fixed potential build issues for windows (winarm, generic, arm32bit, …

…minor fix for avx512)
nvpro-samples · Oct 24, 2024 · 99e368b · 99e368b
1 parent 1e89710
commit 99e368b
Show file tree

Hide file tree

Showing 2 changed files with 14 additions and 17 deletions.
diff --git a/vk_video_decoder/libs/NvVideoParser/CMakeLists.txt b/vk_video_decoder/libs/NvVideoParser/CMakeLists.txt
@@ -49,36 +49,34 @@ include_directories(BEFORE ${VK_VIDEO_COMMON_LIBS_SOURCE_ROOT})
 
 add_library(${VULKAN_VIDEO_PARSER_LIB} SHARED ${LIBNVPARSER})
 
-if (CMAKE_SYSTEM_PROCESSOR MATCHES "^aarch64")
+if ((CMAKE_SYSTEM_PROCESSOR MATCHES "^aarch64") OR (CMAKE_SYSTEM_PROCESSOR MATCHES "^arm64") OR (CMAKE_SYSTEM_PROCESSOR MATCHES "^ARM64"))
   if(WIN32)
-    set(GENERIC_CPU_FEATURE "/arch:armv8.0+nosimd")
     set(NEON_CPU_FEATURE "/arch:armv8.0")
-    set(SVE_CPU_FEATURE "/arch:armv8.0+sve")
+    # set(SVE_CPU_FEATURE "/arch:armv8.0+sve") # clang-cl limitation (SVE intrinsics are not supported by MSVC at the moment)
   elseif(UNIX)
-    set(GENERIC_CPU_FEATURE "-march=armv8-a+nosimd")
     set(NEON_CPU_FEATURE "-march=armv8-a")
     set(SVE_CPU_FEATURE "-march=armv8-a+sve")
   endif()
     add_library(next_start_code_c OBJECT ${CMAKE_CURRENT_SOURCE_DIR}/src/NextStartCodeC.cpp include)
-    set_target_properties(next_start_code_c PROPERTIES COMPILE_FLAGS ${GENERIC_CPU_FEATURE} )
     target_include_directories(next_start_code_c PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/include)
     add_library(next_start_code_neon OBJECT ${CMAKE_CURRENT_SOURCE_DIR}/src/NextStartCodeNEON.cpp include)
     set_target_properties(next_start_code_neon PROPERTIES COMPILE_FLAGS ${NEON_CPU_FEATURE} )
     target_include_directories(next_start_code_neon PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/include)
+  if(WIN32) # clang-cl limitation (SVE intrinsics are not supported by MSVC at the moment)
+    target_link_libraries(${VULKAN_VIDEO_PARSER_LIB} next_start_code_c next_start_code_neon)
+  elseif(UNIX)
     add_library(next_start_code_sve OBJECT ${CMAKE_CURRENT_SOURCE_DIR}/src/NextStartCodeSVE.cpp include)
     set_target_properties(next_start_code_sve PROPERTIES COMPILE_FLAGS ${SVE_CPU_FEATURE} )
     target_include_directories(next_start_code_sve PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/include)
     target_link_libraries(${VULKAN_VIDEO_PARSER_LIB} next_start_code_c next_start_code_neon next_start_code_sve)
-elseif (CMAKE_SYSTEM_PROCESSOR MATCHES "^aarch32")
+  endif()
+elseif ((CMAKE_SYSTEM_PROCESSOR MATCHES "^arm") OR (CMAKE_SYSTEM_PROCESSOR MATCHES "^ARM"))
   if(WIN32)
-    set(GENERIC_CPU_FEATURE "/arch:ARMv7VE")
     set(NEON_CPU_FEATURE "/arch:VFPv4")
   elseif(UNIX)
-    set(GENERIC_CPU_FEATURE "-march=armv7-a+nosimd")
     set(NEON_CPU_FEATURE "-march=armv7-a+simd")
   endif()
     add_library(next_start_code_c OBJECT ${CMAKE_CURRENT_SOURCE_DIR}/src/NextStartCodeC.cpp include)
-    set_target_properties(next_start_code_c PROPERTIES COMPILE_FLAGS ${GENERIC_CPU_FEATURE} )
