libpng: Enable intrinsics on x86/SSE2, ppc64/VSX, and all arm/NEON

9 files changed, 1361 insertions, 27 deletions

diff --git a/drivers/png/SCsub b/drivers/png/SCsub
index fe8c8fa8cc..dd4777a19b 100644
--- a/drivers/png/SCsub
+++ b/drivers/png/SCsub
@@ -33,30 +33,30 @@ if env["builtin_libpng"]:
     # Needed for drivers includes and in platform/web.
     env.Prepend(CPPPATH=[thirdparty_dir])
 
-    # Currently .ASM filter_neon.S does not compile on NT.
-    import os
-
-    # Enable ARM NEON instructions on 32-bit Android to compile more optimized code.
-    use_neon = env["platform"] == "android" and env["arch"] == "arm32" and os.name != "nt"
-    if use_neon:
-        env_png.Append(CPPDEFINES=[("PNG_ARM_NEON_OPT", 2)])
-    else:
-        env_png.Append(CPPDEFINES=[("PNG_ARM_NEON_OPT", 0)])
-
     env_thirdparty = env_png.Clone()
     env_thirdparty.disable_warnings()
     env_thirdparty.add_source_files(thirdparty_obj, thirdparty_sources)
 
-    if use_neon:
-        env_neon = env_thirdparty.Clone()
-        if "S_compiler" in env:
-            env_neon["CC"] = env["S_compiler"]
-        neon_sources = []
-        neon_sources.append(env_neon.Object(thirdparty_dir + "/arm/arm_init.c"))
-        neon_sources.append(env_neon.Object(thirdparty_dir + "/arm/filter_neon_intrinsics.c"))
-        neon_sources.append(env_neon.Object(thirdparty_dir + "/arm/filter_neon.S"))
-        neon_sources.append(env_neon.Object(thirdparty_dir + "/arm/palette_neon_intrinsics.c"))
-        thirdparty_obj += neon_sources
+    if env["arch"].startswith("arm"):
+        if env.msvc:  # Can't compile assembly files with MSVC.
+            env_thirdparty.Append(CPPDEFINES=[("PNG_ARM_NEON_OPT"), 0])
+        else:
+            env_neon = env_thirdparty.Clone()
+            if "S_compiler" in env:
+                env_neon["CC"] = env["S_compiler"]
+            neon_sources = []
+            neon_sources.append(env_neon.Object(thirdparty_dir + "/arm/arm_init.c"))
+            neon_sources.append(env_neon.Object(thirdparty_dir + "/arm/filter_neon_intrinsics.c"))
+            neon_sources.append(env_neon.Object(thirdparty_dir + "/arm/filter_neon.S"))
+            neon_sources.append(env_neon.Object(thirdparty_dir + "/arm/palette_neon_intrinsics.c"))
+            thirdparty_obj += neon_sources
+    elif env["arch"].startswith("x86"):
+        env_thirdparty.Append(CPPDEFINES=["PNG_INTEL_SSE"])
+        env_thirdparty.add_source_files(thirdparty_obj, thirdparty_dir + "/intel/intel_init.c")
+        env_thirdparty.add_source_files(thirdparty_obj, thirdparty_dir + "/intel/filter_sse2_intrinsics.c")
+    elif env["arch"] == "ppc64":
+        env_thirdparty.add_source_files(thirdparty_obj, thirdparty_dir + "/powerpc/powerpc_init.c")
+        env_thirdparty.add_source_files(thirdparty_obj, thirdparty_dir + "/powerpc/filter_vsx_intrinsics.c")
 
