Android: Enable armeabi-v7a builds

CMakeLists.txt

@@ -196,6 +196,15 @@ elseif(${CMAKE_SYSTEM_PROCESSOR} STREQUAL "aarch64")
   set(CPU_ARCH "aarch64")
 elseif(${CMAKE_SYSTEM_PROCESSOR} STREQUAL "arm" OR ${CMAKE_SYSTEM_PROCESSOR} STREQUAL "armv7-a" OR ${CMAKE_SYSTEM_PROCESSOR} STREQUAL "armv7l")
   set(CPU_ARCH "aarch32")
+  if(ANDROID)
+    # Force ARM mode, since apparently ANDROID_ARM_MODE isn't working..
+    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -marm")
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -marm")
+  else()
+    # Enable NEON.
+    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -marm -march=armv7-a+simd -mfpu=neon-vfpv3")
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -marm -march=armv7-a+simd -mfpu=neon-vfpv3")
+  endif()
 else()
   message(FATAL_ERROR "Unknown system processor: " ${CMAKE_SYSTEM_PROCESSOR})
 endif()

android/app/build.gradle

@@ -30,8 +30,8 @@ android {
     defaultConfig {
         externalNativeBuild {
             cmake {
-                arguments "-DCMAKE_BUILD_TYPE=RelWithDebInfo"
-                abiFilters "arm64-v8a"
+                arguments "-DCMAKE_BUILD_TYPE=Release -DANDROID_ARM_NEON=ON"
+                abiFilters "arm64-v8a", "armeabi-v7a"
             }
         }
     }

src/core/gpu_sw.cpp

@@ -11,6 +11,8 @@ Log_SetChannel(GPU_SW);
 
 #if defined(CPU_X64)
 #include <emmintrin.h>
+#elif defined(CPU_AARCH32)
+#include <arm_neon.h>
 #elif defined(CPU_AARCH64)
 #ifdef _MSC_VER
 #include <arm64_neon.h>
@@ -155,7 +157,7 @@ ALWAYS_INLINE void CopyOutRow16<HostDisplayPixelFormat::RGBA5551, u16>(const u16
     _mm_storeu_si128(reinterpret_cast<__m128i*>(dst_ptr), value);
     dst_ptr += 8;
   }
-#elif defined(CPU_AARCH64)
+#elif defined(CPU_AARCH32) || defined(CPU_AARCH64)
   const u32 aligned_width = Common::AlignDownPow2(width, 8);
   for (; col < aligned_width; col += 8)
   {
@@ -195,7 +197,7 @@ ALWAYS_INLINE void CopyOutRow16<HostDisplayPixelFormat::RGB565, u16>(const u16*
     _mm_storeu_si128(reinterpret_cast<__m128i*>(dst_ptr), value);
     dst_ptr += 8;
   }
-#elif defined(CPU_AARCH64)
+#elif defined(CPU_AARCH32) || defined(CPU_AARCH64)
   const u32 aligned_width = Common::AlignDownPow2(width, 8);
   const uint16x8_t single_mask = vdupq_n_u16(0x1F);
   for (; col < aligned_width; col += 8)