build: add armv7 support

Author: ebraraktas

CMakeLists.txt

@@ -208,10 +208,10 @@ macro(ct2_compile_kernels_for_isa isa flag)
   list(APPEND SOURCES ${CMAKE_CURRENT_BINARY_DIR}/kernels_${isa}.cc)
 endmacro()
 
-if(CMAKE_SYSTEM_PROCESSOR MATCHES "(arm64)|(aarch64)"
+if(CMAKE_SYSTEM_PROCESSOR MATCHES "(arm64)|(aarch64)|(armv7-a)"
     OR (APPLE AND CMAKE_OSX_ARCHITECTURES STREQUAL "arm64"))
-  add_definitions(-DCT2_ARM64_BUILD)
-  set(CT2_BUILD_ARCH "arm64")
+  add_definitions(-DCT2_ARM_BUILD)
+  set(CT2_BUILD_ARCH "arm")
 elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "(x86_64)|(amd64)|(AMD64)")
   add_definitions(-DCT2_X86_BUILD)
   set(CT2_BUILD_ARCH "x86_64")
@@ -240,7 +240,7 @@ if(ENABLE_CPU_DISPATCH)
       ct2_compile_kernels_for_isa(avx2 "-mavx2 -mfma")
       ct2_compile_kernels_for_isa(avx512 "-mavx512f -mavx512cd -mavx512vl -mavx512bw -mavx512dq")
     endif()
-  elseif(CT2_BUILD_ARCH STREQUAL "arm64")
+  elseif(CT2_BUILD_ARCH STREQUAL "arm")
     ct2_compile_kernels_for_isa(neon "-DUSE_NEON")
   endif()
 endif()

src/cpu/cpu_info.cc

@@ -42,7 +42,7 @@ namespace ctranslate2 {
   }
 }
 
-#elif defined(CT2_ARM64_BUILD)
+#elif defined(CT2_ARM_BUILD)
 
 namespace ctranslate2 {
   namespace cpu {

src/cpu/cpu_info.h

@@ -12,7 +12,7 @@ namespace ctranslate2 {
     bool cpu_supports_avx();
     bool cpu_supports_avx2();
     bool cpu_supports_avx512();
-#elif defined(CT2_ARM64_BUILD)
+#elif defined(CT2_ARM_BUILD)
     bool cpu_supports_neon();
 #endif
 

src/cpu/cpu_isa.cc

@@ -32,7 +32,7 @@ namespace ctranslate2 {
         return "AVX2";
       case CpuIsa::AVX512:
         return "AVX512";
-#elif defined(CT2_ARM64_BUILD)
+#elif defined(CT2_ARM_BUILD)
       case CpuIsa::NEON:
         return "NEON";
 #endif
@@ -51,7 +51,7 @@ namespace ctranslate2 {
           return try_isa(env_isa, CpuIsa::AVX2, cpu_supports_avx2());
         if (env_isa == "AVX")
           return try_isa(env_isa, CpuIsa::AVX, cpu_supports_avx());
-#elif defined(CT2_ARM64_BUILD)
+#elif defined(CT2_ARM_BUILD)
         if (env_isa == "NEON")
           return try_isa(env_isa, CpuIsa::NEON, cpu_supports_neon());
 #endif
@@ -68,7 +68,7 @@ namespace ctranslate2 {
         return CpuIsa::AVX2;
       if (cpu_supports_avx())
         return CpuIsa::AVX;
-#  elif defined(CT2_ARM64_BUILD)
+#  elif defined(CT2_ARM_BUILD)
       if (cpu_supports_neon())
         return CpuIsa::NEON;
 #  endif

src/cpu/cpu_isa.h

@@ -11,7 +11,7 @@ namespace ctranslate2 {
       AVX,
       AVX2,
       AVX512,
-#elif defined(CT2_ARM64_BUILD)
+#elif defined(CT2_ARM_BUILD)
       NEON,
 #endif
     };
@@ -48,7 +48,7 @@ namespace ctranslate2 {
     CPU_ISA_CASE(cpu::CpuIsa::AVX, SINGLE_ARG(STMTS))         \
     CPU_ISA_DEFAULT(cpu::CpuIsa::GENERIC, SINGLE_ARG(STMTS))  \
   }
-#elif defined(CT2_ARM64_BUILD)
+#elif defined(CT2_ARM_BUILD)
 #  define CPU_ISA_DISPATCH(STMTS)                             \
   switch (cpu::get_cpu_isa()) {                               \
     CPU_ISA_CASE(cpu::CpuIsa::NEON, SINGLE_ARG(STMTS))        \

src/cpu/vec_neon.h

@@ -144,19 +144,35 @@ namespace ctranslate2 {
       }
 
       static inline value_type div(value_type a, value_type b) {
+#ifdef __aarch64__
         return vdivq_f32(a, b);
+#else
+        return a / b;
+#endif
       }
 
       static inline value_type mul_add(value_type a, value_type b, value_type c) {
+#ifdef __aarch64__
         return vfmaq_f32(c, a, b);
+#else
+        return a * b + c;
+#endif
       }
 
       static inline float reduce_add(value_type a) {
+#ifdef __aarch64__
         return vaddvq_f32(a);
+#else
+        return a[0] + a[1] + a[2] + a[3];
+#endif
       }
 
       static inline float reduce_max(value_type a) {
+#ifdef __aarch64__
         return vmaxvq_f32(a);
+#else
+        return std::max({a[0], a[1], a[2], a[3]});
+#endif
       }
 
     };

src/utils.cc

@@ -38,7 +38,7 @@ namespace ctranslate2 {
                  cpu::cpu_supports_avx(),
                  cpu::cpu_supports_avx2(),
                  cpu::cpu_supports_avx512());
-#elif defined(CT2_ARM64_BUILD)
+#elif defined(CT2_ARM_BUILD)
     spdlog::info("CPU: {} (NEON={})",
                  cpu::cpu_vendor(),
                  cpu::cpu_supports_neon());