersTuple,
+        const ManagedBuffer<OutType> &OutBuffer,
+        std::size_t BatchSize) noexcept {
+    const auto InViewsTuple = std::apply(
+        [&](auto &...InBuffers) {
+          return std::make_tuple(
+              llvm::ArrayRef(InBuffers.data(), BatchSize)...);
+        },
+        InBuffersTuple);
+    const auto OutView = llvm::ArrayRef<OutType>(OutBuffer.data(), BatchSize);
+
+    return Checker::check(InViewsTuple, OutView);
+  }
+
+  std::shared_ptr<DeviceContext> Context;
+  DeviceKernel<KernelSignature> Kernel;
+};
+} // namespace mathtest
diff --git a/offload/unittests/Conformance/include/mathtest/HostRefChecker.hpp b/offload/unittests/Conformance/include/mathtest/HostRefChecker.hpp
new file mode 100644
index 0000000000000..c45137d652df2
--- /dev/null
+++ b/offload/unittests/Conformance/include/mathtest/HostRefChecker.hpp
@@ -0,0 +1,82 @@
+#pragma once
+
+#include "mathtest/Numerics.hpp"
+#include "mathtest/Support.hpp"
+#include "mathtest/TestResult.hpp"
+
+#include "llvm/ADT/ArrayRef.h"
+#include "llvm/ADT/Sequence.h"
+#include "llvm/Support/Parallel.h"
+
+#include <cstddef>
+#include <tuple>
+#include <utility>
+
+namespace mathtest {
+
+template <auto Func> class HostRefChecker {
+  using FuncTraits = FunctionTraits<Func>;
+  using InTypesTuple = typename FuncTraits::ArgTypesTuple;
+
+  using FuncConfig = FunctionConfig<Func>;
+
+  template <typename... Ts>
+  using BuffersTupleType = std::tuple<llvm::ArrayRef<Ts>...>;
+
+public:
+  using OutType = typename FuncTraits::ReturnType;
+
+private:
+  template <typename... Ts>
+  using PartialResultType = TestResult<OutType, Ts...>;
+
+public:
+  using ResultType = ApplyTupleTypes_t<InTypesTuple, PartialResultType>;
+  using InBuffersTupleType = ApplyTupleTypes_t<InTypesTuple, BuffersTupleType>;
+
+  HostRefChecker() = delete;
+
+  static ResultType check(InBuffersTupleType InBuffersTuple,
+                          llvm::ArrayRef<OutType> OutBuffer) noexcept {
+    const std::size_t BufferSize = OutBuffer.size();
+    std::apply(
+        [&](const auto &...InBuffers) {
+          assert(
+              ((InBuffers.size() == BufferSize) && ...) &&
+              "All input buffers must have the same size as the output buffer");
+        },
+        InBuffersTuple);
+
+    assert((BufferSize != 0) && "Buffer size cannot be zero");
+
+    ResultType Init;
+
+    auto Transform = [&](std::size_t Index) {
+      auto CurrentInputsTuple = std::apply(
+          [&](const auto &...InBuffers) {
+            return std::make_tuple(InBuffers[Index]...);
+          },
+          InBuffersTuple);
+
+      const OutType Actual = OutBuffer[Index];
+      const OutType Expected = std::apply(Func, CurrentInputsTuple);
+
+      const auto UlpDistance = computeUlpDistance(Actual, Expected);
+      const bool IsFailure = UlpDistance > FuncConfig::UlpTolerance;
+
+      return ResultType(UlpDistance, IsFailure,
+                        typename ResultType::TestCase(
+                            std::move(CurrentInputsTuple), Actual, Expected));
+    };
+
+    auto Reduce = [](ResultType A, const ResultType &B) {
+      A.accumulate(B);
+      return A;
+    };
+
+    const auto Indexes = llvm::seq(BufferSize);
+    return llvm::parallelTransformReduce(Indexes.begin(), Indexes.end(), Init,
+                                         Reduce, Transform);
+  }
+};
+} // namespace mathtest
diff --git a/offload/unittests/Conformance/include/mathtest/IndexedRange.hpp b/offload/unittests/Conformance/include/mathtest/IndexedRange.hpp
new file mode 100644
index 0000000000000..0f33978c8d30d
--- /dev/null
+++ b/offload/unittests/Conformance/include/mathtest/IndexedRange.hpp
@@ -0,0 +1,91 @@
+#pragma once
+
+#include "mathtest/Numerics.hpp"
+
+#include <cassert>
+#include <cstdint>
+#include <limits>
+#include <type_traits>
+
+namespace mathtest {
+
+template <typename T> class [[nodiscard]] IndexedRange {
+  static_assert(IsFloatingPoint_v<T> || std::is_integral_v<T>,
+                "Type T must be an integral or floating-point type");
+  static_assert(sizeof(T) <= sizeof(uint64_t),
+                "Type T must be no wider than uint64_t");
+
+public:
+  constexpr IndexedRange() noexcept
+      : IndexedRange(getMinOrNegInf<T>(), getMaxOrInf<T>(), true) {}
+
+  explicit constexpr IndexedRange(T Begin, T End, bool Inclusive) noexcept
+      : MappedBegin(mapToOrderedUnsigned(Begin)),
+        MappedEnd(mapToOrderedUnsigned(End)) {
+    if (Inclusive) {
+      assert((Begin <= End) && "Begin must be less than or equal to End");
+    } else {
+      assert((Begin < End) && "Begin must be less than End");
+      --MappedEnd;
+    }
+
+    assert(((MappedEnd - MappedBegin) < std::numeric_limits<uint64_t>::max()) &&
+           "The range is too large to index");
+  }
+
+  [[nodiscard]] constexpr uint64_t getSize() const noexcept {
+    return static_cast<uint64_t>(MappedEnd) - MappedBegin + 1;
+  }
+
+  [[nodiscard]] constexpr T operator[](uint64_t Index) const noexcept {
+    assert((Index < getSize()) && "Index is out of range");
+
+    StorageType MappedValue = MappedBegin + Index;
+    return mapFromOrderedUnsigned(MappedValue);
+  }
+
+private:
+  using StorageType = StorageTypeOf_t<T>;
+
+  // Linearise T values into an ordered unsigned space:
+  //  * The mapping is monotonic: a >= b if, and only if, map(a) >= map(b)
+  //  * The difference |map(a) − map(b)| equals the number of representable
+  //    values between a and b within the same type
+  static constexpr StorageType mapToOrderedUnsigned(T Value) {
+    if constexpr (IsFloatingPoint_v<T>) {
+      StorageType SignMask = FPUtils<T>::SignMask;
+      StorageType Bits = FPUtils<T>::getAsBits(Value);
+      return (Bits & SignMask) ? SignMask - (Bits - SignMask) - 1
+                               : SignMask + Bits;
+    }
+
+    if constexpr (std::is_signed_v<T>) {
+      StorageType SignMask = maskLeadingOnes<StorageType, 1>();
+      return __builtin_bit_cast(StorageType, Value) ^ SignMask;
+    }
+
+    return Value;
+  }
+
+  static constexpr T mapFromOrderedUnsigned(StorageType MappedValue) {
+    if constexpr (IsFloatingPoint_v<T>) {
+      StorageType SignMask = FPUtils<T>::SignMask;
+      StorageType Bits = (MappedValue < SignMask)
+                             ? (SignMask - MappedValue) + SignMask - 1
+                             : MappedValue - SignMask;
+
+      return FPUtils<T>::createFromBits(Bits);
+    }
+
+    if constexpr (std::is_signed_v<T>) {
+      StorageType SignMask = maskLeadingOnes<StorageType, 1>();
+      return __builtin_bit_cast(T, MappedValue ^ SignMask);
+    }
+
+    return MappedValue;
+  }
+
+  StorageType MappedBegin;
+  StorageType MappedEnd;
+};
+} // namespace mathtest
diff --git a/offload/unittests/Conformance/include/mathtest/InputGenerator.hpp b/offload/unittests/Conformance/include/mathtest/InputGenerator.hpp
new file mode 100644
index 0000000000000..d9365d4b14423
--- /dev/null
+++ b/offload/unittests/Conformance/include/mathtest/InputGenerator.hpp
@@ -0,0 +1,14 @@
+#pragma once
+
+#include "llvm/ADT/ArrayRef.h"
+
+namespace mathtest {
+
+template <typename... InTypes> class InputGenerator {
+public:
+  virtual ~InputGenerator() noexcept = default;
