laag-libpng 1.6.34.0

data/vendor/git.code.sf.net/p/libpng/code/intel/filter_sse2_intrinsics.c

@@ -0,0 +1,406 @@
+
+/* filter_sse2_intrinsics.c - SSE2 optimized filter functions
+ *
+ * Copyright (c) 2016-2017 Glenn Randers-Pehrson
+ * Written by Mike Klein and Matt Sarett
+ * Derived from arm/filter_neon_intrinsics.c
+ *
+ * Last changed in libpng 1.6.31 [July 27, 2017]
+ *
+ * This code is released under the libpng license.
+ * For conditions of distribution and use, see the disclaimer
+ * and license in png.h
+ */
+
+#include "../pngpriv.h"
+
+#ifdef PNG_READ_SUPPORTED
+
+#if PNG_INTEL_SSE_IMPLEMENTATION > 0
+
+#include <immintrin.h>
+
+/* Functions in this file look at most 3 pixels (a,b,c) to predict the 4th (d).
+ * They're positioned like this:
+ *    prev:  c b
+ *    row:   a d
+ * The Sub filter predicts d=a, Avg d=(a+b)/2, and Paeth predicts d to be
+ * whichever of a, b, or c is closest to p=a+b-c.
+ */
+
+static __m128i load4(const void* p) {
+   return _mm_cvtsi32_si128(*(const int*)p);
+}
+
+static void store4(void* p, __m128i v) {
+   *(int*)p = _mm_cvtsi128_si32(v);
+}
+
+static __m128i load3(const void* p) {
+   /* We'll load 2 bytes, then 1 byte,
+    * then mask them together, and finally load into SSE.
+    */
+   const png_uint_16* p01 = (png_const_uint_16p)p;
+   const png_byte*    p2  = (const png_byte*)(p01+1);
+
+   png_uint_32 v012 = (png_uint_32)(*p01)
+                    | (png_uint_32)(*p2) << 16;
+   return load4(&v012);
+}
+
+static void store3(void* p, __m128i v) {
+   /* We'll pull from SSE as a 32-bit int, then write
+    * its bottom two bytes, then its third byte.
+    */
+   png_uint_32 v012;
+   png_uint_16* p01;
+   png_byte*    p2;
+
+   store4(&v012, v);
+
+   p01 = (png_uint_16p)p;
+   p2  = (png_byte*)(p01+1);
+   *p01 = (png_uint_16)v012;
+   *p2  = (png_byte)(v012 >> 16);
+}
+
+void png_read_filter_row_sub3_sse2(png_row_infop row_info, png_bytep row,
+   png_const_bytep prev)
+{
+   /* The Sub filter predicts each pixel as the previous pixel, a.
+    * There is no pixel to the left of the first pixel.  It's encoded directly.
+    * That works with our main loop if we just say that left pixel was zero.
+    */
+   png_size_t rb;
+
+   __m128i a, d = _mm_setzero_si128();
+
+   png_debug(1, "in png_read_filter_row_sub3_sse2");
+
+   rb = row_info->rowbytes;
+   while (rb >= 4) {
+      a = d; d = load4(row);
+      d = _mm_add_epi8(d, a);
+      store3(row, d);
+
+      row += 3;
+      rb  -= 3;
+   }
+   if (rb > 0) {
+      a = d; d = load3(row);
+      d = _mm_add_epi8(d, a);
+      store3(row, d);
+
+      row += 3;
+      rb  -= 3;
+   }
+   PNG_UNUSED(prev)
+}
+
+void png_read_filter_row_sub4_sse2(png_row_infop row_info, png_bytep row,
+   png_const_bytep prev)
+{
+   /* The Sub filter predicts each pixel as the previous pixel, a.
+    * There is no pixel to the left of the first pixel.  It's encoded directly.
+    * That works with our main loop if we just say that left pixel was zero.
+    */
+   png_size_t rb;
+
+   __m128i a, d = _mm_setzero_si128();
+
+   png_debug(1, "in png_read_filter_row_sub4_sse2");
+
+   rb = row_info->rowbytes+4;
+   while (rb > 4) {
+      a = d; d = load4(row);
+      d = _mm_add_epi8(d, a);
+      store4(row, d);
+
+      row += 4;
+      rb  -= 4;
+   }
+   PNG_UNUSED(prev)
+}
+
+void png_read_filter_row_avg3_sse2(png_row_infop row_info, png_bytep row,
+   png_const_bytep prev)
+{
+   /* The Avg filter predicts each pixel as the (truncated) average of a and b.