     env.drivers_sources += thirdparty_obj
 
diff --git a/modules/text_server_adv/gdextension_build/SConstruct b/modules/text_server_adv/gdextension_build/SConstruct
index af9dae84e3..4b12a4e41d 100644
--- a/modules/text_server_adv/gdextension_build/SConstruct
+++ b/modules/text_server_adv/gdextension_build/SConstruct
@@ -262,7 +262,6 @@ if env["freetype_enabled"]:
         CPPDEFINES=[
             "FT2_BUILD_LIBRARY",
             "FT_CONFIG_OPTION_USE_PNG",
-            ("PNG_ARM_NEON_OPT", 0),
             "FT_CONFIG_OPTION_SYSTEM_ZLIB",
         ]
     )
diff --git a/modules/text_server_fb/gdextension_build/SConstruct b/modules/text_server_fb/gdextension_build/SConstruct
index 51a6ee06be..fcf8976019 100644
--- a/modules/text_server_fb/gdextension_build/SConstruct
+++ b/modules/text_server_fb/gdextension_build/SConstruct
@@ -257,7 +257,6 @@ if env["freetype_enabled"]:
         CPPDEFINES=[
             "FT2_BUILD_LIBRARY",
             "FT_CONFIG_OPTION_USE_PNG",
-            ("PNG_ARM_NEON_OPT", 0),
             "FT_CONFIG_OPTION_SYSTEM_ZLIB",
         ]
     )
diff --git a/platform/ios/detect.py b/platform/ios/detect.py
index f5501e3d85..a01f4b55db 100644
--- a/platform/ios/detect.py
+++ b/platform/ios/detect.py
@@ -85,19 +85,18 @@ def configure(env: "Environment"):
     env["ENV"]["PATH"] = env["IOS_TOOLCHAIN_PATH"] + "/Developer/usr/bin/:" + env["ENV"]["PATH"]
 
     compiler_path = "$IOS_TOOLCHAIN_PATH/usr/bin/${ios_triple}"
-    s_compiler_path = "$IOS_TOOLCHAIN_PATH/Developer/usr/bin/"
 
     ccache_path = os.environ.get("CCACHE")
     if ccache_path is None:
         env["CC"] = compiler_path + "clang"
         env["CXX"] = compiler_path + "clang++"
-        env["S_compiler"] = s_compiler_path + "gcc"
+        env["S_compiler"] = compiler_path + "clang"
     else:
         # there aren't any ccache wrappers available for iOS,
         # to enable caching we need to prepend the path to the ccache binary
         env["CC"] = ccache_path + " " + compiler_path + "clang"
         env["CXX"] = ccache_path + " " + compiler_path + "clang++"
-        env["S_compiler"] = ccache_path + " " + s_compiler_path + "gcc"
+        env["S_compiler"] = ccache_path + " " + compiler_path + "clang"
     env["AR"] = compiler_path + "ar"
     env["RANLIB"] = compiler_path + "ranlib"
 
diff --git a/thirdparty/README.md b/thirdparty/README.md
index a918acbe77..3a98f676b4 100644
--- a/thirdparty/README.md
+++ b/thirdparty/README.md
@@ -315,14 +315,14 @@ Files extracted from upstream source:
 ## libpng
 
 - Upstream: http://libpng.org/pub/png/libpng.html
-- Version: 1.6.38 (0a158f3506502dfa23edfc42790dfaed82efba17, 2022)
+- Version: 1.6.39 (07b8803110da160b158ebfef872627da6c85cbdf, 2022)
 - License: libpng/zlib
 
 Files extracted from upstream source:
 
 - all .c and .h files of the main directory, except from
   `example.c` and `pngtest.c`
-- the arm/ folder
+- `arm/`, `intel/` and `powerpc/` folders
 - `scripts/pnglibconf.h.prebuilt` as `pnglibconf.h`
 - `LICENSE`
 