+
+  [[nodiscard]] virtual size_t
+  fill(llvm::MutableArrayRef<InTypes>... Buffers) noexcept = 0;
+};
+} // namespace mathtest
diff --git a/offload/unittests/Conformance/include/mathtest/Numerics.hpp b/offload/unittests/Conformance/include/mathtest/Numerics.hpp
new file mode 100644
index 0000000000000..f43a26974dda7
--- /dev/null
+++ b/offload/unittests/Conformance/include/mathtest/Numerics.hpp
@@ -0,0 +1,235 @@
+#pragma once
+
+#include "mathtest/Support.hpp"
+#include "mathtest/TypeExtras.hpp"
+
+#include <climits>
+#include <cstddef>
+#include <cstdint>
+#include <limits>
+#include <math.h>
+#include <type_traits>
+
+namespace mathtest {
+
+//===----------------------------------------------------------------------===//
+// Type Traits
+//===----------------------------------------------------------------------===//
+
+template <typename T> struct StorageTypeOf {
+private:
+  static constexpr auto getStorageType() noexcept {
+    if constexpr (std::is_unsigned_v<T>) {
+      return TypeIdentityOf<T>{};
+    } else if constexpr (std::is_signed_v<T>) {
+      return TypeIdentityOf<std::make_unsigned_t<T>>{};
+    } else {
+      static_assert(!std::is_same_v<T, T>, "Unsupported type");
+    }
+  }
+
+public:
+  using type = typename decltype(getStorageType())::type;
+};
+
+#ifdef MATHTEST_HAS_FLOAT16
+template <> struct StorageTypeOf<float16> {
+  using type = uint16_t;
+};
+#endif
+
+template <> struct StorageTypeOf<float> {
+  using type = uint32_t;
+};
+
+template <> struct StorageTypeOf<double> {
+  using type = uint64_t;
+};
+
+template <typename T> using StorageTypeOf_t = typename StorageTypeOf<T>::type;
+
+template <typename T> struct IsFloatingPoint : std::is_floating_point<T> {};
+
+#ifdef MATHTEST_HAS_FLOAT16
+template <> struct IsFloatingPoint<float16> : std::true_type {};
+#endif
+
+template <typename T>
+inline constexpr bool IsFloatingPoint_v // NOLINT(readability-identifier-naming)
+    = IsFloatingPoint<T>::value;
+
+//===----------------------------------------------------------------------===//
+// Bitmask Utilities
+//===----------------------------------------------------------------------===//
+
+template <typename UIntType, std::size_t Count>
+[[nodiscard]] constexpr UIntType maskLeadingOnes() noexcept {
+  static_assert(std::is_unsigned_v<UIntType>,
+                "UIntType must be an unsigned integer type");
+
+  constexpr unsigned TotalBits = CHAR_BIT * sizeof(UIntType);
+  static_assert(
+      Count <= TotalBits,
+      "Count must be less than or equal to the bit width of UIntType");
+
+  return Count == 0 ? UIntType(0) : (~UIntType(0) << (TotalBits - Count));
+  ;
+}
+
+template <typename UIntType, std::size_t Count>
+[[nodiscard]] constexpr UIntType maskTrailingOnes() noexcept {
+  static_assert(std::is_unsigned_v<UIntType>,
+                "UIntType must be an unsigned integer type");
+
+  constexpr unsigned TotalBits = CHAR_BIT * sizeof(UIntType);
+  static_assert(
+      Count <= TotalBits,
+      "Count must be less than or equal to the bit width of UIntType");
+
+  return Count == 0 ? UIntType(0) : (~UIntType(0) >> (TotalBits - Count));
+}
+
+//===----------------------------------------------------------------------===//
+// Floating-Point Utilities
+//===----------------------------------------------------------------------===//
+
+template <typename FloatType> struct FPLayout;
+
+#ifdef MATHTEST_HAS_FLOAT16
+template <> struct FPLayout<float16> {
+  static constexpr std::size_t SignLen = 1;
+  static constexpr std::size_t ExponentLen = 5;
+  static constexpr std::size_t FractionLen = 10;
+};
+#endif
+
+template <> struct FPLayout<float> {
+  static constexpr std::size_t SignLen = 1;
+  static constexpr std::size_t ExponentLen = 8;
+  static constexpr std::size_t FractionLen = 23;
+};
+
+template <> struct FPLayout<double> {
+  static constexpr std::size_t SignLen = 1;
+  static constexpr std::size_t ExponentLen = 11;
+  static constexpr std::size_t FractionLen = 52;
+};
+
+template <typename FloatType> struct FPUtils : public FPLayout<FloatType> {
+  using FPLayout = FPLayout<FloatType>;
+  using StorageType = StorageTypeOf_t<FloatType>;
+  using FPLayout::ExponentLen;
+  using FPLayout::FractionLen;
+  using FPLayout::SignLen;
+
+  static constexpr StorageType SignMask =
+      maskTrailingOnes<StorageType, SignLen>() << (ExponentLen + FractionLen);
+
+  FPUtils() = delete;
+
+  [[nodiscard]] static constexpr FloatType
+  createFromBits(StorageType Bits) noexcept {
+    return __builtin_bit_cast(FloatType, Bits);
+  }
+
+  [[nodiscard]] static constexpr StorageType
+  getAsBits(FloatType Value) noexcept {
+    return __builtin_bit_cast(StorageType, Value);
+  }
+
+  [[nodiscard]] static constexpr bool isNaN(FloatType Value) noexcept {
+    return __builtin_isnan(Value);
+  }
+
+  [[nodiscard]] static constexpr bool getSignBit(FloatType Value) noexcept {
+    return getAsBits(Value) & SignMask;
+  }
+};
+
+//===----------------------------------------------------------------------===//
+// Numeric Functions
+//===----------------------------------------------------------------------===//
+
+template <typename T> [[nodiscard]] constexpr T getMinOrNegInf() noexcept {
+  static_assert(std::is_arithmetic_v<T>, "Type T must be an arithmetic type");
+
+  if constexpr (std::is_floating_point_v<T> &&
+                std::numeric_limits<T>::has_infinity) {
+    return -std::numeric_limits<T>::infinity();
+  }
+
+  return std::numeric_limits<T>::lowest();
+}
+
+#ifdef MATHTEST_HAS_FLOAT16
+template <> [[nodiscard]] constexpr float16 getMinOrNegInf<float16>() noexcept {
+  using StorageType = StorageTypeOf_t<float16>;
+
+  return __builtin_bit_cast(float16, static_cast<StorageType>(0xFC00U));
+}
+#endif
+
+template <typename T> [[nodiscard]] constexpr T getMaxOrInf() noexcept {
+  static_assert(std::is_arithmetic_v<T>, "Type T must be an arithmetic type");
+
+  if constexpr (std::is_floating_point_v<T> &&
+                std::numeric_limits<T>::has_infinity) {
+    return std::numeric_limits<T>::infinity();
+  }
+
+  return std::numeric_limits<T>::max();
+}
+
+#ifdef MATHTEST_HAS_FLOAT16
+template <> [[nodiscard]] constexpr float16 getMaxOrInf<float16>() noexcept {
+  using StorageType = StorageTypeOf_t<float16>;
+
+  return __builtin_bit_cast(float16, static_cast<StorageType>(0x7C00U));
+}
+#endif
+
+template <typename FloatType>
+[[nodiscard]] uint64_t computeUlpDistance(FloatType X, FloatType Y) noexcept {
+  static_assert(IsFloatingPoint_v<FloatType>,
+                "FloatType must be a floating-point type");
+  using FPUtils = FPUtils<FloatType>;
+  using StorageType = typename FPUtils::StorageType;
+
+  if (X == Y) {
+    if (FPUtils::getSignBit(X) != FPUtils::getSignBit(Y)) [[unlikely]] {
+      // When X == Y, different sign bits imply that X and Y are +0.0 and -0.0
+      // (in any order). Since we want to treat them as unequal in the context
+      // of accuracy testing of mathematical functions, we return the smallest
+      // non-zero value
+      return 1;
+    }
+    return 0;
+  }
+
+  const bool XIsNaN = FPUtils::isNaN(X);
+  const bool YIsNaN = FPUtils::isNaN(Y);
+
+  if (XIsNaN && YIsNaN) {
+    return 0;
+  }
+  if (XIsNaN || YIsNaN) {
