Fix #36 - Separate ARM and ARM SVE tags and add SVE2 detectors

jfalcou · Feb 1, 2024 · 5bd922c · 5bd922c
1 parent 2aa8a82
commit 5bd922c
Show file tree

Hide file tree

Showing 6 changed files with 114 additions and 55 deletions.
diff --git a/.github/workflows/unit.yml b/.github/workflows/unit.yml
@@ -30,7 +30,7 @@ jobs:
         - { comp: gcc   , arch: x86_osx, mode: Release  }
     steps:
       - name: Fetch current branch
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4.1.1
       - name: Running CMake for ${{ matrix.cfg.comp }} in ${{ matrix.cfg.mode }} mode
         run: |
           mkdir build && cd build
@@ -55,7 +55,7 @@ jobs:
 
     steps:
       - name: Fetch current branch
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4.1.1
       - name: Running CMake for MSVC ${{ matrix.cfg.mode }}
         run: |
           mkdir build && cd build
@@ -80,7 +80,7 @@ jobs:
 
     steps:
       - name: Fetch current branch
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4.1.1
       - name: Running CMake for Clang CL ${{ matrix.cfg.mode }}
         run: |
           mkdir build && cd build
@@ -109,7 +109,7 @@ jobs:
         - { comp: clang , arch: wasm, mode: Release  }
     steps:
       - name: Fetch current branch
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4.1.1
       - name: Running CMake for ${{ matrix.cfg.comp }} in ${{ matrix.cfg.mode }} mode
         run: |
           mkdir build && cd build
@@ -135,7 +135,7 @@ jobs:
         - { comp: clang , arch: x86, mode: Release  }
     steps:
       - name: Fetch current branch
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4.1.1
       - name: Running CMake for ${{ matrix.cfg.comp }} in ${{ matrix.cfg.mode }} mode
         run: |
           mkdir build && cd build
@@ -163,11 +163,13 @@ jobs:
         - { comp: gcc , arch: aarch64     , opts: -Wno-psabi, mode: Release }
         - { comp: gcc , arch: aarch64.sve , opts: -Wno-psabi, mode: Debug   }
         - { comp: gcc , arch: aarch64.sve , opts: -Wno-psabi, mode: Release }
+        - { comp: gcc , arch: aarch64.sve2, opts: -Wno-psabi, mode: Debug   }
+        - { comp: gcc , arch: aarch64.sve2, opts: -Wno-psabi, mode: Release }
         - { comp: gcc , arch: ppc64       , opts: -Wno-psabi, mode: Debug   }
         - { comp: gcc , arch: ppc64       , opts: -Wno-psabi, mode: Release }
     steps:
       - name: Fetch current branch
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4.1.1
       - name: Running CMake for ${{ matrix.cfg.comp }} in ${{ matrix.cfg.mode }} mode
         run:  |
           mkdir build && cd build