     target_include_directories(next_start_code_c PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/include)
     add_library(next_start_code_neon OBJECT ${CMAKE_CURRENT_SOURCE_DIR}/src/NextStartCodeNEON.cpp include)
     set_target_properties(next_start_code_neon PROPERTIES COMPILE_FLAGS ${NEON_CPU_FEATURE} )
@@ -88,18 +86,13 @@ else()
   if(WIN32)
     set(SSSE3_CPU_FEATURE "/arch:SSE2")
     set(AVX2_CPU_FEATURE "/arch:AVX2")
-    set(AVX512_CPU_FEATURE "/arch:AVX512")
+    # set(AVX512_CPU_FEATURE "/arch:AVX512")
   elseif(UNIX)
-    set(GENERIC_CPU_FEATURE "-mtune=generic -mno-sse -mno-sse2 -mno-ssse3 -mno-sse4 -mno-sse4.2 -mno-avx \
-                             -mno-avx2 -mno-avx512f -mno-avx512vl -mno-avx512bw -mno-bmi2")
     set(SSSE3_CPU_FEATURE "-mssse3")
     set(AVX2_CPU_FEATURE "-mavx2")
     set(AVX512_CPU_FEATURE "-mavx512f -mavx512bw")
   endif()
   add_library(next_start_code_c OBJECT ${CMAKE_CURRENT_SOURCE_DIR}/src/NextStartCodeC.cpp include)
-  if(NOT WIN32)
-    set_target_properties(next_start_code_c PROPERTIES COMPILE_FLAGS ${GENERIC_CPU_FEATURE} )
-  endif()
   target_include_directories(next_start_code_c PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/include)
   add_library(next_start_code_ssse3 OBJECT ${CMAKE_CURRENT_SOURCE_DIR}/src/NextStartCodeSSSE3.cpp include)
   set_target_properties(next_start_code_ssse3 PROPERTIES COMPILE_FLAGS ${SSSE3_CPU_FEATURE} )
@@ -108,7 +101,9 @@ else()
   set_target_properties(next_start_code_avx2 PROPERTIES COMPILE_FLAGS ${AVX2_CPU_FEATURE} )
   target_include_directories(next_start_code_avx2 PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/include)
   add_library(next_start_code_avx512 OBJECT ${CMAKE_CURRENT_SOURCE_DIR}/src/NextStartCodeAVX512.cpp include)
-  set_target_properties(next_start_code_avx512 PROPERTIES COMPILE_FLAGS ${AVX512_CPU_FEATURE} )
+  if(NOT WIN32)
+    set_target_properties(next_start_code_avx512 PROPERTIES COMPILE_FLAGS ${AVX512_CPU_FEATURE} )
+  endif()
   target_include_directories(next_start_code_avx512 PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/include)
   target_link_libraries(${VULKAN_VIDEO_PARSER_LIB} next_start_code_c next_start_code_ssse3 next_start_code_avx2 next_start_code_avx512)
 endif()

diff --git a/vk_video_decoder/libs/NvVideoParser/src/cpudetect.cpp b/vk_video_decoder/libs/NvVideoParser/src/cpudetect.cpp
@@ -33,8 +33,10 @@ class InstructionSet
     static bool AVX(void) { return CPU_Rep.f_1_ECX_[28]; }
     static bool AVX2(void) { return CPU_Rep.f_7_EBX_[5]; }
     static bool AVX512F(void) { return CPU_Rep.f_7_EBX_[16]; }
+    static bool AVX512DQ(void) { return CPU_Rep.f_7_EBX_[17]; }
+    static bool AVX512CD(void) { return CPU_Rep.f_7_EBX_[28]; }
     static bool AVX512BW(void) { return CPU_Rep.f_7_EBX_[30]; }
-    static bool AVX512VL(void) { return CPU_Rep.f_7_EBX_[31]; } // VL isn't required
+    static bool AVX512VL(void) { return CPU_Rep.f_7_EBX_[31]; } // all the 5 bits should be checked if "/arch:AVX512" option is passed
 
 private:
     static const InstructionSet_Internal CPU_Rep;