Use sse2neon.h in gcore/rasterio functionnality

rouault · Mar 16, 2024 · 30f3844 · 30f3844
1 parent 75626d6
commit 30f3844
Show file tree

Hide file tree

Showing 6 changed files with 57 additions and 27 deletions.
diff --git a/gcore/CMakeLists.txt b/gcore/CMakeLists.txt
@@ -97,7 +97,10 @@ if (NOT GDAL_AUTOLOAD_PLUGINS)
     PROPERTY COMPILE_DEFINITIONS GDAL_NO_AUTOLOAD)
 endif ()
 
-if (HAVE_SSSE3_AT_COMPILE_TIME)
+if (GDAL_ENABLE_ARM_NEON_OPTIMIZATIONS)
+  target_compile_definitions(gcore PRIVATE -DHAVE_SSSE3_AT_COMPILE_TIME -DUSE_NEON_OPTIMIZATIONS)
+  target_sources(gcore PRIVATE rasterio_ssse3.cpp)
+elseif (HAVE_SSSE3_AT_COMPILE_TIME)
   target_compile_definitions(gcore PRIVATE -DHAVE_SSSE3_AT_COMPILE_TIME)
   target_sources(gcore PRIVATE rasterio_ssse3.cpp)
   set_property(

diff --git a/gcore/gdal_priv_templates.hpp b/gcore/gdal_priv_templates.hpp
@@ -513,9 +513,14 @@ inline void GDALCopy8Words(const Tin *pValueIn, Tout *const pValueOut)
 }
 
 // Needs SSE2
-#if defined(__x86_64) || defined(_M_X64) || defined(USE_SSE2)
+#if defined(__x86_64) || defined(_M_X64) || defined(USE_SSE2) ||               \
+    defined(USE_NEON_OPTIMIZATIONS)
 
+#ifdef USE_NEON_OPTIMIZATIONS
+#include "sse2neon.h"
+#else
 #include <emmintrin.h>
+#endif
 
 static inline void GDALCopyXMMToInt32(const __m128i xmm, void *pDest)
 {

diff --git a/gcore/rasterio.cpp b/gcore/rasterio.cpp
@@ -57,6 +57,18 @@
 #include "memdataset.h"
 #include "vrtdataset.h"
 
+#if defined(__x86_64) || defined(_M_X64)
+#include <emmintrin.h>
+#define HAVE_SSE2
+#elif defined(USE_NEON_OPTIMIZATIONS)
+#include "sse2neon.h"
+#define HAVE_SSE2
+#endif
+
+#ifdef HAVE_SSSE3_AT_COMPILE_TIME
+#include "rasterio_ssse3.h"
+#endif
+
 static void GDALFastCopyByte(const GByte *CPL_RESTRICT pSrcData,
                              int nSrcPixelStride, GByte *CPL_RESTRICT pDstData,
                              int nDstPixelStride, GPtrDiff_t nWordCount);
@@ -2192,9 +2204,7 @@ static void inline GDALCopyWordsT_8atatime(
     }
 }
 
-#if defined(__x86_64) || defined(_M_X64)
-
-#include <emmintrin.h>
+#ifdef HAVE_SSE2
 
 template <class Tout>
 void GDALCopyWordsByteTo16Bit(const GByte *const CPL_RESTRICT pSrcData,
@@ -2605,7 +2615,7 @@ void GDALCopyWordsT(const double *const CPL_RESTRICT pSrcData,
                             nDstPixelStride, nWordCount);
 }
 
-#endif  // defined(__x86_64) || defined(_M_X64)
+#endif  // HAVE_SSE2
 
 template <>
 void GDALCopyWordsT(const float *const CPL_RESTRICT pSrcData,
@@ -3043,13 +3053,7 @@ static inline void GDALUnrolledCopy(T *CPL_RESTRICT pDest,
     GDALUnrolledCopyGeneric<T, srcStride, dstStride>(pDest, pSrc, nIters);
 }
 
-#if (defined(__x86_64) || defined(_M_X64))
-
-#ifdef HAVE_SSSE3_AT_COMPILE_TIME
-
-#include "rasterio_ssse3.h"
-
-#endif
+#ifdef HAVE_SSE2
 
 template <>
 void GDALUnrolledCopy<GByte, 2, 1>(GByte *CPL_RESTRICT pDest,
@@ -3150,7 +3154,7 @@ void GDALUnrolledCopy<GByte, 4, 1>(GByte *CPL_RESTRICT pDest,
         pSrc += 4;
     }
 }
-#endif  // defined(__x86_64) || defined(_M_X64)
+#endif  // HAVE_SSE2
 
 /************************************************************************/
 /*                         GDALFastCopy()                               */
@@ -5256,13 +5260,7 @@ bool GDALBufferHasOnlyNoData(const void *pBuffer, double dfNoDataValue,
     return false;
 }
 
-#if defined(__x86_64) || defined(_M_X64)
-
-#include <emmintrin.h>
-
-#ifdef HAVE_SSSE3_AT_COMPILE_TIME
-#include "rasterio_ssse3.h"
-#endif
+#ifdef HAVE_SSE2
 