+    return std::numeric_limits<uint64_t>::max();
+  }
+
+  constexpr StorageType SignMask = FPUtils::SignMask;
+
+  // Linearise FloatType values into an ordered unsigned space:
+  //  * The mapping is monotonic: a >= b if, and only if, map(a) >= map(b)
+  //  * The difference |map(a) − map(b)| equals the number of std::nextafter
+  //    steps between a and b within the same type
+  auto MapToOrderedUnsigned = [](FloatType Value) {
+    const StorageType Bits = FPUtils::getAsBits(Value);
+    return (Bits & SignMask) ? SignMask - (Bits - SignMask) : SignMask + Bits;
+  };
+
+  const StorageType MappedX = MapToOrderedUnsigned(X);
+  const StorageType MappedY = MapToOrderedUnsigned(Y);
+  return static_cast<uint64_t>(MappedX > MappedY ? MappedX - MappedY
+                                                 : MappedY - MappedX);
+}
+} // namespace mathtest
diff --git a/offload/unittests/Conformance/include/mathtest/OffloadForward.hpp b/offload/unittests/Conformance/include/mathtest/OffloadForward.hpp
new file mode 100644
index 0000000000000..099b86af2929d
--- /dev/null
+++ b/offload/unittests/Conformance/include/mathtest/OffloadForward.hpp
@@ -0,0 +1,22 @@
+#pragma once
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct ol_error_struct_t;
+typedef const ol_error_struct_t *ol_result_t;
+#define OL_SUCCESS (static_cast<ol_result_t>(nullptr))
+
+struct ol_device_impl_t;
+typedef struct ol_device_impl_t *ol_device_handle_t;
+
+struct ol_program_impl_t;
+typedef struct ol_program_impl_t *ol_program_handle_t;
+
+struct ol_symbol_impl_t;
+typedef struct ol_symbol_impl_t *ol_symbol_handle_t;
+
+#ifdef __cplusplus
+}
+#endif
diff --git a/offload/unittests/Conformance/include/mathtest/Support.hpp b/offload/unittests/Conformance/include/mathtest/Support.hpp
new file mode 100644
index 0000000000000..e2b41f9dec9ab
--- /dev/null
+++ b/offload/unittests/Conformance/include/mathtest/Support.hpp
@@ -0,0 +1,138 @@
+#pragma once
+
+#include <cstddef>
+#include <tuple>
+#include <type_traits>
+#include <utility>
+
+namespace mathtest {
+
+//===----------------------------------------------------------------------===//
+// Function & Type Traits
+//===----------------------------------------------------------------------===//
+
+namespace detail {
+
+template <typename T> struct FunctionTraitsImpl;
+
+template <typename RetType, typename... ArgTypes>
+struct FunctionTraitsImpl<RetType(ArgTypes...)> {
+  using ReturnType = RetType;
+  using ArgTypesTuple = std::tuple<ArgTypes...>;
+};
+
+template <typename RetType, typename... ArgTypes>
+struct FunctionTraitsImpl<RetType(ArgTypes...) noexcept>
+    : FunctionTraitsImpl<RetType(ArgTypes...)> {};
+
+template <typename FuncType>
+struct FunctionTraitsImpl<FuncType *> : FunctionTraitsImpl<FuncType> {};
+} // namespace detail
+
+template <auto Func>
+using FunctionTraits = detail::FunctionTraitsImpl<
+    std::remove_pointer_t<std::decay_t<decltype(Func)>>>;
+
+template <typename FuncType>
+using FunctionTypeTraits = detail::FunctionTraitsImpl<FuncType>;
+
+template <typename T> struct TypeIdentityOf {
+  using type = T;
+};
+
+template <typename TupleTypes, template <typename...> class Template>
+struct ApplyTupleTypes;
+
+template <template <typename...> class Template, typename... Ts>
+struct ApplyTupleTypes<std::tuple<Ts...>, Template> {
+  using type = Template<Ts...>;
+};
+
+template <typename TupleTypes, template <typename...> class Template>
+using ApplyTupleTypes_t = typename ApplyTupleTypes<TupleTypes, Template>::type;
+
+namespace detail {
+
+template <typename T> struct KernelSignatureOfImpl;
+
+template <typename RetType, typename... ArgTypes>
+struct KernelSignatureOfImpl<RetType(ArgTypes...)> {
+  using type = void(const std::decay_t<ArgTypes> *..., RetType *, std::size_t);
+};
+
+template <typename RetType, typename... ArgTypes>
+struct KernelSignatureOfImpl<RetType(ArgTypes...) noexcept>
+    : KernelSignatureOfImpl<RetType(ArgTypes...)> {};
+} // namespace detail
+
+template <auto Func>
+using KernelSignatureOf = detail::KernelSignatureOfImpl<
+    std::remove_pointer_t<std::decay_t<decltype(Func)>>>;
+
+template <auto Func>
+using KernelSignatureOf_t = typename KernelSignatureOf<Func>::type;
+
+//===----------------------------------------------------------------------===//
+// Kernel Argument Packing
+//===----------------------------------------------------------------------===//
+
+template <typename... ArgTypes> struct KernelArgsPack;
+
+template <typename ArgType> struct KernelArgsPack<ArgType> {
+  std::decay_t<ArgType> Arg;
+
+  constexpr KernelArgsPack(ArgType &&Arg) : Arg(std::forward<ArgType>(Arg)) {}
+};
+
+template <typename ArgType0, typename ArgType1, typename... ArgTypes>
+struct KernelArgsPack<ArgType0, ArgType1, ArgTypes...> {
+  std::decay_t<ArgType0> Arg0;
+  KernelArgsPack<ArgType1, ArgTypes...> Args;
+
+  constexpr KernelArgsPack(ArgType0 &&Arg0, ArgType1 &&Arg1, ArgTypes &&...Args)
+      : Arg0(std::forward<ArgType0>(Arg0)),
+        Args(std::forward<ArgType1>(Arg1), std::forward<ArgTypes>(Args)...) {}
+};
+
+template <typename... ArgTypes>
+KernelArgsPack<ArgTypes...> makeKernelArgsPack(ArgTypes &&...Args) {
+  return KernelArgsPack<ArgTypes...>(std::forward<ArgTypes>(Args)...);
+}
+
+//===----------------------------------------------------------------------===//
+// Configuration Helpers
+//===----------------------------------------------------------------------===//
+
+template <auto Func> struct FunctionConfig;
+
+namespace detail {
+
+template <typename... BufferTypes>
+static constexpr std::size_t getDefaultBufferSize() {
+  static_assert(sizeof...(BufferTypes) > 0,
+                "At least one buffer type must be provided");
+
+  constexpr std::size_t TotalMemoryInBytes = 512ULL << 20; // 512 MiB
+  constexpr std::size_t ElementTupleSize = (sizeof(BufferTypes) + ...);
+
+  static_assert(ElementTupleSize > 0,
+                "Cannot calculate buffer size for empty types");
+
+  return TotalMemoryInBytes / ElementTupleSize;
+}
+} // namespace detail
+
+template <typename BufferType, typename BufferTupleTypes>
+struct DefaultBufferSizeFor;
+
+template <typename BufferType, typename... BufferTypes>
+struct DefaultBufferSizeFor<BufferType, std::tuple<BufferTypes...>> {
+  static constexpr std::size_t value // NOLINT(readability-identifier-naming)
+      = detail::getDefaultBufferSize<BufferType, BufferTypes...>();
+};
+
+template <typename OutType, typename InTypesTuple>
+inline constexpr std::size_t
+    DefaultBufferSizeFor_v // NOLINT(readability-identifier-naming)
+    = DefaultBufferSizeFor<OutType, InTypesTuple>::value;
+} // namespace mathtest
diff --git a/offload/unittests/Conformance/include/mathtest/TestResult.hpp b/offload/unittests/Conformance/include/mathtest/TestResult.hpp
new file mode 100644
index 0000000000000..cdb4f2fa09fa1
--- /dev/null
+++ b/offload/unittests/Conformance/include/mathtest/TestResult.hpp
@@ -0,0 +1,70 @@
+#pragma once
+
+#include <cstdint>
+#include <optional>
+#include <tuple>
+#include <utility>
+
+namespace mathtest {
+
+template <typename OutType, typename... InTypes>
+class [[nodiscard]] TestResult {
+public:
+  struct [[nodiscard]] TestCase {
+    std::tuple<InTypes...> Inputs;