diff --git a/thirdparty/libpng/intel/filter_sse2_intrinsics.c b/thirdparty/libpng/intel/filter_sse2_intrinsics.c
new file mode 100644
index 0000000000..d3c0fe9e2d
--- /dev/null
+++ b/thirdparty/libpng/intel/filter_sse2_intrinsics.c
@@ -0,0 +1,391 @@
+
+/* filter_sse2_intrinsics.c - SSE2 optimized filter functions
+ *
+ * Copyright (c) 2018 Cosmin Truta
+ * Copyright (c) 2016-2017 Glenn Randers-Pehrson
+ * Written by Mike Klein and Matt Sarett
+ * Derived from arm/filter_neon_intrinsics.c
+ *
+ * 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) {
+   int tmp;
+   memcpy(&tmp, p, sizeof(tmp));
+   return _mm_cvtsi32_si128(tmp);
+}
+
+static void store4(void* p, __m128i v) {
+   int tmp = _mm_cvtsi128_si32(v);
+   memcpy(p, &tmp, sizeof(int));
+}
+
+static __m128i load3(const void* p) {
+   png_uint_32 tmp = 0;
+   memcpy(&tmp, p, 3);
+   return _mm_cvtsi32_si128(tmp);
+}
+
+static void store3(void* p, __m128i v) {
+   int tmp = _mm_cvtsi128_si32(v);
+   memcpy(p, &tmp, 3);
+}
+
+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.
+    */
+   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.
+    */
+   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.
+    */
+
+   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.
+    */
+   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.
+    */
+   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.
+    */
+   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 */
diff --git a/thirdparty/libpng/intel/intel_init.c b/thirdparty/libpng/intel/intel_init.c
new file mode 100644
index 0000000000..2f8168b7c4
--- /dev/null
+++ b/thirdparty/libpng/intel/intel_init.c
@@ -0,0 +1,52 @@
+
+/* intel_init.c - SSE2 optimized filter functions
+ *
+ * Copyright (c) 2018 Cosmin Truta
+ * Copyright (c) 2016-2017 Glenn Randers-Pehrson
+ * Written by Mike Klein and Matt Sarett, Google, Inc.
+ * Derived from arm/arm_init.c
+ *
+ * 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 */
diff --git a/thirdparty/libpng/powerpc/filter_vsx_intrinsics.c b/thirdparty/libpng/powerpc/filter_vsx_intrinsics.c
new file mode 100644
index 0000000000..01cf8800dc
--- /dev/null
+++ b/thirdparty/libpng/powerpc/filter_vsx_intrinsics.c
@@ -0,0 +1,768 @@
+/* filter_vsx_intrinsics.c - PowerPC optimised filter functions
+ *
+ * Copyright (c) 2018 Cosmin Truta
+ * Copyright (c) 2017 Glenn Randers-Pehrson
+ * Written by Vadim Barkov, 2017.
+ *
+ * This code is released under the libpng license.
+ * For conditions of distribution and use, see the disclaimer
+ * and license in png.h
+ */
+
+#include <stdio.h>
+#include <stdint.h>
+#include "../pngpriv.h"
+
+#ifdef PNG_READ_SUPPORTED
+
+/* This code requires -maltivec and -mvsx on the command line: */
+#if PNG_POWERPC_VSX_IMPLEMENTATION == 1 /* intrinsics code from pngpriv.h */
+
+#include <altivec.h>
+
+#if PNG_POWERPC_VSX_OPT > 0
+
+#ifndef __VSX__
+#  error "This code requires VSX support (POWER7 and later). Please provide -mvsx compiler flag."
+#endif
+
+#define vec_ld_unaligned(vec,data) vec = vec_vsx_ld(0,data)
+#define vec_st_unaligned(vec,data) vec_vsx_st(vec,0,data)
+
+
+/* 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.
