Merge pull request #73 from Stefan-Olt/master

aarch64 build fix and asm optimization (only autotools)

Makefile.am

@@ -66,6 +66,10 @@ libmvtools_la_LDFLAGS = -no-undefined -avoid-version $(PLUGINLDFLAGS)
 
 libmvtools_la_LIBADD = $(FFTW3F_LIBS)
 
+if MVTOOLS_ARM
+libmvtools_la_SOURCES += src/asm/aarch64-pixel-a.S
+endif
+
 if MVTOOLS_X86
 libmvtools_la_SOURCES += src/asm/const-a.asm \
 						 src/asm/cpu-a.asm \

configure.ac

@@ -10,6 +10,7 @@ LT_INIT([disable-static win32-dll])
 
 AC_PROG_CC
 AC_PROG_CXX
+AM_PROG_AS
 
 AC_CANONICAL_HOST
 
@@ -54,7 +55,7 @@ AS_CASE(
   [i?86],         [BITS="32" NASMFLAGS="$NASMFLAGS -DARCH_X86_64=0" X86="true"],
   [x86_64|amd64], [BITS="64" NASMFLAGS="$NASMFLAGS -DARCH_X86_64=1 -DPIC" X86="true"],
   [powerpc*],     [PPC="true"],
-  [arm*],         [ARM="true"],
+  [arm*|aarch*],  [ARM="true"],
   [AC_MSG_ERROR([Unknown host CPU: $host_cpu.])]
 )
 
@@ -63,6 +64,7 @@ AS_CASE(
    [darwin*],
    [
       NASMFLAGS="$NASMFLAGS -f macho$BITS -DPREFIX"
+      CPPFLAGS="$CPPFLAGS -DPREFIX"
    ],
    [*linux*|gnu*|dragonfly*|*bsd*],
    [

src/CPU.c

@@ -191,7 +191,13 @@ uint32_t cpu_detect(void) {
     return cpu;
 }
 
-#else // not MVTOOLS_X86
+#elif defined(MVTOOLS_ARM)
+
+uint32_t cpu_detect(void) {
+    return ~0; // we just assume NEON is available, as there is no instruction to check
+}
+
+#else // not MVTOOLS_X86 or MVTOOLS_ARM
 
 uint32_t cpu_detect(void) {
     return 0;

src/CPU.h

@@ -57,6 +57,9 @@ enum {
 #ifdef MVTOOLS_X86
     MVOPT_SSE2 = 1,
     MVOPT_AVX2 = 2,
+#elif MVTOOLS_ARM
+    MVOPT_NEON = 1,
+    MVOPT_SSE2 = 1, // SSE2 is converted to Neon
 #endif // MVTOOLS_X86
 };
 

src/DCTFFTW.cpp

@@ -54,9 +54,13 @@ static void Float2Pixels_C(const DCTFFTW *dct, uint8_t *dstp8, int dst_pitch, fl
 }
 
 
-#if defined(MVTOOLS_X86)
+#if defined(MVTOOLS_X86) || defined(MVTOOLS_ARM)
 
+#if defined(MVTOOLS_ARM)
+#include "sse2neon.h"
+#else
 #include <emmintrin.h>
+#endif
 
 template <typename PixelType>
 static void Float2Pixels_SSE2(const DCTFFTW *dct, uint8_t *dstp8, int dst_pitch, float *realdata) {
@@ -153,7 +157,7 @@ void dctInit(DCTFFTW *dct, int sizex, int sizey, int bitsPerSample, int opt) {
         dct->Float2Pixels = Float2Pixels_C<uint16_t>;
 
     if (opt) {
-#if defined(MVTOOLS_X86)
+#if defined(MVTOOLS_X86) || defined(MVTOOLS_ARM)
         if (bitsPerSample == 8)
             dct->Float2Pixels = Float2Pixels_SSE2<uint8_t>;
         else

src/Luma.cpp

@@ -25,9 +25,13 @@ unsigned int luma_c(const uint8_t *pSrc8, intptr_t nSrcPitch) {
 }
 
 
-#if defined(MVTOOLS_X86)
+#if defined(MVTOOLS_X86) || defined(MVTOOLS_ARM)
 
+#if defined(MVTOOLS_ARM)
+#include "sse2neon.h"
+#else
 #include <emmintrin.h>
+#endif
 
 
 #define zeroes _mm_setzero_si128()
@@ -69,7 +73,7 @@ unsigned int luma_sse2(const uint8_t *pSrc, intptr_t nSrcPitch) {
 // opt can fit in four bits, if the width and height need more than eight bits each.
 #define KEY(width, height, bits, opt) (unsigned)(width) << 24 | (height) << 16 | (bits) << 8 | (opt)
 