+    OutType Actual;
+    OutType Expected;
+
+    explicit constexpr TestCase(std::tuple<InTypes...> &&Inputs, OutType Actual,
+                                OutType Expected) noexcept
+        : Inputs(std::move(Inputs)), Actual(std::move(Actual)),
+          Expected(std::move(Expected)) {}
+  };
+
+  TestResult() = default;
+
+  explicit TestResult(uint64_t UlpDistance, bool IsFailure,
+                      TestCase &&Case) noexcept
+      : MaxUlpDistance(UlpDistance), FailureCount(IsFailure ? 1 : 0),
+        TestCaseCount(1) {
+    if (IsFailure) {
+      WorstFailingCase.emplace(std::move(Case));
+    }
+  }
+
+  void accumulate(const TestResult &Other) noexcept {
+    if (Other.MaxUlpDistance > MaxUlpDistance) {
+      MaxUlpDistance = Other.MaxUlpDistance;
+      WorstFailingCase = Other.WorstFailingCase;
+    }
+
+    FailureCount += Other.FailureCount;
+    TestCaseCount += Other.TestCaseCount;
+  }
+
+  [[nodiscard]] bool hasPassed() const noexcept { return FailureCount == 0; }
+
+  [[nodiscard]] uint64_t getMaxUlpDistance() const noexcept {
+    return MaxUlpDistance;
+  }
+
+  [[nodiscard]] uint64_t getFailureCount() const noexcept {
+    return FailureCount;
+  }
+
+  [[nodiscard]] uint64_t getTestCaseCount() const noexcept {
+    return TestCaseCount;
+  }
+
+  [[nodiscard]] const std::optional<TestCase> &
+  getWorstFailingCase() const noexcept {
+    return WorstFailingCase;
+  }
+
+private:
+  uint64_t MaxUlpDistance = 0;
+  uint64_t FailureCount = 0;
+  uint64_t TestCaseCount = 0;
+  std::optional<TestCase> WorstFailingCase;
+};
+} // namespace mathtest
diff --git a/offload/unittests/Conformance/include/mathtest/TestRunner.hpp b/offload/unittests/Conformance/include/mathtest/TestRunner.hpp
new file mode 100644
index 0000000000000..764642647e84b
--- /dev/null
+++ b/offload/unittests/Conformance/include/mathtest/TestRunner.hpp
@@ -0,0 +1,118 @@
+#pragma once
+
+#include "mathtest/Numerics.hpp"
+
+#include "llvm/ADT/Twine.h"
+#include "llvm/Support/FormatVariadic.h"
+#include "llvm/Support/raw_ostream.h"
+
+#include <chrono>
+#include <tuple>
+
+namespace mathtest {
+namespace detail {
+
+template <typename T>
+void printValue(llvm::raw_ostream &OS, const T &Value) noexcept {
+  if constexpr (IsFloatingPoint_v<T>) {
+    using FPUtils = FPUtils<T>;
+
+    if constexpr (sizeof(T) < sizeof(float)) {
+      OS << float(Value);
+    } else {
+      OS << Value;
+    }
+
+    OS << llvm::formatv(" (0x{0})",
+                        llvm::Twine::utohexstr(FPUtils::getAsBits(Value)));
+  } else {
+    OS << Value;
+  }
+}
+
+template <typename... Ts>
+void printValues(llvm::raw_ostream &OS,
+                 const std::tuple<Ts...> &ValuesTuple) noexcept {
+  std::apply(
+      [&OS](const auto &...Values) {
+        bool IsFirst = true;
+        auto Print = [&](const auto &Value) {
+          if (!IsFirst) {
+            OS << ", ";
+          }
+          printValue(OS, Value);
+          IsFirst = false;
+        };
+        (Print(Values), ...);
+      },
+      ValuesTuple);
+}
+
+template <typename TestCaseType>
+void printWorstFailingCase(llvm::raw_ostream &OS,
+                           const TestCaseType &TestCase) noexcept {
+  OS << "--- Worst Failing Case ---\n";
+  OS << llvm::formatv("  {0,-14} : ", "Input(s)");
+  printValues(OS, TestCase.Inputs);
+  OS << "\n";
+
+  OS << llvm::formatv("  {0,-14} : ", "Actual");
+  printValue(OS, TestCase.Actual);
+  OS << "\n";
+
+  OS << llvm::formatv("  {0,-14} : ", "Expected");
+  printValue(OS, TestCase.Expected);
+  OS << "\n";
+}
+
+template <typename TestType, typename ResultType>
+void printReport(const TestType &Test, const ResultType &Result,
+                 const std::chrono::steady_clock::duration &Duration) noexcept {
+  using FunctionConfig = typename TestType::FunctionConfig;
+
+  const bool Passed = Result.hasPassed();
+  const auto ElapsedMilliseconds =
+      std::chrono::duration_cast<std::chrono::milliseconds>(Duration).count();
+
+  llvm::errs() << llvm::formatv("=== Test Report for '{0}' === \n",
+                                FunctionConfig::Name);
+  llvm::errs() << llvm::formatv("{0,-17}: {1} ({2})\n", "Device",
+                                Test.getContext().getName(),
+                                Test.getContext().getPlatform());
+  llvm::errs() << llvm::formatv("{0,-17}: {1} ms\n", "Elapsed time",
+                                ElapsedMilliseconds);
+  llvm::errs() << llvm::formatv("{0,-17}: {1}\n", "ULP tolerance",
+                                FunctionConfig::UlpTolerance);
+  llvm::errs() << llvm::formatv("{0,-17}: {1}\n", "Max ULP distance",
+                                Result.getMaxUlpDistance());
+  llvm::errs() << llvm::formatv("{0,-17}: {1}\n", "Test cases",
+                                Result.getTestCaseCount());
+  llvm::errs() << llvm::formatv("{0,-17}: {1}\n", "Failures",
+                                Result.getFailureCount());
+  llvm::errs() << llvm::formatv("{0,-17}: {1}\n", "Status",
+                                Passed ? "PASSED" : "FAILED");
+
+  if (auto Worst = Result.getWorstFailingCase()) {
+    printWorstFailingCase(llvm::errs(), Worst.value());
+  }
+
+  llvm::errs().flush();
+}
+} // namespace detail
+
+template <typename TestType>
+[[nodiscard]] bool
+runTest(const TestType &Test,
+        typename TestType::GeneratorType &Generator) noexcept {
+  const auto StartTime = std::chrono::steady_clock::now();
+
+  auto Result = Test.run(Generator);
+
+  const auto EndTime = std::chrono::steady_clock::now();
+  const auto Duration = EndTime - StartTime;
+
+  detail::printReport(Test, Result, Duration);
+
+  return Result.hasPassed();
+}
+} // namespace mathtest
diff --git a/offload/unittests/Conformance/include/mathtest/TypeExtras.hpp b/offload/unittests/Conformance/include/mathtest/TypeExtras.hpp
new file mode 100644
index 0000000000000..3242349ce6b4d
--- /dev/null
+++ b/offload/unittests/Conformance/include/mathtest/TypeExtras.hpp
@@ -0,0 +1,9 @@
+#pragma once
+
+namespace mathtest {
+
+#ifdef __FLT16_MAX__
+#define MATHTEST_HAS_FLOAT16
+typedef _Float16 float16;
+#endif
+} // namespace mathtest
diff --git a/offload/unittests/Conformance/lib/CMakeLists.txt b/offload/unittests/Conformance/lib/CMakeLists.txt
new file mode 100644
index 0000000000000..a0402a54fbadf
--- /dev/null
+++ b/offload/unittests/Conformance/lib/CMakeLists.txt
@@ -0,0 +1,5 @@
+add_library(MathTest STATIC DeviceContext.cpp DeviceResources.cpp ErrorHandling.cpp)
+
+target_include_directories(MathTest PUBLIC "${CMAKE_CURRENT_SOURCE_DIR}/../include")
+target_compile_options(MathTest PUBLIC -fno-rtti)
+target_link_libraries(MathTest PUBLIC LLVMOffload LLVMSupport LLVMDemangle)
diff --git a/offload/unittests/Conformance/lib/DeviceContext.cpp b/offload/unittests/Conformance/lib/DeviceContext.cpp
new file mode 100644
index 0000000000000..a0ca8bab3ddf9
--- /dev/null
+++ b/offload/unittests/Conformance/lib/DeviceContext.cpp
@@ -0,0 +1,201 @@
+#include "mathtest/DeviceContext.hpp"
+
+#include "mathtest/ErrorHandling.hpp"
+
+#include "llvm/ADT/SmallString.h"
+#include "llvm/ADT/Twine.h"
+#include "llvm/Support/ErrorHandling.h"
+#include "llvm/Support/ErrorOr.h"
+#include "llvm/Support/MemoryBuffer.h"
