PR12034.patch

From bb1c2d40fa123e41165e407cda8bcb7027b0d41e Mon Sep 17 00:00:00 2001
From: croniccorey <[email protected]>
Date: Thu, 25 May 2017 03:58:40 -0400
Subject: [PATCH] [Kodi] Update NEON support

- Define HAS_NEON when NEON cmake switch is set
- Aarch64 always has neon, Enable cmake switch and change CPUInfo to report it
- Don't set NEON_FLAGS if none exist, Not all targets need extra flags
- Don't use preprocessor to check for neon support, Instead use HAS_NEON set by cmake
---
 cmake/scripts/android/ArchSetup.cmake                       | 1 +
 cmake/scripts/common/ArchSetup.cmake                        | 5 ++++-
 cmake/scripts/ios/ArchSetup.cmake                           | 2 +-
 cmake/scripts/linux/ArchSetup.cmake                         | 2 +-
 xbmc/cores/VideoPlayer/VideoRenderers/LinuxRendererGLES.cpp | 2 +-
 xbmc/cores/VideoPlayer/VideoRenderers/yuv2rgb.neon.S        | 2 +-
 xbmc/cores/VideoPlayer/VideoRenderers/yuv2rgb.neon.h        | 2 +-
 xbmc/guilib/MatrixGLES.cpp                                  | 6 +++---
 xbmc/utils/CPUInfo.cpp                                      | 8 ++++++--
 9 files changed, 19 insertions(+), 11 deletions(-)

diff --git a/cmake/scripts/android/ArchSetup.cmake b/cmake/scripts/android/ArchSetup.cmake
index fe7d4c15835f..17215a7aad80 100644
--- a/cmake/scripts/android/ArchSetup.cmake
+++ b/cmake/scripts/android/ArchSetup.cmake
@@ -17,6 +17,7 @@ else()
     set(NEON_FLAGS "-mfpu=neon -mvectorize-with-neon-quad")
   elseif(CPU STREQUAL arm64-v8a)
     set(ARCH aarch64)
+    set(NEON True)
   elseif(CPU STREQUAL i686)
     set(ARCH i486-linux)
     set(NEON False)
diff --git a/cmake/scripts/common/ArchSetup.cmake b/cmake/scripts/common/ArchSetup.cmake
index 74afc8ccaf7c..8d5dba8bbeed 100644
--- a/cmake/scripts/common/ArchSetup.cmake
+++ b/cmake/scripts/common/ArchSetup.cmake
@@ -146,7 +146,10 @@ if(NOT DEFINED NEON OR NEON)
   option(ENABLE_NEON "Enable NEON optimization" ${NEON})
   if(ENABLE_NEON)
     message(STATUS "NEON optimization enabled")
-    add_options(ALL_LANGUAGES ALL_BUILDS ${NEON_FLAGS})
+    add_definitions(-DHAS_NEON)
+    if(NEON_FLAGS)
+      add_options(ALL_LANGUAGES ALL_BUILDS ${NEON_FLAGS})
+    endif()
   endif()
 endif()
 
diff --git a/cmake/scripts/ios/ArchSetup.cmake b/cmake/scripts/ios/ArchSetup.cmake
index dbb7c4858eef..20d8cd20eefb 100644
--- a/cmake/scripts/ios/ArchSetup.cmake
+++ b/cmake/scripts/ios/ArchSetup.cmake
@@ -15,7 +15,7 @@ else()
   if(CPU STREQUAL armv7 OR CPU STREQUAL arm64)
     set(CMAKE_OSX_ARCHITECTURES ${CPU})
     set(ARCH arm-osx)
-    set(NEON False)
+    set(NEON True)
   else()
     message(SEND_ERROR "Unknown CPU: ${CPU}")
   endif()
diff --git a/cmake/scripts/linux/ArchSetup.cmake b/cmake/scripts/linux/ArchSetup.cmake
index eeed920aff57..53e5f1bd1204 100644
--- a/cmake/scripts/linux/ArchSetup.cmake
+++ b/cmake/scripts/linux/ArchSetup.cmake
@@ -19,7 +19,7 @@ else()
     set(NEON_FLAGS "-mfpu=neon -mvectorize-with-neon-quad")
   elseif(CPU MATCHES aarch64 OR CPU MATCHES arm64)
     set(ARCH aarch64)
-    set(NEON False)
+    set(NEON True)
   else()
     message(SEND_ERROR "Unknown CPU: ${CPU}")
   endif()
diff --git a/xbmc/cores/VideoPlayer/VideoRenderers/LinuxRendererGLES.cpp b/xbmc/cores/VideoPlayer/VideoRenderers/LinuxRendererGLES.cpp
index 2d5e603093ff..4681dff16107 100644
--- a/xbmc/cores/VideoPlayer/VideoRenderers/LinuxRendererGLES.cpp
+++ b/xbmc/cores/VideoPlayer/VideoRenderers/LinuxRendererGLES.cpp
@@ -50,7 +50,7 @@ extern "C" {
 #include "libswscale/swscale.h"
 }
 