+    * There's no pixel to the left of the first pixel.  Luckily, it's
+    * predicted to be half of the pixel above it.  So again, this works
+    * perfectly with our loop if we make sure a starts at zero.
+    */
+
+   png_size_t rb;
+
+   const __m128i zero = _mm_setzero_si128();
+
+   __m128i    b;
+   __m128i a, d = zero;
+
+   png_debug(1, "in png_read_filter_row_avg3_sse2");
+   rb = row_info->rowbytes;
+   while (rb >= 4) {
+      __m128i avg;
+             b = load4(prev);
+      a = d; d = load4(row );
+
+      /* PNG requires a truncating average, so we can't just use _mm_avg_epu8 */
+      avg = _mm_avg_epu8(a,b);
+      /* ...but we can fix it up by subtracting off 1 if it rounded up. */
+      avg = _mm_sub_epi8(avg, _mm_and_si128(_mm_xor_si128(a,b),
+                                            _mm_set1_epi8(1)));
+      d = _mm_add_epi8(d, avg);
+      store3(row, d);
+
+      prev += 3;
+      row  += 3;
+      rb   -= 3;
+   }
+   if (rb > 0) {
+      __m128i avg;
+             b = load3(prev);
+      a = d; d = load3(row );
+
+      /* PNG requires a truncating average, so we can't just use _mm_avg_epu8 */
+      avg = _mm_avg_epu8(a,b);
+      /* ...but we can fix it up by subtracting off 1 if it rounded up. */
+      avg = _mm_sub_epi8(avg, _mm_and_si128(_mm_xor_si128(a,b),
+                                            _mm_set1_epi8(1)));
+
+      d = _mm_add_epi8(d, avg);
+      store3(row, d);
+
+      prev += 3;
+      row  += 3;
+      rb   -= 3;
+   }
+}
+
+void png_read_filter_row_avg4_sse2(png_row_infop row_info, png_bytep row,
+   png_const_bytep prev)
+{
+   /* The Avg filter predicts each pixel as the (truncated) average of a and b.
+    * There's no pixel to the left of the first pixel.  Luckily, it's
+    * predicted to be half of the pixel above it.  So again, this works
+    * perfectly with our loop if we make sure a starts at zero.
+    */
+   png_size_t rb;
+   const __m128i zero = _mm_setzero_si128();
+   __m128i    b;
+   __m128i a, d = zero;
+
+   png_debug(1, "in png_read_filter_row_avg4_sse2");
+
+   rb = row_info->rowbytes+4;
+   while (rb > 4) {
+      __m128i avg;
+             b = load4(prev);
+      a = d; d = load4(row );
+
+      /* PNG requires a truncating average, so we can't just use _mm_avg_epu8 */
+      avg = _mm_avg_epu8(a,b);
+      /* ...but we can fix it up by subtracting off 1 if it rounded up. */
+      avg = _mm_sub_epi8(avg, _mm_and_si128(_mm_xor_si128(a,b),
+                                            _mm_set1_epi8(1)));
+
+      d = _mm_add_epi8(d, avg);
+      store4(row, d);
+
+      prev += 4;
+      row  += 4;
+      rb   -= 4;
+   }
+}
+
+/* Returns |x| for 16-bit lanes. */
+static __m128i abs_i16(__m128i x) {
+#if PNG_INTEL_SSE_IMPLEMENTATION >= 2
+   return _mm_abs_epi16(x);
+#else
+   /* Read this all as, return x<0 ? -x : x.
+   * To negate two's complement, you flip all the bits then add 1.