+#include "llvm/Support/Path.h"
+
+#include <OffloadAPI.h>
+#include <cstddef>
+#include <memory>
+#include <string>
+#include <system_error>
+#include <vector>
+
+using namespace mathtest;
+
+//===----------------------------------------------------------------------===//
+// Helpers
+//===----------------------------------------------------------------------===//
+
+namespace {
+
+// The static 'Wrapper' instance ensures olInit() is called once at program
+// startup and olShutDown() is called once at program termination
+struct OffloadInitWrapper {
+  OffloadInitWrapper() { OL_CHECK(olInit()); }
+  ~OffloadInitWrapper() { OL_CHECK(olShutDown()); }
+};
+static OffloadInitWrapper Wrapper{};
+
+[[nodiscard]] ol_platform_backend_t
+getBackend(ol_device_handle_t DeviceHandle) noexcept {
+  ol_platform_handle_t Platform;
+  OL_CHECK(olGetDeviceInfo(DeviceHandle, OL_DEVICE_INFO_PLATFORM,
+                           sizeof(Platform), &Platform));
+  ol_platform_backend_t Backend = OL_PLATFORM_BACKEND_UNKNOWN;
+  OL_CHECK(olGetPlatformInfo(Platform, OL_PLATFORM_INFO_BACKEND,
+                             sizeof(Backend), &Backend));
+  return Backend;
+}
+
+const std::vector<ol_device_handle_t> &getDevices() {
+  // Thread-safe initialization of a static local variable
+  static std::vector<ol_device_handle_t> Devices =
+      []() -> std::vector<ol_device_handle_t> {
+    std::vector<ol_device_handle_t> TmpDevices;
+
+    // Discovers all devices that are not the host
+    const auto *const ResultFromIterate = olIterateDevices(
+        [](ol_device_handle_t DeviceHandle, void *Data) {
+          if (getBackend(DeviceHandle) != OL_PLATFORM_BACKEND_HOST) {
+            static_cast<std::vector<ol_device_handle_t> *>(Data)->push_back(
+                DeviceHandle);
+          }
+          return true;
+        },
+        &TmpDevices);
+
+    OL_CHECK(ResultFromIterate);
+
+    return TmpDevices;
+  }();
+
+  return Devices;
+}
+} // namespace
+
+std::size_t mathtest::countDevices() { return getDevices().size(); }
+
+void detail::allocManagedMemory(ol_device_handle_t DeviceHandle,
+                                std::size_t Size,
+                                void **AllocationOut) noexcept {
+  OL_CHECK(
+      olMemAlloc(DeviceHandle, OL_ALLOC_TYPE_MANAGED, Size, AllocationOut));
+}
+
+//===----------------------------------------------------------------------===//
+// DeviceContext
+//===----------------------------------------------------------------------===//
+
+DeviceContext::DeviceContext(std::size_t DeviceId)
+    : DeviceId(DeviceId), DeviceHandle(nullptr) {
+  const auto &Devices = getDevices();
+
+  if (DeviceId >= Devices.size()) {
+    FATAL_ERROR("Invalid DeviceId: " + llvm::Twine(DeviceId) + ", but only " +
+                llvm::Twine(Devices.size()) + " devices are available");
+  }
+
+  DeviceHandle = Devices[DeviceId];
+}
+
+[[nodiscard]] std::shared_ptr<DeviceImage>
+DeviceContext::loadBinary(llvm::StringRef Directory, llvm::StringRef BinaryName,
+                          llvm::StringRef Extension) const {
+  llvm::SmallString<128> FullPath(Directory);
+  llvm::sys::path::append(FullPath, llvm::Twine(BinaryName) + Extension);
+
+  // For simplicity, this implementation intentionally reads the binary from
+  // disk on every call.
+  //
+  // Other use cases could benefit from a global, thread-safe cache to avoid
+  // redundant file I/O and GPU program creation.
+
+  llvm::ErrorOr<std::unique_ptr<llvm::MemoryBuffer>> FileOrErr =
+      llvm::MemoryBuffer::getFile(FullPath);
+  if (std::error_code ErrorCode = FileOrErr.getError()) {
+    FATAL_ERROR(llvm::Twine("Failed to read device binary file '") + FullPath +
+                "': " + ErrorCode.message());
+  }
+  std::unique_ptr<llvm::MemoryBuffer> &BinaryData = *FileOrErr;
+
+  ol_program_handle_t ProgramHandle = nullptr;
+  OL_CHECK(olCreateProgram(DeviceHandle, BinaryData->getBufferStart(),
+                           BinaryData->getBufferSize(), &ProgramHandle));
+
+  return std::shared_ptr<DeviceImage>(
+      new DeviceImage(DeviceHandle, ProgramHandle));
+}
+
+[[nodiscard]] std::shared_ptr<DeviceImage>
+DeviceContext::loadBinary(llvm::StringRef Directory,
+                          llvm::StringRef BinaryName) const {
+  llvm::StringRef Extension;
+
+  switch (getBackend(DeviceHandle)) {
+  case OL_PLATFORM_BACKEND_AMDGPU:
+    Extension = ".amdgpu.bin";
+    break;
+  case OL_PLATFORM_BACKEND_CUDA:
+    Extension = ".nvptx64.bin";
+    break;
+  default:
+    llvm_unreachable("Unsupported backend to infer binary extension");
+  }
+
+  return loadBinary(Directory, BinaryName, Extension);
+}
+
+void DeviceContext::getKernelImpl(
+    ol_program_handle_t ProgramHandle, llvm::StringRef KernelName,
+    ol_symbol_handle_t *KernelHandle) const noexcept {
+  llvm::SmallString<32> KernelNameBuffer(KernelName);
+  OL_CHECK(olGetSymbol(ProgramHandle, KernelNameBuffer.c_str(),
+                       OL_SYMBOL_KIND_KERNEL, KernelHandle));
+}
+
+void DeviceContext::launchKernelImpl(
+    ol_symbol_handle_t KernelHandle, const Dim &NumGroups, const Dim &GroupSize,
+    const void *KernelArgs, std::size_t KernelArgsSize) const noexcept {
+  ol_kernel_launch_size_args_t LaunchArgs;
+  LaunchArgs.Dimensions = 3; // It seems this field is not used anywhere.
+                             // Defaulting to the safest value
+  LaunchArgs.NumGroups = {NumGroups[0], NumGroups[1], NumGroups[2]};
+  LaunchArgs.GroupSize = {GroupSize[0], GroupSize[1], GroupSize[2]};
+  LaunchArgs.DynSharedMemory = 0;
+
+  OL_CHECK(olLaunchKernel(nullptr, DeviceHandle, KernelHandle, KernelArgs,
+                          KernelArgsSize, &LaunchArgs, nullptr));
+}
+
+[[nodiscard]] std::string DeviceContext::getName() const {
+  std::size_t PropSize = 0;
+  OL_CHECK(olGetDeviceInfoSize(DeviceHandle, OL_DEVICE_INFO_NAME, &PropSize));
+
+  if (PropSize == 0) {
+    return "";
+  }
+
+  std::string PropValue(PropSize, '\0');
+  OL_CHECK(olGetDeviceInfo(DeviceHandle, OL_DEVICE_INFO_NAME, PropSize,
+                           PropValue.data()));
+  PropValue.pop_back(); // Remove the null terminator
+
+  return PropValue;
+}
+
+[[nodiscard]] std::string DeviceContext::getPlatform() const {
+  ol_platform_handle_t PlatformHandle = nullptr;
+  OL_CHECK(olGetDeviceInfo(DeviceHandle, OL_DEVICE_INFO_PLATFORM,
+                           sizeof(ol_platform_handle_t), &PlatformHandle));
+
+  std::size_t PropSize = 0;
+  OL_CHECK(
+      olGetPlatformInfoSize(PlatformHandle, OL_PLATFORM_INFO_NAME, &PropSize));
+
+  if (PropSize == 0) {
+    return "";
+  }
+
+  std::string PropValue(PropSize, '\0');
+  OL_CHECK(olGetPlatformInfo(PlatformHandle, OL_PLATFORM_INFO_NAME, PropSize,
+                             PropValue.data()));
+  PropValue.pop_back(); // Remove the null terminator
+
+  return PropValue;
+}
diff --git a/offload/unittests/Conformance/lib/DeviceResources.cpp b/offload/unittests/Conformance/lib/DeviceResources.cpp
new file mode 100644
index 0000000000000..5d1c94dca7677
--- /dev/null
+++ b/offload/unittests/Conformance/lib/DeviceResources.cpp
@@ -0,0 +1,54 @@
+#include "mathtest/DeviceResources.hpp"
+
+#include "mathtest/ErrorHandling.hpp"
+
+#include <OffloadAPI.h>