+ * ( this is taken from ../intel/filter_sse2_intrinsics.c )
+ */
+
+#define vsx_declare_common_vars(row_info,row,prev_row,offset) \
+   png_byte i;\
+   png_bytep rp = row + offset;\
+   png_const_bytep pp = prev_row;\
+   size_t unaligned_top = 16 - (((size_t)rp % 16));\
+   size_t istop;\
+   if(unaligned_top == 16)\
+      unaligned_top = 0;\
+   istop = row_info->rowbytes;\
+   if((unaligned_top < istop))\
+      istop -= unaligned_top;\
+   else{\
+      unaligned_top = istop;\
+      istop = 0;\
+   }
+
+void png_read_filter_row_up_vsx(png_row_infop row_info, png_bytep row,
+                                png_const_bytep prev_row)
+{
+   vector unsigned char rp_vec;
+   vector unsigned char pp_vec;
+   vsx_declare_common_vars(row_info,row,prev_row,0)
+
+   /* Altivec operations require 16-byte aligned data
+    * but input can be unaligned. So we calculate
+    * unaligned part as usual.
+    */
+   for (i = 0; i < unaligned_top; i++)
+   {
+      *rp = (png_byte)(((int)(*rp) + (int)(*pp++)) & 0xff);
+      rp++;
+   }
+
+   /* Using SIMD while we can */
+   while( istop >= 16 )
+   {
+      rp_vec = vec_ld(0,rp);
+      vec_ld_unaligned(pp_vec,pp);
+
+      rp_vec = vec_add(rp_vec,pp_vec);
+
+      vec_st(rp_vec,0,rp);
+
+      pp += 16;
+      rp += 16;
+      istop -= 16;
+   }
+
+   if(istop > 0)
+   {
+      /* If byte count of row is not divisible by 16
+       * we will process remaining part as usual
+       */
+      for (i = 0; i < istop; i++)
+      {
+         *rp = (png_byte)(((int)(*rp) + (int)(*pp++)) & 0xff);
+         rp++;
+      }
+}
+
+}
+
+static const vector unsigned char VSX_LEFTSHIFTED1_4 = {16,16,16,16, 0, 1, 2, 3,16,16,16,16,16,16,16,16};
+static const vector unsigned char VSX_LEFTSHIFTED2_4 = {16,16,16,16,16,16,16,16, 4, 5, 6, 7,16,16,16,16};
+static const vector unsigned char VSX_LEFTSHIFTED3_4 = {16,16,16,16,16,16,16,16,16,16,16,16, 8, 9,10,11};
+
+static const vector unsigned char VSX_LEFTSHIFTED1_3 = {16,16,16, 0, 1, 2,16,16,16,16,16,16,16,16,16,16};
+static const vector unsigned char VSX_LEFTSHIFTED2_3 = {16,16,16,16,16,16, 3, 4, 5,16,16,16,16,16,16,16};
+static const vector unsigned char VSX_LEFTSHIFTED3_3 = {16,16,16,16,16,16,16,16,16, 6, 7, 8,16,16,16,16};
+static const vector unsigned char VSX_LEFTSHIFTED4_3 = {16,16,16,16,16,16,16,16,16,16,16,16, 9,10,11,16};
+
+static const vector unsigned char VSX_NOT_SHIFTED1_4 = {16,16,16,16, 4, 5, 6, 7,16,16,16,16,16,16,16,16};
+static const vector unsigned char VSX_NOT_SHIFTED2_4 = {16,16,16,16,16,16,16,16, 8, 9,10,11,16,16,16,16};
+static const vector unsigned char VSX_NOT_SHIFTED3_4 = {16,16,16,16,16,16,16,16,16,16,16,16,12,13,14,15};
+
+static const vector unsigned char VSX_NOT_SHIFTED1_3 = {16,16,16, 3, 4, 5,16,16,16,16,16,16,16,16,16,16};
+static const vector unsigned char VSX_NOT_SHIFTED2_3 = {16,16,16,16,16,16, 6, 7, 8,16,16,16,16,16,16,16};
+static const vector unsigned char VSX_NOT_SHIFTED3_3 = {16,16,16,16,16,16,16,16,16, 9,10,11,16,16,16,16};
+static const vector unsigned char VSX_NOT_SHIFTED4_3 = {16,16,16,16,16,16,16,16,16,16,16,16,12,13,14,16};
+
