Rework test harness

.github/workflows/ci.yml

@@ -64,3 +64,6 @@ jobs:
 
     - name: 'CMake Build Release'
       run: cmake --build ${{ github.workspace }}/release --config Release --parallel 4
+
+    - name: 'Test Release'
+      run: ctest --test-dir ${{ github.workspace }}/release -C Release --output-on-failure

CMakeLists.txt

@@ -44,6 +44,7 @@ if(FASTSIMD_DISPATCH_CLASS)
 endif()
 
 if(FASTSIMD_TESTS)
+    enable_testing()
     add_subdirectory(tests)
 endif()
 

include/FastSIMD/ToolSet/x86/128/f32x4.h

@@ -224,7 +224,8 @@ namespace FS
         }
         else
         {
-            return _mm_xor_ps( ifFalse.native, _mm_and_ps( mask.native, _mm_xor_ps( ifTrue.native, ifFalse.native ) ) );
+            // MSVC 19.51 miscompiles the equivalent xor( ifFalse, and( mask, xor( ifTrue, ifFalse ) ) ) inside FS::ATan2
+            return _mm_xor_ps( ifTrue.native, _mm_andnot_ps( mask.native, _mm_xor_ps( ifTrue.native, ifFalse.native ) ) );
         }
     }
 

tests/CMakeLists.txt

@@ -1,13 +1,60 @@
 
-fastsimd_create_dispatch_library(simd_test SOURCES "test.inl" FEATURE_SETS SCALAR SSE2 SSE41 AVX2 AVX512 NEON AARCH64 WASM)
-fastsimd_create_dispatch_library(simd_test_relaxed RELAXED SOURCES "test.inl" FEATURE_SETS SCALAR SSE2 SSE41 AVX2 AVX512 NEON AARCH64 WASM)
+set(FASTSIMD_TEST_FEATURE_SETS SCALAR SSE2 SSE41 AVX2 AVX512 NEON AARCH64 WASM)
 
-add_executable(test "test.cpp")
-target_link_libraries(test PRIVATE FastSIMD simd_test simd_test_relaxed)
+# Each register size is its own pair of dispatch libraries, test.inl instantiates FASTSIMD_TEST_REGISTER_BYTES
+# All sizes in one translation unit needs ~4.5GB per unit in a GCC Debug build, too much for 4 parallel jobs on a CI runner
+# Keep in sync with kTestRegisterBytes in test.h
+set(simd_test_libraries "")
+foreach(register_bytes 16 32 64 128)
+  fastsimd_create_dispatch_library(simd_test_${register_bytes} SOURCES "test.inl" FEATURE_SETS ${FASTSIMD_TEST_FEATURE_SETS})
+  fastsimd_create_dispatch_library(simd_test_relaxed_${register_bytes} RELAXED SOURCES "test.inl" FEATURE_SETS ${FASTSIMD_TEST_FEATURE_SETS})
+
+  foreach(simd_test_library simd_test_${register_bytes} simd_test_relaxed_${register_bytes})
+    target_compile_definitions(${simd_test_library} PRIVATE FASTSIMD_TEST_REGISTER_BYTES=${register_bytes})
+
+    # Every test is too many sections for an unoptimised COFF object file
+    if(MSVC)
+      target_compile_options(${simd_test_library} PRIVATE /bigobj)
+    elseif(WIN32 AND CMAKE_CXX_COMPILER_ID STREQUAL "GNU")
+      # GNU as (MinGW) has the same limit, clang's integrated assembler switches to big objects itself
+      target_compile_options(${simd_test_library} PRIVATE -Wa,-mbig-obj)
+    endif()
+
+    list(APPEND simd_test_libraries ${simd_test_library})
+  endforeach()
+endforeach()
+
+# "test" is a reserved target name when CTest is enabled
+add_executable(FastSIMD_tests "test.cpp" "test_inputs.cpp")
+target_link_libraries(FastSIMD_tests PRIVATE FastSIMD ${simd_test_libraries})
+
+# Random inputs must be the same on every platform for a seed, stop arm64 compilers fusing the range maths into FMA
+if(MSVC AND CMAKE_CXX_COMPILER_ID STREQUAL "Clang")
+  target_compile_options(FastSIMD_tests PRIVATE /clang:-ffp-contract=off)
+elseif(NOT CMAKE_CXX_COMPILER_ID STREQUAL "MSVC")
+  target_compile_options(FastSIMD_tests PRIVATE -ffp-contract=off)
+endif()
+
+# Fewer random inputs keeps ctest (and CI) quick, run FastSIMD_tests directly for a full run
+set(FASTSIMD_CTEST_INPUT_COUNT 65536 CACHE STRING "FastSIMD random input values per test when run through ctest")
+add_test(NAME FastSIMD_tests COMMAND FastSIMD_tests --count ${FASTSIMD_CTEST_INPUT_COUNT})
 
 if(CMAKE_SYSTEM_NAME STREQUAL "Emscripten")
-  set(CMAKE_EXECUTABLE_SUFFIX ".html")
-  target_link_options(test PRIVATE -sALLOW_MEMORY_GROWTH=1 -sSINGLE_FILE)
+  # Runs under node, which the Emscripten toolchain sets as the ctest emulator
+  # EXIT_RUNTIME makes main's return value the exit code
+  # Debug info makes the single file output over 1GB, larger than node can load, so it's left out when linking
+  target_link_options(FastSIMD_tests PRIVATE -sALLOW_MEMORY_GROWTH=1 -sSINGLE_FILE -sEXIT_RUNTIME=1 $<$<CONFIG:Debug>:-g0>)
+
+  # The toolchain uses whichever node is first on PATH, emsdk's own node supports the relaxed SIMD the relaxed tests use
+  if(DEFINED ENV{EMSDK_NODE})
+    set_target_properties(FastSIMD_tests PROPERTIES CROSSCOMPILING_EMULATOR "$ENV{EMSDK_NODE}")
+  endif()
+
+  # Single page to run the tests in a browser
+  add_executable(FastSIMD_tests_web "test.cpp" "test_inputs.cpp")
+  target_link_libraries(FastSIMD_tests_web PRIVATE FastSIMD ${simd_test_libraries})
+  set_target_properties(FastSIMD_tests_web PROPERTIES SUFFIX ".html")
+  target_link_options(FastSIMD_tests_web PRIVATE -sALLOW_MEMORY_GROWTH=1 -sSINGLE_FILE $<$<CONFIG:Debug>:-g0>)
 endif()
 
 # Enable Warnings