+
+using namespace mathtest;
+
+//===----------------------------------------------------------------------===//
+// Helpers
+//===----------------------------------------------------------------------===//
+
+void detail::freeDeviceMemory(void *Address) noexcept {
+  if (Address) {
+    OL_CHECK(olMemFree(Address));
+  }
+}
+
+//===----------------------------------------------------------------------===//
+// DeviceImage
+//===----------------------------------------------------------------------===//
+
+DeviceImage::~DeviceImage() noexcept {
+  if (Handle) {
+    OL_CHECK(olDestroyProgram(Handle));
+  }
+}
+
+DeviceImage &DeviceImage::operator=(DeviceImage &&Other) noexcept {
+  if (this == &Other)
+    return *this;
+
+  if (Handle) {
+    OL_CHECK(olDestroyProgram(Handle));
+  }
+
+  DeviceHandle = Other.DeviceHandle;
+  Handle = Other.Handle;
+
+  Other.DeviceHandle = nullptr;
+  Other.Handle = nullptr;
+
+  return *this;
+}
+
+DeviceImage::DeviceImage(DeviceImage &&Other) noexcept
+    : DeviceHandle(Other.DeviceHandle), Handle(Other.Handle) {
+  Other.DeviceHandle = nullptr;
+  Other.Handle = nullptr;
+}
+
+DeviceImage::DeviceImage(ol_device_handle_t DeviceHandle,
+                         ol_program_handle_t Handle) noexcept
+    : DeviceHandle(DeviceHandle), Handle(Handle) {}
diff --git a/offload/unittests/Conformance/lib/ErrorHandling.cpp b/offload/unittests/Conformance/lib/ErrorHandling.cpp
new file mode 100644
index 0000000000000..0f85260bc7926
--- /dev/null
+++ b/offload/unittests/Conformance/lib/ErrorHandling.cpp
@@ -0,0 +1,37 @@
+#include "mathtest/ErrorHandling.hpp"
+
+#include "llvm/ADT/Twine.h"
+#include "llvm/Support/ErrorHandling.h"
+
+#include <OffloadAPI.h>
+
+using namespace mathtest;
+
+[[noreturn]] void detail::reportFatalError(const llvm::Twine &Message,
+                                           const char *File, int Line,
+                                           const char *FuncName) {
+  // clang-format off
+  llvm::report_fatal_error(
+      llvm::Twine("Fatal error in '") + FuncName +
+          "' at " + File + ":" + llvm::Twine(Line) +
+          "\n  Message: " + Message,
+      /*gen_crash_diag=*/false);
+  // clang-format on
+}
+
+[[noreturn]] void detail::reportOffloadError(const char *ResultExpr,
+                                             ol_result_t Result,
+                                             const char *File, int Line,
+                                             const char *FuncName) {
+  // clang-format off
+  llvm::report_fatal_error(
+      llvm::Twine("OL_CHECK failed") +
+          "\n  Location: " + File + ":" + llvm::Twine(Line) +
+          "\n  Function: " + FuncName +
+          "\n  Expression: " + ResultExpr +
+          "\n  Error code: " + llvm::Twine(Result->Code) +
+          "\n  Details: " +
+          (Result->Details ? Result->Details : "No details provided"),
+      /*gen_crash_diag=*/false);
+  // clang-format on
+}
diff --git a/offload/unittests/Conformance/sin.cpp b/offload/unittests/Conformance/sin.cpp
deleted file mode 100644
index 9e15690a9e9d7..0000000000000
--- a/offload/unittests/Conformance/sin.cpp
+++ /dev/null
@@ -1,8 +0,0 @@
-#include "llvm/Support/MemoryBuffer.h"
-#include "llvm/Support/raw_ostream.h"
-#include <OffloadAPI.h>
-#include <math.h>
-
-llvm::StringRef DeviceBinsDirectory = DEVICE_CODE_PATH;
-
-int main() { llvm::errs() << sin(0.0) << "\n"; }
diff --git a/offload/unittests/Conformance/tests/CMakeLists.txt b/offload/unittests/Conformance/tests/CMakeLists.txt
new file mode 100644
index 0000000000000..b5da56f46fc05
--- /dev/null
+++ b/offload/unittests/Conformance/tests/CMakeLists.txt
@@ -0,0 +1,2 @@
+add_conformance_test(Hypotf16Test Hypotf16Test.cpp)
+add_conformance_test(LogfTest LogfTest.cpp)
diff --git a/offload/unittests/Conformance/tests/Hypotf16Test.cpp b/offload/unittests/Conformance/tests/Hypotf16Test.cpp
new file mode 100644
index 0000000000000..8c6b5054e2d4c
--- /dev/null
+++ b/offload/unittests/Conformance/tests/Hypotf16Test.cpp
@@ -0,0 +1,51 @@
+#include "mathtest/TypeExtras.hpp"
+
+#ifdef MATHTEST_HAS_FLOAT16
+#include "mathtest/DeviceContext.hpp"
+#include "mathtest/ExhaustiveGenerator.hpp"
+#include "mathtest/GpuMathTest.hpp"
+#include "mathtest/IndexedRange.hpp"
+#include "mathtest/TestRunner.hpp"
+
+#include "llvm/ADT/StringRef.h"
+
+#include <cstdlib>
+#include <math.h>
+#include <memory>
+
+using namespace mathtest;
+
+extern "C" {
+
+float16 hypotf16(float16, float16);
+}
+
+namespace mathtest {
+
+template <> struct FunctionConfig<hypotf16> {
+  static constexpr llvm::StringRef Name = "hypotf16";
+  static constexpr llvm::StringRef KernelName = "hypotf16Kernel";
+
+  // Source: The Khronos Group, The OpenCL C Specification v3.0.19, Sec. 7.4,
+  //         Table 69 (Full Profile), Khronos Registry [July 10, 2025].
+  static constexpr uint64_t UlpTolerance = 2;
+};
+} // namespace mathtest
+
+int main() {
+  // TODO: Add command-line arguments parsing for test configuration.
+  auto Context = std::make_shared<DeviceContext>(/*DeviceId=*/0);
+  const llvm::StringRef Provider = "llvm-libm";
+  const llvm::StringRef DeviceBinsDirectory = DEVICE_CODE_PATH;
+
+  GpuMathTest<hypotf16> Hypotf16Test(Context, Provider, DeviceBinsDirectory);
+
+  IndexedRange<float16> RangeX;
+  IndexedRange<float16> RangeY;
+  ExhaustiveGenerator<float16, float16> Generator(RangeX, RangeY);
+
+  const auto Passed = runTest(Hypotf16Test, Generator);
+
+  return Passed ? EXIT_SUCCESS : EXIT_FAILURE;
+}
+#endif // MATHTEST_HAS_FLOAT16
diff --git a/offload/unittests/Conformance/tests/LogfTest.cpp b/offload/unittests/Conformance/tests/LogfTest.cpp
new file mode 100644
index 0000000000000..1af5e844ccdb1
--- /dev/null
+++ b/offload/unittests/Conformance/tests/LogfTest.cpp
@@ -0,0 +1,44 @@
+#include "mathtest/DeviceContext.hpp"
+#include "mathtest/ExhaustiveGenerator.hpp"
+#include "mathtest/GpuMathTest.hpp"
+#include "mathtest/IndexedRange.hpp"
+#include "mathtest/TestRunner.hpp"
+
+#include "llvm/ADT/StringRef.h"
+
+#include <cstdlib>
+#include <limits>
+#include <math.h>
+#include <memory>
+
+namespace mathtest {
+
+template <> struct FunctionConfig<logf> {
+  static constexpr llvm::StringRef Name = "logf";
+  static constexpr llvm::StringRef KernelName = "logfKernel";
+
+  // Source: The Khronos Group, The OpenCL C Specification v3.0.19, Sec. 7.4,
+  //         Table 65, Khronos Registry [July 10, 2025].
+  static constexpr uint64_t UlpTolerance = 3;
+};
+} // namespace mathtest
+
+int main() {
+  using namespace mathtest;
+
+  // TODO: Add command-line arguments parsing for test configuration.
+  auto Context = std::make_shared<DeviceContext>(/*DeviceId=*/0);
+  const llvm::StringRef Provider = "llvm-libm";
+  const llvm::StringRef DeviceBinsDirectory = DEVICE_CODE_PATH;
+
+  GpuMathTest<logf> LogfTest(Context, Provider, DeviceBinsDirectory);
+
+  IndexedRange<float> Range(/*Begin=*/0.0f,
+                            /*End=*/std::numeric_limits<float>::infinity(),
+                            /*Inclusive=*/true);
+  ExhaustiveGenerator<float> Generator(Range);
+
+  const auto Passed = runTest(LogfTest, Generator);
+
+  return Passed ? EXIT_SUCCESS : EXIT_FAILURE;
+}
\ No newline at end of file