+static const vector unsigned char VSX_CHAR_ZERO = {0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
+#ifdef __LITTLE_ENDIAN__
+
+static const vector unsigned char VSX_CHAR_TO_SHORT1_4 = { 4,16, 5,16, 6,16, 7,16,16,16,16,16,16,16,16,16};
+static const vector unsigned char VSX_CHAR_TO_SHORT2_4 = { 8,16, 9,16,10,16,11,16,16,16,16,16,16,16,16,16};
+static const vector unsigned char VSX_CHAR_TO_SHORT3_4 = {12,16,13,16,14,16,15,16,16,16,16,16,16,16,16,16};
+
+static const vector unsigned char VSX_SHORT_TO_CHAR1_4 = {16,16,16,16, 0, 2, 4, 6,16,16,16,16,16,16,16,16};
+static const vector unsigned char VSX_SHORT_TO_CHAR2_4 = {16,16,16,16,16,16,16,16, 0, 2, 4, 6,16,16,16,16};
+static const vector unsigned char VSX_SHORT_TO_CHAR3_4 = {16,16,16,16,16,16,16,16,16,16,16,16, 0, 2, 4, 6};
+
+static const vector unsigned char VSX_CHAR_TO_SHORT1_3 = { 3,16, 4,16, 5,16,16,16,16,16,16,16,16,16,16,16};
+static const vector unsigned char VSX_CHAR_TO_SHORT2_3 = { 6,16, 7,16, 8,16,16,16,16,16,16,16,16,16,16,16};
+static const vector unsigned char VSX_CHAR_TO_SHORT3_3 = { 9,16,10,16,11,16,16,16,16,16,16,16,16,16,16,16};
+static const vector unsigned char VSX_CHAR_TO_SHORT4_3 = {12,16,13,16,14,16,16,16,16,16,16,16,16,16,16,16};
+
+static const vector unsigned char VSX_SHORT_TO_CHAR1_3 = {16,16,16, 0, 2, 4,16,16,16,16,16,16,16,16,16,16};
+static const vector unsigned char VSX_SHORT_TO_CHAR2_3 = {16,16,16,16,16,16, 0, 2, 4,16,16,16,16,16,16,16};
+static const vector unsigned char VSX_SHORT_TO_CHAR3_3 = {16,16,16,16,16,16,16,16,16, 0, 2, 4,16,16,16,16};
+static const vector unsigned char VSX_SHORT_TO_CHAR4_3 = {16,16,16,16,16,16,16,16,16,16,16,16, 0, 2, 4,16};
+
+#elif defined(__BIG_ENDIAN__)
+
+static const vector unsigned char VSX_CHAR_TO_SHORT1_4 = {16, 4,16, 5,16, 6,16, 7,16,16,16,16,16,16,16,16};
+static const vector unsigned char VSX_CHAR_TO_SHORT2_4 = {16, 8,16, 9,16,10,16,11,16,16,16,16,16,16,16,16};
+static const vector unsigned char VSX_CHAR_TO_SHORT3_4 = {16,12,16,13,16,14,16,15,16,16,16,16,16,16,16,16};
+
+static const vector unsigned char VSX_SHORT_TO_CHAR1_4 = {16,16,16,16, 1, 3, 5, 7,16,16,16,16,16,16,16,16};
+static const vector unsigned char VSX_SHORT_TO_CHAR2_4 = {16,16,16,16,16,16,16,16, 1, 3, 5, 7,16,16,16,16};
+static const vector unsigned char VSX_SHORT_TO_CHAR3_4 = {16,16,16,16,16,16,16,16,16,16,16,16, 1, 3, 5, 7};
+
+static const vector unsigned char VSX_CHAR_TO_SHORT1_3 = {16, 3,16, 4,16, 5,16,16,16,16,16,16,16,16,16,16};
+static const vector unsigned char VSX_CHAR_TO_SHORT2_3 = {16, 6,16, 7,16, 8,16,16,16,16,16,16,16,16,16,16};
+static const vector unsigned char VSX_CHAR_TO_SHORT3_3 = {16, 9,16,10,16,11,16,16,16,16,16,16,16,16,16,16};
+static const vector unsigned char VSX_CHAR_TO_SHORT4_3 = {16,12,16,13,16,14,16,16,16,16,16,16,16,16,16,16};
+
+static const vector unsigned char VSX_SHORT_TO_CHAR1_3 = {16,16,16, 1, 3, 5,16,16,16,16,16,16,16,16,16,16};
+static const vector unsigned char VSX_SHORT_TO_CHAR2_3 = {16,16,16,16,16,16, 1, 3, 5,16,16,16,16,16,16,16};
+static const vector unsigned char VSX_SHORT_TO_CHAR3_3 = {16,16,16,16,16,16,16,16,16, 1, 3, 5,16,16,16,16};