 /************************************************************************/
 /*                    GDALDeinterleave3Byte()                           */
@@ -5276,6 +5274,12 @@ GDALDeinterleave3Byte(const GByte *CPL_RESTRICT pabySrc,
                       GByte *CPL_RESTRICT pabyDest0,
                       GByte *CPL_RESTRICT pabyDest1,
                       GByte *CPL_RESTRICT pabyDest2, size_t nIters)
+#ifdef USE_NEON_OPTIMIZATIONS
+{
+    return GDALDeinterleave3Byte_SSSE3(pabySrc, pabyDest0, pabyDest1, pabyDest2,
+                                       nIters);
+}
+#else
 {
 #ifdef HAVE_SSSE3_AT_COMPILE_TIME
     if (CPLHaveRuntimeSSSE3())
@@ -5323,6 +5327,7 @@ GDALDeinterleave3Byte(const GByte *CPL_RESTRICT pabySrc,
         pabyDest2[i] = pabySrc[3 * i + 2];
     }
 }
+#endif
 
 /************************************************************************/
 /*                    GDALDeinterleave4Byte()                           */
@@ -5378,6 +5383,12 @@ static void GDALDeinterleave4Byte(const GByte *CPL_RESTRICT pabySrc,
                                   GByte *CPL_RESTRICT pabyDest1,
                                   GByte *CPL_RESTRICT pabyDest2,
                                   GByte *CPL_RESTRICT pabyDest3, size_t nIters)
+#ifdef USE_NEON_OPTIMIZATIONS
+{
+    return GDALDeinterleave4Byte_SSSE3(pabySrc, pabyDest0, pabyDest1, pabyDest2,
+                                       pabyDest3, nIters);
+}
+#else
 {
 #ifdef HAVE_SSSE3_AT_COMPILE_TIME
     if (CPLHaveRuntimeSSSE3())
@@ -5426,6 +5437,7 @@ static void GDALDeinterleave4Byte(const GByte *CPL_RESTRICT pabySrc,
         pabyDest3[i] = pabySrc[4 * i + 3];
     }
 }
+#endif
 #else
 // GCC autovectorizer does an excellent job
 __attribute__((optimize("tree-vectorize"))) static void GDALDeinterleave4Byte(
@@ -5553,8 +5565,7 @@ void GDALDeinterleave(const void *pSourceBuffer, GDALDataType eSourceDT,
         }
 #if ((defined(__GNUC__) && !defined(__clang__)) ||                             \
      defined(__INTEL_CLANG_COMPILER)) &&                                       \
-    (defined(__x86_64) || defined(_M_X64)) &&                                  \
-    defined(HAVE_SSSE3_AT_COMPILE_TIME)
+    defined(HAVE_SSE2) && defined(HAVE_SSSE3_AT_COMPILE_TIME)
         else if ((eSourceDT == GDT_Int16 || eSourceDT == GDT_UInt16) &&
                  CPLHaveRuntimeSSSE3())
         {

diff --git a/gcore/rasterio_ssse3.cpp b/gcore/rasterio_ssse3.cpp
@@ -28,12 +28,18 @@
 
 #include "cpl_port.h"
 
-#if defined(HAVE_SSSE3_AT_COMPILE_TIME) &&                                     \
-    (defined(__x86_64) || defined(_M_X64))
+#if (defined(HAVE_SSSE3_AT_COMPILE_TIME) &&                                    \
+     (defined(__x86_64) || defined(_M_X64))) ||                                \
+    defined(USE_NEON_OPTIMIZATIONS)
 
 #include "rasterio_ssse3.h"
 
+#ifdef USE_NEON_OPTIMIZATIONS
+#include "sse2neon.h"
+#else
 #include <tmmintrin.h>
+#endif
+
 #include "gdal_priv_templates.hpp"
 
 void GDALUnrolledCopy_GByte_3_1_SSSE3(GByte *CPL_RESTRICT pDest,

diff --git a/gcore/rasterio_ssse3.h b/gcore/rasterio_ssse3.h
@@ -32,7 +32,7 @@
 #include "cpl_port.h"
 
 #if defined(HAVE_SSSE3_AT_COMPILE_TIME) &&                                     \
-    (defined(__x86_64) || defined(_M_X64))
+    (defined(__x86_64) || defined(_M_X64) || defined(USE_NEON_OPTIMIZATIONS))
 
 void GDALUnrolledCopy_GByte_3_1_SSSE3(GByte *CPL_RESTRICT pDest,
                                       const GByte *CPL_RESTRICT pSrc,

diff --git a/port/cpl_cpu_features.h b/port/cpl_cpu_features.h
@@ -46,7 +46,12 @@ bool CPLHaveRuntimeSSE();
 #endif
 #endif
 
-#ifdef HAVE_SSSE3_AT_COMPILE_TIME
+#ifdef USE_NEON_OPTIMIZATIONS
+static bool inline CPLHaveRuntimeSSSE3()
+{
+    return true;
+}
+#elif defined(HAVE_SSSE3_AT_COMPILE_TIME)
 #if __SSSE3__
 #define HAVE_INLINE_SSSE3
 static bool inline CPLHaveRuntimeSSSE3()