>From 7a80f2152ca9f714b5b4a3439c197e8d939613cb Mon Sep 17 00:00:00 2001
From: Leandro Augusto Lacerda Campos <leandrolcampos at yahoo.com.br>
Date: Thu, 17 Jul 2025 01:08:40 -0300
Subject: [PATCH 2/5] Add trailing newline

---
 offload/unittests/Conformance/tests/LogfTest.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/offload/unittests/Conformance/tests/LogfTest.cpp b/offload/unittests/Conformance/tests/LogfTest.cpp
index 1af5e844ccdb1..1c866720843ff 100644
--- a/offload/unittests/Conformance/tests/LogfTest.cpp
+++ b/offload/unittests/Conformance/tests/LogfTest.cpp
@@ -41,4 +41,4 @@ int main() {
   const auto Passed = runTest(LogfTest, Generator);
 
   return Passed ? EXIT_SUCCESS : EXIT_FAILURE;
-}
\ No newline at end of file
+}

>From 2ad044d4ad1c3aafb7c1361bd9363016091d4f86 Mon Sep 17 00:00:00 2001
From: Leandro Augusto Lacerda Campos <leandrolcampos at yahoo.com.br>
Date: Thu, 17 Jul 2025 01:15:44 -0300
Subject: [PATCH 3/5] Pass correct GPU architecture to device compiler

---
 offload/unittests/CMakeLists.txt | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/offload/unittests/CMakeLists.txt b/offload/unittests/CMakeLists.txt
index bea49387528b0..ad64f800da014 100644
--- a/offload/unittests/CMakeLists.txt
+++ b/offload/unittests/CMakeLists.txt
@@ -39,7 +39,7 @@ function(add_offload_test_device_code test_filename test_name)
       add_custom_command(
         OUTPUT ${output_file}
         COMMAND ${CMAKE_C_COMPILER}
-        --target=nvptx64-nvidia-cuda -march=native
+        --target=nvptx64-nvidia-cuda -march=${nvptx_arch}
         -stdlib -nogpulib --cuda-path=${CUDA_ROOT} -flto -fno-builtin ${ARGN}
         ${SRC_PATH} -o ${output_file}
         DEPENDS ${SRC_PATH}

>From c0d5b217941401c44c804199eded5ea65ef1a2f7 Mon Sep 17 00:00:00 2001
From: Leandro Augusto Lacerda Campos <leandrolcampos at yahoo.com.br>
Date: Thu, 17 Jul 2025 01:35:15 -0300
Subject: [PATCH 4/5] Annotate `#endif` directives for clarity