+    */
+   __m128i is_negative = _mm_cmplt_epi16(x, _mm_setzero_si128());
+
+   /* Flip negative lanes. */
+   x = _mm_xor_si128(x, is_negative);
+
+   /* +1 to negative lanes, else +0. */
+   x = _mm_sub_epi16(x, is_negative);
+   return x;
+#endif
+}
+
+/* Bytewise c ? t : e. */
+static __m128i if_then_else(__m128i c, __m128i t, __m128i e) {
+#if PNG_INTEL_SSE_IMPLEMENTATION >= 3
+   return _mm_blendv_epi8(e,t,c);
+#else
+   return _mm_or_si128(_mm_and_si128(c, t), _mm_andnot_si128(c, e));
+#endif
+}
+
+void png_read_filter_row_paeth3_sse2(png_row_infop row_info, png_bytep row,
+   png_const_bytep prev)
+{
+   /* Paeth tries to predict pixel d using the pixel to the left of it, a,
+    * and two pixels from the previous row, b and c:
+    *   prev: c b
+    *   row:  a d
+    * The Paeth function predicts d to be whichever of a, b, or c is nearest to
+    * p=a+b-c.
+    *
+    * The first pixel has no left context, and so uses an Up filter, p = b.
+    * This works naturally with our main loop's p = a+b-c if we force a and c
+    * to zero.
+    * Here we zero b and d, which become c and a respectively at the start of
+    * the loop.
+    */
+   png_size_t rb;
+   const __m128i zero = _mm_setzero_si128();
+   __m128i c, b = zero,
+           a, d = zero;
+
+   png_debug(1, "in png_read_filter_row_paeth3_sse2");
+
+   rb = row_info->rowbytes;
+   while (rb >= 4) {
+      /* It's easiest to do this math (particularly, deal with pc) with 16-bit
+       * intermediates.
+       */
+      __m128i pa,pb,pc,smallest,nearest;
+      c = b; b = _mm_unpacklo_epi8(load4(prev), zero);
+      a = d; d = _mm_unpacklo_epi8(load4(row ), zero);
+
+      /* (p-a) == (a+b-c - a) == (b-c) */
+   
+      pa = _mm_sub_epi16(b,c);
+
+      /* (p-b) == (a+b-c - b) == (a-c) */
+      pb = _mm_sub_epi16(a,c);
+
+      /* (p-c) == (a+b-c - c) == (a+b-c-c) == (b-c)+(a-c) */
+      pc = _mm_add_epi16(pa,pb);
+
+      pa = abs_i16(pa);  /* |p-a| */
+      pb = abs_i16(pb);  /* |p-b| */
+      pc = abs_i16(pc);  /* |p-c| */
+
+      smallest = _mm_min_epi16(pc, _mm_min_epi16(pa, pb));
+
+      /* Paeth breaks ties favoring a over b over c. */
+      nearest  = if_then_else(_mm_cmpeq_epi16(smallest, pa), a,
+                 if_then_else(_mm_cmpeq_epi16(smallest, pb), b,
+                                                             c));
+
+      /* Note `_epi8`: we need addition to wrap modulo 255. */
+      d = _mm_add_epi8(d, nearest);
+      store3(row, _mm_packus_epi16(d,d));
+
+      prev += 3;
+      row  += 3;
+      rb   -= 3;
+   }
+   if (rb > 0) {
+      /* It's easiest to do this math (particularly, deal with pc) with 16-bit
+       * intermediates.
+       */
+      __m128i pa,pb,pc,smallest,nearest;
+      c = b; b = _mm_unpacklo_epi8(load3(prev), zero);
+      a = d; d = _mm_unpacklo_epi8(load3(row ), zero);
+
+      /* (p-a) == (a+b-c - a) == (b-c) */
+      pa = _mm_sub_epi16(b,c);
+
+      /* (p-b) == (a+b-c - b) == (a-c) */
+      pb = _mm_sub_epi16(a,c);
+
+      /* (p-c) == (a+b-c - c) == (a+b-c-c) == (b-c)+(a-c) */
+      pc = _mm_add_epi16(pa,pb);
+
+      pa = abs_i16(pa);  /* |p-a| */
+      pb = abs_i16(pb);  /* |p-b| */
+      pc = abs_i16(pc);  /* |p-c| */
+
+      smallest = _mm_min_epi16(pc, _mm_min_epi16(pa, pb));
+
+      /* Paeth breaks ties favoring a over b over c. */
+      nearest  = if_then_else(_mm_cmpeq_epi16(smallest, pa), a,
+                         if_then_else(_mm_cmpeq_epi16(smallest, pb), b,
+                                                                     c));
+
+      /* Note `_epi8`: we need addition to wrap modulo 255. */
+      d = _mm_add_epi8(d, nearest);
+      store3(row, _mm_packus_epi16(d,d));
+
+      prev += 3;
+      row  += 3;
+      rb   -= 3;
+   }
+}
+
+void png_read_filter_row_paeth4_sse2(png_row_infop row_info, png_bytep row,
+   png_const_bytep prev)
+{
+   /* Paeth tries to predict pixel d using the pixel to the left of it, a,
+    * and two pixels from the previous row, b and c:
+    *   prev: c b
+    *   row:  a d
+    * The Paeth function predicts d to be whichever of a, b, or c is nearest to
+    * p=a+b-c.
