Fix and simplify SIMD tests

The tests for SSE and AVX must only be done if the correct compiler
flags were used.

Signed-off-by: Stefan Weil <sw@weilnetz.de>

Author: stweil

src/arch/Makefile.am

@@ -19,13 +19,17 @@ noinst_LTLIBRARIES += libtesseract_avx.la libtesseract_avx2.la
 noinst_LTLIBRARIES += libtesseract_sse.la
 noinst_LTLIBRARIES += libtesseract_arch.la
 
+libtesseract_arch_la_CPPFLAGS = $(AM_CPPFLAGS)
 if AVX_OPT
+libtesseract_arch_la_CPPFLAGS += -DAVX
 libtesseract_avx_la_CXXFLAGS = -ffast-math -mavx
 endif
 if AVX2_OPT
+libtesseract_arch_la_CPPFLAGS += -DAVX2
 libtesseract_avx2_la_CXXFLAGS = -ffast-math -mavx2
 endif
 if SSE41_OPT
+libtesseract_arch_la_CPPFLAGS += -DSSE4_1
 libtesseract_sse_la_CXXFLAGS = -ffast-math -msse4.1
 endif
 

src/arch/dotproductavx.cpp

@@ -16,21 +16,7 @@
 // limitations under the License.
 ///////////////////////////////////////////////////////////////////////
 
-#if !defined(__AVX__)
-// Implementation for non-avx archs.
-
-#include "dotproductavx.h"
-#include <cstdio>
-#include <cstdlib>
-
-namespace tesseract {
-double DotProductAVX(const double* u, const double* v, int n) {
-  fprintf(stderr, "DotProductAVX can't be used on Android\n");
-  abort();
-}
-}  // namespace tesseract
-
-#else  // !defined(__AVX__)
+#if defined(__AVX__)
 // Implementation for avx capable archs.
 #include <immintrin.h>
 #include <cstdint>
@@ -111,4 +97,4 @@ double DotProductAVX(const double* u, const double* v, int n) {
 
 }  // namespace tesseract.
 
-#endif  // ANDROID_BUILD
+#endif  // __AVX__

src/arch/dotproductsse.cpp

@@ -16,26 +16,7 @@
 // limitations under the License.
 ///////////////////////////////////////////////////////////////////////
 
-#if !defined(__SSE4_1__)
-// This code can't compile with "-msse4.1", so use dummy stubs.
-
-#include "dotproductsse.h"
-#include <cstdio>
-#include <cstdlib>
-
-namespace tesseract {
-double DotProductSSE(const double* u, const double* v, int n) {
-  fprintf(stderr, "DotProductSSE can't be used on Android\n");
-  abort();
-}
-int32_t IntDotProductSSE(const int8_t* u, const int8_t* v, int n) {
-  fprintf(stderr, "IntDotProductSSE can't be used on Android\n");
-  abort();
-}
-}  // namespace tesseract
-
-#else  // !defined(__SSE4_1__)
-// Non-Android code here
+#if defined(__SSE4_1__)
 
 #include <emmintrin.h>
 #include <smmintrin.h>
@@ -137,4 +118,4 @@ int32_t IntDotProductSSE(const int8_t* u, const int8_t* v, int n) {
 
 }  // namespace tesseract.
 
-#endif  // ANDROID_BUILD
+#endif  // __SSE4_1__

src/arch/simddetect.cpp

@@ -22,19 +22,10 @@
 #include "params.h"   // for STRING_VAR
 #include "tprintf.h"  // for tprintf
 