-#if defined(MVTOOLS_X86)
+#if defined(MVTOOLS_X86) || defined(MVTOOLS_ARM)
 #define LUMA_SSE2(width, height) \
     { KEY(width, height, 8, SSE2), luma_sse2<width, height> },
 #else
@@ -110,7 +114,7 @@ static const std::unordered_map<uint32_t, LUMAFunction> luma_functions = {
 LUMAFunction selectLumaFunction(unsigned width, unsigned height, unsigned bits, int opt) {
     LUMAFunction luma = luma_functions.at(KEY(width, height, bits, Scalar));
 
-#if defined(MVTOOLS_X86)
+#if defined(MVTOOLS_X86) || defined(MVTOOLS_ARM)
     if (opt) {
         try {
             luma = luma_functions.at(KEY(width, height, bits, SSE2));

src/MVDegrains.cpp

@@ -365,7 +365,7 @@ static void VS_CC mvdegrainFree(void *instanceData, VSCore *core, const VSAPI *v
 // opt can fit in four bits, if the width and height need more than eight bits each.
 #define KEY(width, height, bits, opt) (unsigned)(width) << 24 | (height) << 16 | (bits) << 8 | (opt)
 
-#if defined(MVTOOLS_X86)
+#if defined(MVTOOLS_X86) || defined(MVTOOLS_ARM)
 #define DEGRAIN_SSE2(radius, width, height) \
     { KEY(width, height, 8, MVOPT_SSE2), Degrain_sse2<radius, width, height> },
 
@@ -458,17 +458,18 @@ static const std::unordered_map<uint32_t, DenoiseFunction> degrain_functions_sse
 static DenoiseFunction selectDegrainFunction(unsigned radius, unsigned width, unsigned height, unsigned bits, int opt) {
     DenoiseFunction degrain = degrain_functions[radius - 1].at(KEY(width, height, bits, MVOPT_SCALAR));
 
-#if defined(MVTOOLS_X86)
+#if defined(MVTOOLS_X86) || defined(MVTOOLS_ARM)
     if (opt) {
         try {
             degrain = degrain_functions_sse2[radius - 1].at(KEY(width, height, bits, MVOPT_SSE2));
         } catch (std::out_of_range &) { }
-
+#if defined(MVTOOLS_X86)
         if (g_cpuinfo & X264_CPU_AVX2) {
             DenoiseFunction tmp = selectDegrainFunctionAVX2(radius, width, height, bits);
             if (tmp)
                 degrain = tmp;
         }
+#endif
     }
 #endif
 
@@ -496,7 +497,7 @@ static void selectFunctions(MVDegrainData *d) {
 
         d->ToPixels = ToPixels_uint16_t_uint8_t;
 
-#if defined(MVTOOLS_X86)
+#if defined(MVTOOLS_X86) || defined(MVTOOLS_ARM)
         if (d->opt) {
             d->LimitChanges = LimitChanges_sse2;
         }

src/MVDegrains.h

@@ -53,10 +53,16 @@ static void Degrain_C(uint8_t * __restrict pDst8, int nDstPitch, const uint8_t *
 }
 
 
-#if defined(MVTOOLS_X86)
+#if defined(MVTOOLS_X86) || defined(MVTOOLS_ARM)
 
+#if defined(MVTOOLS_ARM)
+#include "sse2neon.h"
+#else
 #include <emmintrin.h>
 
+DenoiseFunction selectDegrainFunctionAVX2(unsigned radius, unsigned width, unsigned height, unsigned bits);
+#endif
+
 // XXX Moves the pointers passed in pRefs. This is okay because they are not
 // used after this function is done with them.
 template <int radius, int blockWidth, int blockHeight>
@@ -133,8 +139,6 @@ static void Degrain_sse2(uint8_t *pDst, int nDstPitch, const uint8_t *pSrc, int
     }
 }
 
-DenoiseFunction selectDegrainFunctionAVX2(unsigned radius, unsigned width, unsigned height, unsigned bits);
-
 static void LimitChanges_sse2(uint8_t *pDst, intptr_t nDstPitch, const uint8_t *pSrc, intptr_t nSrcPitch, intptr_t nWidth, intptr_t nHeight, intptr_t nLimit) {
     __m128i bytes_limit = _mm_set1_epi8(nLimit);
 

src/MVFrame.cpp

@@ -32,11 +32,7 @@
 #define min(a, b) (((a) < (b)) ? (a) : (b))
 #endif
 
-#if defined(MVTOOLS_X86)
-
-#include <emmintrin.h>
-
-#define zeroes _mm_setzero_si128()
+#if defined(MVTOOLS_X86) || defined(MVTOOLS_ARM)
 