+    *
+    * The first pixel has no left context, and so uses an Up filter, p = b.
+    * This works naturally with our main loop's p = a+b-c if we force a and c
+    * to zero.
+    * Here we zero b and d, which become c and a respectively at the start of
+    * the loop.
+    */
+   png_size_t rb;
+   const __m128i zero = _mm_setzero_si128();
+   __m128i pa,pb,pc,smallest,nearest;
+   __m128i c, b = zero,
+           a, d = zero;
+
+   png_debug(1, "in png_read_filter_row_paeth4_sse2");
+
+   rb = row_info->rowbytes+4;
+   while (rb > 4) {
+      /* It's easiest to do this math (particularly, deal with pc) with 16-bit
+       * intermediates.
+       */
+      c = b; b = _mm_unpacklo_epi8(load4(prev), zero);
+      a = d; d = _mm_unpacklo_epi8(load4(row ), zero);
+
+      /* (p-a) == (a+b-c - a) == (b-c) */
+      pa = _mm_sub_epi16(b,c);
+
+      /* (p-b) == (a+b-c - b) == (a-c) */
+      pb = _mm_sub_epi16(a,c);
+
+      /* (p-c) == (a+b-c - c) == (a+b-c-c) == (b-c)+(a-c) */
+      pc = _mm_add_epi16(pa,pb);
+
+      pa = abs_i16(pa);  /* |p-a| */
+      pb = abs_i16(pb);  /* |p-b| */
+      pc = abs_i16(pc);  /* |p-c| */
+
+      smallest = _mm_min_epi16(pc, _mm_min_epi16(pa, pb));
+
+      /* Paeth breaks ties favoring a over b over c. */
+      nearest  = if_then_else(_mm_cmpeq_epi16(smallest, pa), a,
+                         if_then_else(_mm_cmpeq_epi16(smallest, pb), b,
+                                                                     c));
+
+      /* Note `_epi8`: we need addition to wrap modulo 255. */
+      d = _mm_add_epi8(d, nearest);
+      store4(row, _mm_packus_epi16(d,d));
+
+      prev += 4;
+      row  += 4;
+      rb   -= 4;
+   }
+}
+
+#endif /* PNG_INTEL_SSE_IMPLEMENTATION > 0 */
+#endif /* READ */

data/vendor/git.code.sf.net/p/libpng/code/intel/intel_init.c

@@ -0,0 +1,53 @@
+
+/* intel_init.c - SSE2 optimized filter functions
+ *
+ * Copyright (c) 2016-2017 Glenn Randers-Pehrson
+ * Written by Mike Klein and Matt Sarett, Google, Inc.
+ * Derived from arm/arm_init.c
+ *
+ * Last changed in libpng 1.6.29 [March 16, 2017]
+ *
+ * This code is released under the libpng license.
+ * For conditions of distribution and use, see the disclaimer
+ * and license in png.h
+ */
+
+#include "../pngpriv.h"
+
+#ifdef PNG_READ_SUPPORTED
+#if PNG_INTEL_SSE_IMPLEMENTATION > 0
+
+void
+png_init_filter_functions_sse2(png_structp pp, unsigned int bpp)
+{
+   /* The techniques used to implement each of these filters in SSE operate on
+    * one pixel at a time.
+    * So they generally speed up 3bpp images about 3x, 4bpp images about 4x.
+    * They can scale up to 6 and 8 bpp images and down to 2 bpp images,
+    * but they'd not likely have any benefit for 1bpp images.
+    * Most of these can be implemented using only MMX and 64-bit registers,
+    * but they end up a bit slower than using the equally-ubiquitous SSE2.