-#undef X86_BUILD
-#if defined(__x86_64__) || defined(__i386__) || defined(_WIN32)
-#  if !defined(ANDROID_BUILD)
-#    define X86_BUILD 1
-#  endif  // !ANDROID_BUILD
-#endif    // x86 target
-
-#if defined(X86_BUILD)
-#  if defined(__GNUC__)
-#    include <cpuid.h>
-#  elif defined(_WIN32)
-#    include <intrin.h>
-#  endif
+#if defined(__GNUC__)
+# include <cpuid.h>
+#elif defined(_WIN32)
+# include <intrin.h>
 #endif
 
 namespace tesseract {
@@ -84,13 +75,15 @@ SIMDDetect::SIMDDetect() {
   // The fallback is a generic dot product calculation.
   SetDotProduct(DotProductGeneric);
 
-#if defined(X86_BUILD)
-#  if defined(__GNUC__)
+#if defined(__GNUC__)
   unsigned int eax, ebx, ecx, edx;
   if (__get_cpuid(1, &eax, &ebx, &ecx, &edx) != 0) {
     // Note that these tests all use hex because the older compilers don't have
     // the newer flags.
+#if defined(SSE4_1)
     sse_available_ = (ecx & 0x00080000) != 0;
+#endif
+#if defined(AVX)
     avx_available_ = (ecx & 0x10000000) != 0;
     if (avx_available_) {
       // There is supposed to be a __get_cpuid_count function, but this is all
@@ -101,30 +94,38 @@ SIMDDetect::SIMDDetect() {
       avx512F_available_ = (ebx & 0x00010000) != 0;
       avx512BW_available_ = (ebx & 0x40000000) != 0;
     }
+#endif
   }
 #  elif defined(_WIN32)
   int cpuInfo[4];
   __cpuid(cpuInfo, 0);
   if (cpuInfo[0] >= 1) {
     __cpuid(cpuInfo, 1);
+#if defined(SSE4_1)
     sse_available_ = (cpuInfo[2] & 0x00080000) != 0;
+#endif
+#if defined(AVX)
     avx_available_ = (cpuInfo[2] & 0x10000000) != 0;
+#endif
   }
-#  else
-#    error "I don't know how to test for SIMD with this compiler"
-#  endif
-#endif  // X86_BUILD
+#else
+#error "I don't know how to test for SIMD with this compiler"
+#endif
 
-#if defined(X86_BUILD)
   // Select code for calculation of dot product based on autodetection.
-  if (avx_available_) {
+  if (false) {
+    // This is a dummy to support conditional compilation.
+#if defined(AVX)
+  } else if (avx_available_) {
     // AVX detected.
     SetDotProduct(DotProductAVX);
+#endif
+#if defined(SSE4_1)
   } else if (sse_available_) {
     // SSE detected.
     SetDotProduct(DotProductSSE);
+#endif
   }
-#endif  // X86_BUILD
 }
 
 void SIMDDetect::Update() {
@@ -141,26 +142,29 @@ void SIMDDetect::Update() {
     // Native optimized code selected by config variable.
     SetDotProduct(DotProductNative);
     dotproduct_method = "native";
-  }
-#if defined(X86_BUILD)
-  else if (!strcmp(dotproduct.string(), "avx")) {
+#if defined(AVX)
+  } else if (!strcmp(dotproduct.string(), "avx")) {
     // AVX selected by config variable.
     SetDotProduct(DotProductAVX);
     dotproduct_method = "avx";
+#endif
+#if defined(SSE4_1)
   } else if (!strcmp(dotproduct.string(), "sse")) {
     // SSE selected by config variable.
     SetDotProduct(DotProductSSE);
     dotproduct_method = "sse";
-  }
-#endif  // X86_BUILD
-  else {
+#endif
+  } else {
     // Unsupported value of config variable.
     tprintf("Warning, ignoring unsupported config variable value: dotproduct=%s\n",
             dotproduct.string());
     tprintf("Support values for dotproduct: auto generic native"
-#if defined(X86_BUILD)
-            " avx sse"
-#endif  // X86_BUILD
+#if defined(AVX)
+            " avx"
+#endif
+#if defined(SSE4_1)
+            " sse"
+#endif
             ".\n");
   }