-#if defined(__ARM_NEON__) && !defined(__LP64__)
+#if defined(HAS_NEON) && !defined(__LP64__)
 #include "yuv2rgb.neon.h"
 #include "utils/CPUInfo.h"
 #endif
diff --git a/xbmc/cores/VideoPlayer/VideoRenderers/yuv2rgb.neon.S b/xbmc/cores/VideoPlayer/VideoRenderers/yuv2rgb.neon.S
index 48bbe378ae8a..df7218b4fab0 100644
--- a/xbmc/cores/VideoPlayer/VideoRenderers/yuv2rgb.neon.S
+++ b/xbmc/cores/VideoPlayer/VideoRenderers/yuv2rgb.neon.S
@@ -7,7 +7,7 @@
 //
 //
 //
-#if defined(__ARM_NEON__) && !defined(__LP64__)
+#if defined(HAS_NEON) && !defined(__LP64__)
 
  /* Initial ARM Neon implementation of core YUV2RGB functions. */
         
diff --git a/xbmc/cores/VideoPlayer/VideoRenderers/yuv2rgb.neon.h b/xbmc/cores/VideoPlayer/VideoRenderers/yuv2rgb.neon.h
index 6aeb368478a7..145f61e3704a 100644
--- a/xbmc/cores/VideoPlayer/VideoRenderers/yuv2rgb.neon.h
+++ b/xbmc/cores/VideoPlayer/VideoRenderers/yuv2rgb.neon.h
@@ -22,7 +22,7 @@
 extern "C" {
 #endif
 
-#if defined(__ARM_NEON__) && !defined(__LP64__)
+#if defined(HAS_NEON) && !defined(__LP64__)
   void yuv420_2_rgb8888_neon
   (
     uint8_t *dst_ptr,
diff --git a/xbmc/guilib/MatrixGLES.cpp b/xbmc/guilib/MatrixGLES.cpp
index dbc8081d48de..0c73a56191c4 100644
--- a/xbmc/guilib/MatrixGLES.cpp
+++ b/xbmc/guilib/MatrixGLES.cpp
@@ -27,7 +27,7 @@
 #include <cmath>
 #include "MatrixGLES.h"
 #include "utils/log.h"
-#if defined(__ARM_NEON__)
+#if defined(HAS_NEON)
 #include "utils/CPUInfo.h"
 #endif
 
@@ -139,7 +139,7 @@ void CMatrixGL::Rotatef(GLfloat angle, GLfloat x, GLfloat y, GLfloat z)
   MultMatrixf(matrix);
 }
 
-#if defined(__ARM_NEON__) && !defined(__LP64__)
+#if defined(HAS_NEON) && !defined(__LP64__)
   
 inline void Matrix4Mul(const float* src_mat_1, const float* src_mat_2, float* dst_mat)
 {
@@ -182,7 +182,7 @@ inline void Matrix4Mul(const float* src_mat_1, const float* src_mat_2, float* ds
 #endif
 void CMatrixGL::MultMatrixf(const GLfloat *matrix)
 {
-#if defined(__ARM_NEON__) && !defined(__LP64__)
+#if defined(HAS_NEON) && !defined(__LP64__)
     if ((g_cpuInfo.GetCPUFeatures() & CPU_FEATURE_NEON) == CPU_FEATURE_NEON)
     {
       GLfloat m[16];
diff --git a/xbmc/utils/CPUInfo.cpp b/xbmc/utils/CPUInfo.cpp
index 5e2ebbd18d85..ff3a188b12de 100644
--- a/xbmc/utils/CPUInfo.cpp
+++ b/xbmc/utils/CPUInfo.cpp
@@ -43,7 +43,7 @@
 #include <sys/resource.h>
 #endif
 
-#if defined(TARGET_LINUX) && defined(__ARM_NEON__) && !defined(TARGET_ANDROID)
+#if defined(TARGET_LINUX) && defined(HAS_NEON) && !defined(TARGET_ANDROID)
 #include <fcntl.h>
 #include <unistd.h>
 #include <elf.h>
@@ -957,7 +957,10 @@ bool CCPUInfo::HasNeon()
 #elif defined(TARGET_DARWIN_IOS)
   has_neon = 1;
 
-#elif defined(TARGET_LINUX) && defined(__ARM_NEON__)
+#elif defined(TARGET_LINUX) && defined(HAS_NEON)
+#if defined(__LP64__)
+  has_neon = 1;
+#else
   if (has_neon == -1)
   {
     has_neon = 0;
@@ -978,6 +981,7 @@ bool CCPUInfo::HasNeon()
       close(fd);
     }
   }
+#endif
 
 #endif