 /* TODO: port these
    extern "C" void  VerticalBicubic_iSSE(uint8_t *pDst, const uint8_t *pSrc, intptr_t nDstPitch,
@@ -49,6 +45,12 @@
    extern "C" void  RB2FilteredHorizontalInplaceLine_SSE(uint8_t *pSrc, intptr_t nWidthMMX);
    */
 
+#if defined(MVTOOLS_ARM)
+#include "sse2neon.h"
+#else
+
+#include <emmintrin.h>
+
 void Average2_avx2(uint8_t *pDst, const uint8_t *pSrc1, const uint8_t *pSrc2, intptr_t nPitch, intptr_t nWidth, intptr_t nHeight);
 void VerticalBilinear_avx2(uint8_t *pDst, const uint8_t *pSrc, intptr_t nPitch,
                            intptr_t nWidth, intptr_t nHeight, intptr_t bitsPerSample);
@@ -61,6 +63,12 @@ void VerticalWiener_avx2(uint8_t *pDst, const uint8_t *pSrc, intptr_t nPitch,
 void HorizontalWiener_avx2(uint8_t *pDst, const uint8_t *pSrc, intptr_t nPitch,
                            intptr_t nWidth, intptr_t nHeight, intptr_t bitsPerSample) ;
 
+#endif
+
+
+
+#define zeroes _mm_setzero_si128()
+
 static void Average2_sse2(uint8_t *pDst, const uint8_t *pSrc1, const uint8_t *pSrc2, intptr_t nPitch, intptr_t nWidth, intptr_t nHeight) {
     for (int y = 0; y < nHeight; y++) {
         for (int x = 0; x < nWidth; x += 16) {
@@ -707,7 +715,7 @@ static void RB2BilinearFilteredVertical(uint8_t *pDst8, const uint8_t *pSrc8, in
         int xstart = 0;
 
         if (sizeof(PixelType) == 1 && opt && nWidthMMX >= 8) {
-#if defined(MVTOOLS_X86)
+#if defined(MVTOOLS_X86) || defined(MVTOOLS_ARM)
             RB2BilinearFilteredVerticalLine_sse2((uint8_t *)pDst, (const uint8_t *)pSrc, nSrcPitch, nWidthMMX);
             xstart = nWidthMMX;
 #endif
@@ -745,7 +753,7 @@ static void RB2BilinearFilteredHorizontalInplace(uint8_t *pSrc8, int nSrcPitch,
         int xstart = 1;
 
         if (sizeof(PixelType) == 1 && opt) {
-#if defined(MVTOOLS_X86)
+#if defined(MVTOOLS_X86) || defined(MVTOOLS_ARM)
             RB2BilinearFilteredHorizontalInplaceLine_sse2((uint8_t *)pSrc, nWidthMMX); /* very first is skipped */
             xstart = nWidthMMX;
 #endif
@@ -797,7 +805,7 @@ static void RB2QuadraticVertical(uint8_t *pDst8, const uint8_t *pSrc8, int nDstP
         int xstart = 0;
 
         if (sizeof(PixelType) == 1 && opt && nWidthMMX >= 8) {
-#if defined(MVTOOLS_X86)
+#if defined(MVTOOLS_X86) || defined(MVTOOLS_ARM)
             RB2QuadraticVerticalLine_sse2((uint8_t *)pDst, (const uint8_t *)pSrc, nSrcPitch, nWidthMMX);
             xstart = nWidthMMX;
 #endif
@@ -848,7 +856,7 @@ static void RB2QuadraticHorizontalInplace(uint8_t *pSrc8, int nSrcPitch, int nWi
         int xstart = 1;
 
         if (sizeof(PixelType) == 1 && opt) {
-#if defined(MVTOOLS_X86)
+#if defined(MVTOOLS_X86) || defined(MVTOOLS_ARM)
             RB2QuadraticHorizontalInplaceLine_sse2((uint8_t *)pSrc, nWidthMMX);
             xstart = nWidthMMX;
 #endif
@@ -913,7 +921,7 @@ static void RB2CubicVertical(uint8_t *pDst8, const uint8_t *pSrc8, int nDstPitch
         int xstart = 0;
 
         if (sizeof(PixelType) == 1 && opt && nWidthMMX >= 8) {
-#if defined(MVTOOLS_X86)
+#if defined(MVTOOLS_X86) || defined(MVTOOLS_ARM)
             RB2CubicVerticalLine_sse2((uint8_t *)pDst, (const uint8_t *)pSrc, nSrcPitch, nWidthMMX);
             xstart = nWidthMMX;
 #endif
@@ -964,7 +972,7 @@ static void RB2CubicHorizontalInplace(uint8_t *pSrc8, int nSrcPitch, int nWidth,
         int xstart = 1;
 