diff --git a/include/spy.hpp b/include/spy.hpp
@@ -793,26 +793,32 @@ namespace avx512
 #endif
 }
 }
-#if !defined(SPY_SIMD_DETECTED) && defined(__ARM_FEATURE_SVE)
+#if defined(__ARM_FEATURE_SVE2) || defined(__ARM_FEATURE_SVE)
 # if !defined(__ARM_FEATURE_SVE_BITS) || (__ARM_FEATURE_SVE_BITS == 0)
-#   define SPY_SIMD_IS_ARM_FLEXIBLE_SVE
-#   define SPY_SIMD_DETECTED ::spy::detail::simd_version::sve_
+#   define SPY_SIMD_IS_ARM_FLEXIBLE_SVE_CARDINAL
 # elif defined(__ARM_FEATURE_SVE_BITS)
 #   if(__ARM_FEATURE_SVE_BITS == 128)
-#     define SPY_SIMD_IS_ARM_FIXED_SVE
-#     define SPY_SIMD_DETECTED ::spy::detail::simd_version::fixed_sve_
+#     define SPY_SIMD_IS_ARM_FIXED_SVE_CARDINAL
 #   elif(__ARM_FEATURE_SVE_BITS == 256)
-#     define SPY_SIMD_IS_ARM_FIXED_SVE
-#     define SPY_SIMD_DETECTED ::spy::detail::simd_version::fixed_sve_
+#     define SPY_SIMD_IS_ARM_FIXED_SVE_CARDINAL
 #   elif(__ARM_FEATURE_SVE_BITS == 512)
-#     define SPY_SIMD_IS_ARM_FIXED_SVE
-#     define SPY_SIMD_DETECTED ::spy::detail::simd_version::fixed_sve_
+#     define SPY_SIMD_IS_ARM_FIXED_SVE_CARDINAL
 #   elif(__ARM_FEATURE_SVE_BITS == 1024)
-#     define SPY_SIMD_IS_ARM_FIXED_SVE
-#     define SPY_SIMD_DETECTED ::spy::detail::simd_version::fixed_sve_
+#     define SPY_SIMD_IS_ARM_FIXED_SVE_CARDINAL
+#   else
+#   error "[SPY] - No support for non-power of 2 SVE cardinals"
 #   endif
 # endif
 #endif
+#if !defined(SPY_SIMD_DETECTED) && defined(__ARM_FEATURE_SVE2)
+# define SPY_SIMD_IS_ARM_SVE2
+# define SPY_SIMD_DETECTED ::spy::detail::simd_version::sve2_
+# define SPY_SIMD_VENDOR ::spy::detail::simd_isa::arm_sve_
+#elif !defined(SPY_SIMD_DETECTED) && defined(__ARM_FEATURE_SVE)
+# define SPY_SIMD_IS_ARM_SVE
+# define SPY_SIMD_DETECTED ::spy::detail::simd_version::sve_
+# define SPY_SIMD_VENDOR ::spy::detail::simd_isa::arm_sve_
+#endif
 #if !defined(SPY_SIMD_DETECTED) && defined(__aarch64__)
 #  define SPY_SIMD_IS_ARM_ASIMD
 #  define SPY_SIMD_DETECTED ::spy::detail::simd_version::asimd_
@@ -867,7 +873,7 @@ namespace avx512
 #endif
 namespace spy::detail
 {
-  enum class simd_isa { undefined_ = -1, x86_ = 1000, ppc_ = 2000, arm_ = 3000, wasm_ = 4000 };
+  enum class simd_isa { undefined_ = -1, x86_ = 1000, ppc_ = 2000, arm_ = 3000, arm_sve_ = 3500, wasm_ = 4000 };
   enum class simd_version { undefined_  = -1
                           , sse1_       = 1110, sse2_  = 1120, sse3_ = 1130, ssse3_ = 1131
                           , sse41_      = 1141, sse42_ = 1142
@@ -879,10 +885,12 @@ namespace spy::detail
                           , vsx_        = 3000
                           , vsx_2_06_   = 3206, vsx_2_07_ = 3207, vsx_3_00_ = 3300, vsx_3_01_ = 3301
                           , neon_       = 4001, asimd_    = 4002
-                          , sve_        = 5000, fixed_sve_  = 5100
+                          , sve_        = 5000, sve2_     = 5500
                           , simd128_    = 6000
                           };
-  template<simd_isa InsSetArch = simd_isa::undefined_, simd_version Version = simd_version::undefined_>
+  template< simd_isa InsSetArch  = simd_isa::undefined_
+          , simd_version Version = simd_version::undefined_
+          >
   struct simd_info
   {
     static constexpr auto           isa     = InsSetArch;
@@ -896,7 +904,7 @@ namespace spy::detail
                   )                                                                       return 128;
       else  if constexpr(Version == simd_version::avx_ || Version == simd_version::avx2_) return 256;
       else  if constexpr(Version == simd_version::avx512_     )                           return 512;
-      else  if constexpr(Version == simd_version::fixed_sve_  )
+      else  if constexpr(Version >= simd_version::sve_)
       {
 #if defined(__ARM_FEATURE_SVE_BITS)
         return __ARM_FEATURE_SVE_BITS;
@@ -906,6 +914,7 @@ namespace spy::detail
       }
       else return -1;
     }();
+    static constexpr bool has_fixed_cardinal() { return width != -1; }
     friend std::ostream& operator<<(std::ostream& os, simd_info const&)
     {
             if constexpr ( Version == simd_version::simd128_  ) os << "WASM SIMD128";
@@ -930,10 +939,15 @@ namespace spy::detail
       }
       else  if constexpr ( Version == simd_version::neon_     ) os  << "ARM NEON";
       else  if constexpr ( Version == simd_version::asimd_    ) os  << "ARM ASIMD";
-      else  if constexpr ( Version == simd_version::sve_      ) os  << "ARM SVE (dyn. bits)";
-      else  if constexpr ( Version == simd_version::fixed_sve_) os  << "ARM SVE ("
-                                                                    << simd_info::width
-                                                                    << " bits)";
+      else  if constexpr ( Version >= simd_version::sve_      )
+      {
+        if constexpr ( Version == simd_version::sve2_ ) os  << "ARM SVE2 (";
+        else                                            os  << "ARM SVE  (";
+        constexpr auto fc = has_fixed_cardinal();
+        if constexpr(fc)  os << simd_info::width;
+        else              os << "dyn.";
+        os << " bits)";
+      }
       else return os << "Undefined SIMD instructions set";
       if constexpr (spy::supports::fma_)     os << " (with FMA3 support)";
       if constexpr (spy::supports::fma4_)    os << " (with FMA4 support)";
@@ -1017,11 +1031,13 @@ namespace spy
   constexpr inline auto vsx_3_01_ = ppc_simd_info<detail::simd_version::vsx_3_01_>{};
   template<detail::simd_version V = detail::simd_version::undefined_>
   using arm_simd_info = detail::simd_info<detail::simd_isa::arm_,V>;
+  template<detail::simd_version V = detail::simd_version::undefined_>
+  using sve_simd_info = detail::simd_info<detail::simd_isa::arm_sve_,V>;
   constexpr inline auto arm_simd_   = arm_simd_info<>{};
   constexpr inline auto neon_       = arm_simd_info<detail::simd_version::neon_ >{};
   constexpr inline auto asimd_      = arm_simd_info<detail::simd_version::asimd_>{};
-  constexpr inline auto sve_        = arm_simd_info<detail::simd_version::sve_>{};
-  constexpr inline auto fixed_sve_  = arm_simd_info<detail::simd_version::fixed_sve_>{};
+  constexpr inline auto sve_        = sve_simd_info<detail::simd_version::sve_>{};
+  constexpr inline auto sve2_       = sve_simd_info<detail::simd_version::sve2_>{};
 }
 #include <cstddef>
 namespace spy::detail