---
 offload/unittests/Conformance/device_code/LLVMLibm.c   |  4 ++--
 .../Conformance/include/mathtest/Numerics.hpp          | 10 +++++-----
 .../Conformance/include/mathtest/OffloadForward.hpp    |  4 ++--
 .../Conformance/include/mathtest/TypeExtras.hpp        |  2 +-
 4 files changed, 10 insertions(+), 10 deletions(-)

diff --git a/offload/unittests/Conformance/device_code/LLVMLibm.c b/offload/unittests/Conformance/device_code/LLVMLibm.c
index 5f436a235eb08..ad08344b1c538 100644
--- a/offload/unittests/Conformance/device_code/LLVMLibm.c
+++ b/offload/unittests/Conformance/device_code/LLVMLibm.c
@@ -6,7 +6,7 @@
 #ifdef __FLT16_MAX__
 #define HAS_FLOAT16
 typedef _Float16 float16;
-#endif
+#endif // __FLT16_MAX__
 
 #ifdef HAS_FLOAT16
 __gpu_kernel void hypotf16Kernel(const float16 *X, float16 *Y, float16 *Out,
@@ -18,7 +18,7 @@ __gpu_kernel void hypotf16Kernel(const float16 *X, float16 *Y, float16 *Out,
     Out[Index] = hypotf16(X[Index], Y[Index]);
   }
 }
-#endif
+#endif // HAS_FLOAT16
 
 __gpu_kernel void logfKernel(const float *X, float *Out, size_t NumElements) {
   uint32_t Index =
diff --git a/offload/unittests/Conformance/include/mathtest/Numerics.hpp b/offload/unittests/Conformance/include/mathtest/Numerics.hpp
index f43a26974dda7..3957b2432b12e 100644
--- a/offload/unittests/Conformance/include/mathtest/Numerics.hpp
+++ b/offload/unittests/Conformance/include/mathtest/Numerics.hpp
@@ -36,7 +36,7 @@ template <typename T> struct StorageTypeOf {
 template <> struct StorageTypeOf<float16> {
   using type = uint16_t;
 };
-#endif
+#endif // MATHTEST_HAS_FLOAT16
 
 template <> struct StorageTypeOf<float> {
   using type = uint32_t;
@@ -52,7 +52,7 @@ template <typename T> struct IsFloatingPoint : std::is_floating_point<T> {};
 
 #ifdef MATHTEST_HAS_FLOAT16
 template <> struct IsFloatingPoint<float16> : std::true_type {};
-#endif
+#endif // MATHTEST_HAS_FLOAT16
 
 template <typename T>
 inline constexpr bool IsFloatingPoint_v // NOLINT(readability-identifier-naming)
@@ -101,7 +101,7 @@ template <> struct FPLayout<float16> {
   static constexpr std::size_t ExponentLen = 5;
   static constexpr std::size_t FractionLen = 10;
 };
-#endif
+#endif // MATHTEST_HAS_FLOAT16
 
 template <> struct FPLayout<float> {
   static constexpr std::size_t SignLen = 1;
@@ -167,7 +167,7 @@ template <> [[nodiscard]] constexpr float16 getMinOrNegInf<float16>() noexcept {
 
   return __builtin_bit_cast(float16, static_cast<StorageType>(0xFC00U));
 }
-#endif
+#endif // MATHTEST_HAS_FLOAT16
 
 template <typename T> [[nodiscard]] constexpr T getMaxOrInf() noexcept {
   static_assert(std::is_arithmetic_v<T>, "Type T must be an arithmetic type");
@@ -186,7 +186,7 @@ template <> [[nodiscard]] constexpr float16 getMaxOrInf<float16>() noexcept {
 
   return __builtin_bit_cast(float16, static_cast<StorageType>(0x7C00U));
 }
-#endif
+#endif // MATHTEST_HAS_FLOAT16
 
 template <typename FloatType>
 [[nodiscard]] uint64_t computeUlpDistance(FloatType X, FloatType Y) noexcept {
diff --git a/offload/unittests/Conformance/include/mathtest/OffloadForward.hpp b/offload/unittests/Conformance/include/mathtest/OffloadForward.hpp
index 099b86af2929d..9609bd19f06b2 100644
--- a/offload/unittests/Conformance/include/mathtest/OffloadForward.hpp
+++ b/offload/unittests/Conformance/include/mathtest/OffloadForward.hpp
@@ -2,7 +2,7 @@
 
 #ifdef __cplusplus
 extern "C" {
-#endif
+#endif // __cplusplus
 
 struct ol_error_struct_t;
 typedef const ol_error_struct_t *ol_result_t;
@@ -19,4 +19,4 @@ typedef struct ol_symbol_impl_t *ol_symbol_handle_t;
 
 #ifdef __cplusplus
 }
-#endif
+#endif // __cplusplus
diff --git a/offload/unittests/Conformance/include/mathtest/TypeExtras.hpp b/offload/unittests/Conformance/include/mathtest/TypeExtras.hpp
index 3242349ce6b4d..bfd58be5050ea 100644
--- a/offload/unittests/Conformance/include/mathtest/TypeExtras.hpp
+++ b/offload/unittests/Conformance/include/mathtest/TypeExtras.hpp
@@ -5,5 +5,5 @@ namespace mathtest {
 #ifdef __FLT16_MAX__
 #define MATHTEST_HAS_FLOAT16
 typedef _Float16 float16;
-#endif
+#endif // __FLT16_MAX__
 } // namespace mathtest

>From 0a6defa45f12d963aec1093e3e755a3354fd5423 Mon Sep 17 00:00:00 2001
From: Leandro Augusto Lacerda Campos <leandrolcampos at yahoo.com.br>
Date: Fri, 18 Jul 2025 14:59:58 -0300
Subject: [PATCH 5/5] Rename aliases for consistency

---
 .../Conformance/include/mathtest/HostRefChecker.hpp    | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)

diff --git a/offload/unittests/Conformance/include/mathtest/HostRefChecker.hpp b/offload/unittests/Conformance/include/mathtest/HostRefChecker.hpp
index c45137d652df2..a0083b1d1cedf 100644
--- a/offload/unittests/Conformance/include/mathtest/HostRefChecker.hpp
+++ b/offload/unittests/Conformance/include/mathtest/HostRefChecker.hpp
@@ -15,16 +15,16 @@
 namespace mathtest {
 
 template <auto Func> class HostRefChecker {
-  using FuncTraits = FunctionTraits<Func>;
-  using InTypesTuple = typename FuncTraits::ArgTypesTuple;
+  using FunctionTraits = FunctionTraits<Func>;
+  using InTypesTuple = typename FunctionTraits::ArgTypesTuple;
 
-  using FuncConfig = FunctionConfig<Func>;
+  using FunctionConfig = FunctionConfig<Func>;
 
   template <typename... Ts>
   using BuffersTupleType = std::tuple<llvm::ArrayRef<Ts>...>;
 
 public:
-  using OutType = typename FuncTraits::ReturnType;
+  using OutType = typename FunctionTraits::ReturnType;
 
 private:
   template <typename... Ts>
@@ -62,7 +62,7 @@ template <auto Func> class HostRefChecker {
       const OutType Expected = std::apply(Func, CurrentInputsTuple);
 
       const auto UlpDistance = computeUlpDistance(Actual, Expected);
-      const bool IsFailure = UlpDistance > FuncConfig::UlpTolerance;
+      const bool IsFailure = UlpDistance > FunctionConfig::UlpTolerance;
 
       return ResultType(UlpDistance, IsFailure,
                         typename ResultType::TestCase(



More information about the llvm-commits mailing list