         if (sizeof(PixelType) == 1 && opt) {
-#if defined(MVTOOLS_X86)
+#if defined(MVTOOLS_X86) || defined(MVTOOLS_ARM)
             RB2CubicHorizontalInplaceLine_sse2((uint8_t *)pSrc, nWidthMMX);
             xstart = nWidthMMX;
 #endif
@@ -1395,16 +1403,17 @@ void mvpRefine(MVPlane *mvp, int sharp) {
             refine[2] = DiagonalBilinear<uint8_t>;
 
             if (mvp->opt) {
-#if defined(MVTOOLS_X86)
+#if defined(MVTOOLS_X86) || defined(MVTOOLS_ARM)
                 refine[0] = HorizontalBilinear_sse2;
                 refine[1] = VerticalBilinear_sse2;
                 refine[2] = DiagonalBilinear_sse2;
-
+#if defined(MVTOOLS_X86)
                 if (g_cpuinfo & X264_CPU_AVX2) {
                     refine[0] = HorizontalBilinear_avx2;
                     refine[1] = VerticalBilinear_avx2;
                     refine[2] = DiagonalBilinear_avx2;
                 }
+#endif
 #endif
             }
         } else {
@@ -1434,14 +1443,15 @@ void mvpRefine(MVPlane *mvp, int sharp) {
             refine[1] = VerticalWiener<uint8_t>;
 
             if (mvp->opt) {
-#if defined(MVTOOLS_X86)
+#if defined(MVTOOLS_X86) || defined(MVTOOLS_ARM)
                 refine[0] = refine[2] = HorizontalWiener_sse2;
                 refine[1] = VerticalWiener_sse2;
-
+#if defined(MVTOOLS_X86)
                 if (g_cpuinfo & X264_CPU_AVX2) {
                     refine[0] = refine[2] = HorizontalWiener_avx2;
                     refine[1] = VerticalWiener_avx2;
                 }
+#endif
 #endif
             }
         } else {
@@ -1485,11 +1495,12 @@ void mvpRefine(MVPlane *mvp, int sharp) {
             avg = Average2<uint8_t>;
 
             if (mvp->opt) {
-#if defined(MVTOOLS_X86)
+#if defined(MVTOOLS_X86) || defined(MVTOOLS_ARM)
                 avg = Average2_sse2;
-
+#if defined(MVTOOLS_X86)
                 if (g_cpuinfo & X264_CPU_AVX2)
                     avg = Average2_avx2;
+#endif
 #endif
             }
         } else {

src/Overlap.cpp

@@ -158,9 +158,13 @@ void overlaps_c(uint8_t *pDst8, intptr_t nDstPitch, const uint8_t *pSrc8, intptr
 }
 
 
-#if defined(MVTOOLS_X86)
+#if defined(MVTOOLS_X86) || defined(MVTOOLS_ARM)
 
+#if defined(MVTOOLS_ARM)
+#include "sse2neon.h"
+#else
 #include <emmintrin.h>
+#endif
 
 
 #define zeroes _mm_setzero_si128()
@@ -236,7 +240,7 @@ struct OverlapsWrapper<4, blockHeight> {
 // opt can fit in four bits, if the width and height need more than eight bits each.
 #define KEY(width, height, bits, opt) (unsigned)(width) << 24 | (height) << 16 | (bits) << 8 | (opt)
 
-#if defined(MVTOOLS_X86)
+#if defined(MVTOOLS_X86) || defined(MVTOOLS_ARM)
 #define OVERS_SSE2(width, height) \
     { KEY(width, height, 8, MVOPT_SSE2), OverlapsWrapper<width, height>::overlaps_sse2 },
 #else
@@ -305,17 +309,18 @@ static const std::unordered_map<uint32_t, OverlapsFunction> overlaps_functions =
 OverlapsFunction selectOverlapsFunction(unsigned width, unsigned height, unsigned bits, int opt) {
     OverlapsFunction overs = overlaps_functions.at(KEY(width, height, bits, MVOPT_SCALAR));
 
-#if defined(MVTOOLS_X86)
+#if defined(MVTOOLS_X86) || defined(MVTOOLS_ARM)
     if (opt) {
         try {
             overs = overlaps_functions.at(KEY(width, height, bits, MVOPT_SSE2));
         } catch (std::out_of_range &) { }
-
+#ifdef MVTOOLS_X86
         if (g_cpuinfo & X264_CPU_AVX2) {
             OverlapsFunction tmp = selectOverlapsFunctionAVX2(width, height, bits);
             if (tmp)
                 overs = tmp;
         }
+#endif
     }
 #endif