diff --git a/src/spy/simd.hpp b/src/spy/simd.hpp
@@ -15,7 +15,7 @@
 
 namespace spy::detail
 {
-  enum class simd_isa { undefined_ = -1, x86_ = 1000, ppc_ = 2000, arm_ = 3000, wasm_ = 4000 };
+  enum class simd_isa { undefined_ = -1, x86_ = 1000, ppc_ = 2000, arm_ = 3000, arm_sve_ = 3500, wasm_ = 4000 };
 
   enum class simd_version { undefined_  = -1
                           , sse1_       = 1110, sse2_  = 1120, sse3_ = 1130, ssse3_ = 1131
@@ -28,15 +28,18 @@ namespace spy::detail
                           , vsx_        = 3000
                           , vsx_2_06_   = 3206, vsx_2_07_ = 3207, vsx_3_00_ = 3300, vsx_3_01_ = 3301
                           , neon_       = 4001, asimd_    = 4002
-                          , sve_        = 5000, fixed_sve_  = 5100
+                          , sve_        = 5000, sve2_     = 5500
                           , simd128_    = 6000
                           };
 
-  template<simd_isa InsSetArch = simd_isa::undefined_, simd_version Version = simd_version::undefined_>
+  template< simd_isa InsSetArch  = simd_isa::undefined_
+          , simd_version Version = simd_version::undefined_
+          >
   struct simd_info
   {
     static constexpr auto           isa     = InsSetArch;
     static constexpr auto           version = Version;
+
     static constexpr std::ptrdiff_t width   = []()
     {
       if constexpr(   Version == simd_version::simd128_
@@ -46,7 +49,7 @@ namespace spy::detail
                   )                                                                       return 128;
       else  if constexpr(Version == simd_version::avx_ || Version == simd_version::avx2_) return 256;
       else  if constexpr(Version == simd_version::avx512_     )                           return 512;
-      else  if constexpr(Version == simd_version::fixed_sve_  )
+      else  if constexpr(Version >= simd_version::sve_)
       {
 #if defined(__ARM_FEATURE_SVE_BITS)
         return __ARM_FEATURE_SVE_BITS;
@@ -57,6 +60,8 @@ namespace spy::detail
       else return -1;
     }();
 
