Merge branch 'armv7_neon' of https://github.com/mastertheknife/ZoneMinder into storageareas

CMakeLists.txt

@@ -66,12 +66,23 @@ set(CMAKE_CXX_FLAGS_DEBUG "-Wall -D__STDC_CONSTANT_MACROS -g")
 set(CMAKE_INCLUDE_CURRENT_DIR ON)
 set(CMAKE_MODULE_PATH ${CMAKE_MODULE_PATH} "${CMAKE_SOURCE_DIR}/cmake/Modules/")
 
-IF(${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm")
-add_definitions (-mfpu=neon)
-ELSE(${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm")
-
-ENDIF(${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm")
-
+# GCC below 6.0 doesn't support __target__("fpu=neon") attribute, required for compiling ARM Neon code, otherwise compilation fails.
+# Must use -mfpu=neon compiler flag instead, but only do that for processors that support neon, otherwise strip the neon code alltogether,
+# because passing -fmpu=neon is unsafe to processors that don't support neon
+IF(CMAKE_SYSTEM_PROCESSOR MATCHES "^arm" AND CMAKE_SYSTEM_NAME MATCHES "Linux")
+	IF(CMAKE_COMPILER_IS_GNUCXX AND CMAKE_CXX_COMPILER_VERSION VERSION_LESS 6.0)
+		EXEC_PROGRAM(grep ARGS " neon " "/proc/cpuinfo" OUTPUT_VARIABLE neonoutput RETURN_VALUE neonresult)
+		IF(neonresult EQUAL 0)
+			set(CMAKE_C_FLAGS_RELEASE "${CMAKE_C_FLAGS_RELEASE} -mfpu=neon")
+			set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} -mfpu=neon")
+			set(CMAKE_C_FLAGS_DEBUG "${CMAKE_C_FLAGS_DEBUG} -mfpu=neon")
+			set(CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG} -mfpu=neon")
+		ELSE(neonresult EQUAL 0)
+			add_definitions(-DZM_STRIP_NEON=1)
+			message(STATUS "ARM Neon is not available on this processor. Neon functions will be absent")
+		ENDIF(neonresult EQUAL 0)
+	ENDIF(CMAKE_COMPILER_IS_GNUCXX AND CMAKE_CXX_COMPILER_VERSION VERSION_LESS 6.0)
+ENDIF(CMAKE_SYSTEM_PROCESSOR MATCHES "^arm" AND CMAKE_SYSTEM_NAME MATCHES "Linux")
 
 # Modules that we need:
 include (GNUInstallDirs)

src/zm_image.cpp

@@ -3291,11 +3291,11 @@ __attribute__((noinline)) void std_fastblend(const uint8_t* col1, const uint8_t*
 }
 
 /* FastBlend Neon for AArch32 */
-#if defined(__arm__)
+#if (defined(__arm__) && !defined(ZM_STRIP_NEON))
 __attribute__((noinline,__target__("fpu=neon")))
 #endif
 void neon32_armv7_fastblend(const uint8_t* col1, const uint8_t* col2, uint8_t* result, unsigned long count, double blendpercent) {
-#if defined(__arm__)
+#if (defined(__arm__) && !defined(ZM_STRIP_NEON))
   static int8_t divider = 0;
   static double current_blendpercent = 0.0;
 
@@ -3348,7 +3348,7 @@ void neon32_armv7_fastblend(const uint8_t* col1, const uint8_t* col2, uint8_t* r
   : "%r12", "%q0", "%q1", "%q2", "%q3", "cc", "memory"
   );
 #else
-  Panic("Neon function called on a non ARM platform");
+  Panic("Neon function called on a non-ARM platform or Neon code is absent");
 #endif
 }
 
@@ -3575,11 +3575,11 @@ __attribute__((noinline)) void std_delta8_abgr(const uint8_t* col1, const uint8_
 }
 
 /* Grayscale Neon for AArch32 */
-#if defined(__arm__)
+#if (defined(__arm__) && !defined(ZM_STRIP_NEON))
 __attribute__((noinline,__target__("fpu=neon")))
 #endif
 void neon32_armv7_delta8_gray8(const uint8_t* col1, const uint8_t* col2, uint8_t* result, unsigned long count) {
-#if defined(__arm__)
+#if (defined(__arm__) && !defined(ZM_STRIP_NEON))
 