+   */
+   png_debug(1, "in png_init_filter_functions_sse2");
+   if (bpp == 3)
+   {
+      pp->read_filter[PNG_FILTER_VALUE_SUB-1] = png_read_filter_row_sub3_sse2;
+      pp->read_filter[PNG_FILTER_VALUE_AVG-1] = png_read_filter_row_avg3_sse2;
+      pp->read_filter[PNG_FILTER_VALUE_PAETH-1] =
+         png_read_filter_row_paeth3_sse2;
+   }
+   else if (bpp == 4)
+   {
+      pp->read_filter[PNG_FILTER_VALUE_SUB-1] = png_read_filter_row_sub4_sse2;
+      pp->read_filter[PNG_FILTER_VALUE_AVG-1] = png_read_filter_row_avg4_sse2;
+      pp->read_filter[PNG_FILTER_VALUE_PAETH-1] =
+          png_read_filter_row_paeth4_sse2;
+   }
+
+   /* No need optimize PNG_FILTER_VALUE_UP.  The compiler should
+    * autovectorize.
+    */
+}
+
+#endif /* PNG_INTEL_SSE_IMPLEMENTATION > 0 */
+#endif /* PNG_READ_SUPPORTED */

data/vendor/git.code.sf.net/p/libpng/code/libpng-config.in

@@ -0,0 +1,127 @@
+#! /bin/sh
+
+# libpng-config
+# provides configuration info for libpng.
+
+# Copyright (C) 2002, 2004, 2006, 2007 Glenn Randers-Pehrson
+
+# This code is released under the libpng license.
+# For conditions of distribution and use, see the disclaimer
+# and license in png.h
+
+# Modeled after libxml-config.
+
+version="@PNGLIB_VERSION@"
+prefix="@prefix@"
+exec_prefix="@exec_prefix@"
+libdir="@libdir@"
+includedir="@includedir@/libpng@PNGLIB_MAJOR@@PNGLIB_MINOR@"
+libs="-lpng@PNGLIB_MAJOR@@PNGLIB_MINOR@"
+all_libs="-lpng@PNGLIB_MAJOR@@PNGLIB_MINOR@ @LIBS@"
+I_opts="-I${includedir}"
+L_opts="-L${libdir}"
+R_opts=""
+cppflags=""
+ccopts=""
+ldopts=""
+
+usage()
+{
+    cat <<EOF
+Usage: $0 [OPTION] ...
+
+Known values for OPTION are:
+
+  --prefix        print libpng prefix
+  --libdir        print path to directory containing library
+  --libs          print library linking information
+  --ccopts        print compiler options
+  --cppflags      print pre-processor flags
+  --cflags        print preprocessor flags, I_opts, and compiler options
+  --I_opts        print "-I" include options
+  --L_opts        print linker "-L" flags for dynamic linking
+  --R_opts        print dynamic linker "-R" or "-rpath" flags
+  --ldopts        print linker options
+  --ldflags       print linker flags (ldopts, L_opts, R_opts, and libs)
+  --static        revise subsequent outputs for static linking
+  --help          print this help and exit
+  --version       print version information
+EOF
+
+    exit $1
+}
+
+if test $# -eq 0; then
+    usage 1
+fi
+
+while test $# -gt 0; do
+    case "$1" in
+
+    --prefix)
+        echo ${prefix}
+        ;;
+
+    --version)
+        echo ${version}
+        exit 0
+        ;;
+
+    --help)
+        usage 0
+        ;;
+
+    --ccopts)
+        echo ${ccopts}
+        ;;
+
+    --cppflags)
+        echo ${cppflags}
+        ;;
+
+    --cflags)
+        echo ${I_opts} ${cppflags} ${ccopts}
+        ;;
+
+    --libdir)
+        echo ${libdir}
+        ;;
+
+    --libs)
+        echo ${libs}
+        ;;
+
+    --I_opts)
+        echo ${I_opts}
+        ;;
+
+    --L_opts)
+        echo ${L_opts}
+        ;;
+
+    --R_opts)
+        echo ${R_opts}
+        ;;
+
+    --ldopts)
+        echo ${ldopts}
+        ;;
+
+    --ldflags)
+        echo ${ldopts} ${L_opts} ${R_opts} ${libs}
+        ;;
+
+    --static)
+        R_opts=""
+        libs=${all_libs}
+        ;;
+
+    *)
+        usage
+        exit 1
+        ;;
+    esac
+    shift
+done
+
+exit 0