+    static constexpr bool has_fixed_cardinal() { return width != -1; }
+
     friend std::ostream& operator<<(std::ostream& os, simd_info const&)
     {
             if constexpr ( Version == simd_version::simd128_  ) os << "WASM SIMD128";
@@ -81,10 +86,17 @@ namespace spy::detail
       }
       else  if constexpr ( Version == simd_version::neon_     ) os  << "ARM NEON";
       else  if constexpr ( Version == simd_version::asimd_    ) os  << "ARM ASIMD";
-      else  if constexpr ( Version == simd_version::sve_      ) os  << "ARM SVE (dyn. bits)";
-      else  if constexpr ( Version == simd_version::fixed_sve_) os  << "ARM SVE ("
-                                                                    << simd_info::width
-                                                                    << " bits)";
+      else  if constexpr ( Version >= simd_version::sve_      )
+      {
+        if constexpr ( Version == simd_version::sve2_ ) os  << "ARM SVE2 (";
+        else                                            os  << "ARM SVE  (";
+
+        constexpr auto fc = has_fixed_cardinal();
+        if constexpr(fc)  os << simd_info::width;
+        else              os << "dyn.";
+
+        os << " bits)";
+      }
       else return os << "Undefined SIMD instructions set";
 
       if constexpr (spy::supports::fma_)     os << " (with FMA3 support)";
@@ -195,9 +207,12 @@ namespace spy
   template<detail::simd_version V = detail::simd_version::undefined_>
   using arm_simd_info = detail::simd_info<detail::simd_isa::arm_,V>;
 
+  template<detail::simd_version V = detail::simd_version::undefined_>
+  using sve_simd_info = detail::simd_info<detail::simd_isa::arm_sve_,V>;
+
   constexpr inline auto arm_simd_   = arm_simd_info<>{};
   constexpr inline auto neon_       = arm_simd_info<detail::simd_version::neon_ >{};
   constexpr inline auto asimd_      = arm_simd_info<detail::simd_version::asimd_>{};
-  constexpr inline auto sve_        = arm_simd_info<detail::simd_version::sve_>{};
-  constexpr inline auto fixed_sve_  = arm_simd_info<detail::simd_version::fixed_sve_>{};
+  constexpr inline auto sve_        = sve_simd_info<detail::simd_version::sve_>{};
+  constexpr inline auto sve2_       = sve_simd_info<detail::simd_version::sve2_>{};
 }
diff --git a/src/spy/simd/arm.hpp b/src/spy/simd/arm.hpp
@@ -7,29 +7,36 @@
 //==================================================================================================
 #pragma once
 
-#if !defined(SPY_SIMD_DETECTED) && defined(__ARM_FEATURE_SVE)
+#if defined(__ARM_FEATURE_SVE2) || defined(__ARM_FEATURE_SVE)
 // Flexible SVE has no SVE_BITS or SVE_BITS set at 0 and is set via -march=armv8-a+sve
 # if !defined(__ARM_FEATURE_SVE_BITS) || (__ARM_FEATURE_SVE_BITS == 0)
-#   define SPY_SIMD_IS_ARM_FLEXIBLE_SVE
-#   define SPY_SIMD_DETECTED ::spy::detail::simd_version::sve_
+#   define SPY_SIMD_IS_ARM_FLEXIBLE_SVE_CARDINAL
 // Fixed-size SVE has SVE_BITS set at expected size via -msve-vector-bits
 # elif defined(__ARM_FEATURE_SVE_BITS)
 #   if(__ARM_FEATURE_SVE_BITS == 128)
-#     define SPY_SIMD_IS_ARM_FIXED_SVE
-#     define SPY_SIMD_DETECTED ::spy::detail::simd_version::fixed_sve_
+#     define SPY_SIMD_IS_ARM_FIXED_SVE_CARDINAL
 #   elif(__ARM_FEATURE_SVE_BITS == 256)
-#     define SPY_SIMD_IS_ARM_FIXED_SVE
-#     define SPY_SIMD_DETECTED ::spy::detail::simd_version::fixed_sve_
+#     define SPY_SIMD_IS_ARM_FIXED_SVE_CARDINAL
 #   elif(__ARM_FEATURE_SVE_BITS == 512)
-#     define SPY_SIMD_IS_ARM_FIXED_SVE
-#     define SPY_SIMD_DETECTED ::spy::detail::simd_version::fixed_sve_
+#     define SPY_SIMD_IS_ARM_FIXED_SVE_CARDINAL
 #   elif(__ARM_FEATURE_SVE_BITS == 1024)
-#     define SPY_SIMD_IS_ARM_FIXED_SVE
-#     define SPY_SIMD_DETECTED ::spy::detail::simd_version::fixed_sve_
+#     define SPY_SIMD_IS_ARM_FIXED_SVE_CARDINAL
+#   else
+#   error "[SPY] - No support for non-power of 2 SVE cardinals"
 #   endif
 # endif
 #endif
 