   /* Q0(D0,D1) = col1 */
   /* Q1(D2,D3) = col2 */
@@ -3597,16 +3597,16 @@ void neon32_armv7_delta8_gray8(const uint8_t* col1, const uint8_t* col2, uint8_t
   : "%q0", "%q1", "cc", "memory"
   );
 #else
-  Panic("Neon function called on a non ARM platform");
+  Panic("Neon function called on a non-ARM platform or Neon code is absent");
 #endif
 }
 
 /* RGB32 Neon for AArch32 */
-#if defined(__arm__)
+#if (defined(__arm__) && !defined(ZM_STRIP_NEON))
 __attribute__((noinline,__target__("fpu=neon")))
 #endif
 void neon32_armv7_delta8_rgb32(const uint8_t* col1, const uint8_t* col2, uint8_t* result, unsigned long count, uint32_t multiplier) {
-#if defined(__arm__)
+#if (defined(__arm__) && !defined(ZM_STRIP_NEON))
 
   /* Q0(D0,D1) = col1 */
   /* Q1(D2,D3) = col2 */
@@ -3632,7 +3632,7 @@ void neon32_armv7_delta8_rgb32(const uint8_t* col1, const uint8_t* col2, uint8_t
   : "%r12", "%q0", "%q1", "%q2", "cc", "memory"
   );
 #else
-  Panic("Neon function called on a non ARM platform");
+  Panic("Neon function called on a non-ARM platform or Neon code is absent");
 #endif
 }
 

src/zm_utils.cpp

@@ -250,30 +250,50 @@ void hwcaps_detect() {
   /* x86 or x86-64 processor */
   uint32_t r_edx, r_ecx, r_ebx;
 
+#ifdef __x86_64__
   __asm__ __volatile__(
+  "push %%rbx\n\t"
   "mov $0x0,%%ecx\n\t"
   "mov $0x7,%%eax\n\t"
   "cpuid\n\t"
-#ifdef __x86_64__
   "push %%rbx\n\t"
-#else
-  "push %%ebx\n\t"
-#endif
   "mov $0x1,%%eax\n\t"
   "cpuid\n\t"
-#ifdef __x86_64__
+  "pop %%rax\n\t"
   "pop %%rbx\n\t"
-#else
-  "pop %%ebx\n\t"
-#endif
-  : "=d" (r_edx), "=c" (r_ecx), "=b" (r_ebx)
+  : "=d" (r_edx), "=c" (r_ecx), "=a" (r_ebx)
+  :
   :
-  : "%eax"
   );
-	
+#else
+  __asm__ __volatile__(
+  "push %%ebx\n\t"
+  "mov $0x0,%%ecx\n\t"
+  "mov $0x7,%%eax\n\t"
+  "cpuid\n\t"
+  "push %%ebx\n\t"
+  "mov $0x1,%%eax\n\t"
+  "cpuid\n\t"
+  "pop %%eax\n\t"
+  "pop %%ebx\n\t"
+  : "=d" (r_edx), "=c" (r_ecx), "=a" (r_ebx)
+  :
+  :
+  );
+#endif
+
   if (r_ebx & 0x00000020) {
     sseversion = 52; /* AVX2 */
     Debug(1,"Detected a x86\\x86-64 processor with AVX2");
+  } else if (r_ecx & 0x10000000) {
+    sseversion = 51; /* AVX */
+    Debug(1,"Detected a x86\\x86-64 processor with AVX");
+  } else if (r_ecx & 0x00100000) {
+    sseversion = 42; /* SSE4.2 */
+    Debug(1,"Detected a x86\\x86-64 processor with SSE4.2");
+  } else if (r_ecx & 0x00080000) {
+    sseversion = 41; /* SSE4.1 */
+    Debug(1,"Detected a x86\\x86-64 processor with SSE4.1");
   } else if (r_ecx & 0x00000200) {
     sseversion = 35; /* SSSE3 */
     Debug(1,"Detected a x86\\x86-64 processor with SSSE3");