+#if !defined(SPY_SIMD_DETECTED) && defined(__ARM_FEATURE_SVE2)
+# define SPY_SIMD_IS_ARM_SVE2
+# define SPY_SIMD_DETECTED ::spy::detail::simd_version::sve2_
+# define SPY_SIMD_VENDOR ::spy::detail::simd_isa::arm_sve_
+#elif !defined(SPY_SIMD_DETECTED) && defined(__ARM_FEATURE_SVE)
+# define SPY_SIMD_IS_ARM_SVE
+# define SPY_SIMD_DETECTED ::spy::detail::simd_version::sve_
+# define SPY_SIMD_VENDOR ::spy::detail::simd_isa::arm_sve_
+#endif
+
 #if !defined(SPY_SIMD_DETECTED) && defined(__aarch64__)
 #  define SPY_SIMD_IS_ARM_ASIMD
 #  define SPY_SIMD_DETECTED ::spy::detail::simd_version::asimd_

diff --git a/test/simd.cpp b/test/simd.cpp
@@ -12,6 +12,7 @@
 int main()
 {
   std::cout << "Supported SIMD instructions set: " << spy::simd_instruction_set << std::endl;
+  std::cout << "Supported SIMD instructions set: " << typeid(spy::simd_instruction_set).name() << std::endl;
   std::cout << "Check that X86 SIMD extension detection is correct: " << std::endl;
   {
     std::cout << "X86 SIMD status: " << std::boolalpha << (spy::simd_instruction_set == spy::x86_simd_) << std::endl;
@@ -50,14 +51,19 @@ int main()
     std::cout << "ARM SIMD status: "  << std::boolalpha << (spy::simd_instruction_set == spy::arm_simd_ ) << std::endl;
     std::cout << "NEON     status: "  << std::boolalpha << (spy::simd_instruction_set >= spy::neon_ )     << std::endl;
     std::cout << "ASIMD    status: "  << std::boolalpha << (spy::simd_instruction_set >= spy::asimd_)     << std::endl;
-    std::cout << "SVE      status: "  << std::boolalpha << (spy::simd_instruction_set >= spy::sve_);
-    if constexpr(spy::simd_instruction_set >= spy::fixed_sve_)
-    {
-      std::cout << " - fixed size: " << spy::simd_instruction_set.width << " bits" << std::endl;
-    }
-    else if constexpr(spy::simd_instruction_set >= spy::sve_)
+
+    std::cout << "SVE      status: "  << std::boolalpha << (spy::simd_instruction_set == spy::sve_ )      << std::endl;
+    std::cout << "SVE2     status: "  << std::boolalpha << (spy::simd_instruction_set == spy::sve2_)      << std::endl;
+    if constexpr(spy::simd_instruction_set >= spy::sve_)
     {
-      std::cout << " - flexible size" << std::endl;
+      if constexpr(spy::simd_instruction_set.has_fixed_cardinal())
+      {
+        std::cout << "SVE uses fixed size: " << spy::simd_instruction_set.width << " bits" << std::endl;
+      }
+      else
+      {
+        std::cout << "SVE uses flexible size" << std::endl;
+      }
     }
   }
   std::cout << std::endl;

diff --git a/test/toolchain/gcc.aarch64.sve2.cmake b/test/toolchain/gcc.aarch64.sve2.cmake
@@ -0,0 +1,13 @@
+##==================================================================================================
+##  SPY - C++ Informations Broker
+##  Copyright : SPY Project Contributors
+##  SPDX-License-Identifier: BSL-1.0
+##==================================================================================================
+set(CMAKE_SYSTEM_NAME Linux)
+set(CMAKE_SYSTEM_PROCESSOR arm)
+
+set(CMAKE_C_COMPILER    aarch64-linux-gnu-gcc-12 )
+set(CMAKE_CXX_COMPILER  aarch64-linux-gnu-g++-12 )
+set(CMAKE_CXX_FLAGS     "-march=armv8-a+sve2 -msve-vector-bits=512" )
+
+set(CMAKE_CROSSCOMPILING_CMD qemu-aarch64)