+static const vector unsigned char VSX_SHORT_TO_CHAR4_3 = {16,16,16,16,16,16,16,16,16,16,16,16, 1, 3, 5,16};
+
+#endif
+
+#define vsx_char_to_short(vec,offset,bpp) (vector unsigned short)vec_perm((vec),VSX_CHAR_ZERO,VSX_CHAR_TO_SHORT##offset##_##bpp)
+#define vsx_short_to_char(vec,offset,bpp) vec_perm(((vector unsigned char)(vec)),VSX_CHAR_ZERO,VSX_SHORT_TO_CHAR##offset##_##bpp)
+
+#ifdef PNG_USE_ABS
+#  define vsx_abs(number) abs(number)
+#else
+#  define vsx_abs(number) (number > 0) ? (number) : -(number)
+#endif
+
+void png_read_filter_row_sub4_vsx(png_row_infop row_info, png_bytep row,
+                                  png_const_bytep prev_row)
+{
+   png_byte bpp = 4;
+
+   vector unsigned char rp_vec;
+   vector unsigned char part_vec;
+
+   vsx_declare_common_vars(row_info,row,prev_row,bpp)
+
+   PNG_UNUSED(pp)
+
+   /* Altivec operations require 16-byte aligned data
+    * but input can be unaligned. So we calculate
+    * unaligned part as usual.
+    */
+   for (i = 0; i < unaligned_top; i++)
+   {
+      *rp = (png_byte)(((int)(*rp) + (int)(*(rp-bpp))) & 0xff);
+      rp++;
+   }
+
+   /* Using SIMD while we can */
+   while( istop >= 16 )
+   {
+      for(i=0;i < bpp ; i++)
+      {
+         *rp = (png_byte)(((int)(*rp) + (int)(*(rp-bpp))) & 0xff);
+         rp++;
+      }
+      rp -= bpp;
+
+      rp_vec = vec_ld(0,rp);
+      part_vec = vec_perm(rp_vec,VSX_CHAR_ZERO,VSX_LEFTSHIFTED1_4);
+      rp_vec = vec_add(rp_vec,part_vec);
+
+      part_vec = vec_perm(rp_vec,VSX_CHAR_ZERO,VSX_LEFTSHIFTED2_4);
+      rp_vec = vec_add(rp_vec,part_vec);
+
+      part_vec = vec_perm(rp_vec,VSX_CHAR_ZERO,VSX_LEFTSHIFTED3_4);
+      rp_vec = vec_add(rp_vec,part_vec);
+
+      vec_st(rp_vec,0,rp);
+
+      rp += 16;
+      istop -= 16;
+   }
+
+   if(istop > 0)
+      for (i = 0; i < istop % 16; i++)
+      {
+         *rp = (png_byte)(((int)(*rp) + (int)(*(rp - bpp))) & 0xff);
+         rp++;
+      }
+
+}
+
+void png_read_filter_row_sub3_vsx(png_row_infop row_info, png_bytep row,
+                                  png_const_bytep prev_row)
+{
+   png_byte bpp = 3;
+
+   vector unsigned char rp_vec;
+   vector unsigned char part_vec;
+
+   vsx_declare_common_vars(row_info,row,prev_row,bpp)
+
+   PNG_UNUSED(pp)
+
+   /* Altivec operations require 16-byte aligned data
+    * but input can be unaligned. So we calculate
+    * unaligned part as usual.
+    */
+   for (i = 0; i < unaligned_top; i++)
+   {
+      *rp = (png_byte)(((int)(*rp) + (int)(*(rp-bpp))) & 0xff);
+      rp++;
+   }
+
+   /* Using SIMD while we can */
+   while( istop >= 16 )
+   {
+      for(i=0;i < bpp ; i++)
+      {
+         *rp = (png_byte)(((int)(*rp) + (int)(*(rp-bpp))) & 0xff);
+         rp++;
+      }
+      rp -= bpp;
+
+      rp_vec = vec_ld(0,rp);
+      part_vec = vec_perm(rp_vec,VSX_CHAR_ZERO,VSX_LEFTSHIFTED1_3);
+      rp_vec = vec_add(rp_vec,part_vec);
+
+      part_vec = vec_perm(rp_vec,VSX_CHAR_ZERO,VSX_LEFTSHIFTED2_3);
+      rp_vec = vec_add(rp_vec,part_vec);
+
+      part_vec = vec_perm(rp_vec,VSX_CHAR_ZERO,VSX_LEFTSHIFTED3_3);
+      rp_vec = vec_add(rp_vec,part_vec);
+
+      part_vec = vec_perm(rp_vec,VSX_CHAR_ZERO,VSX_LEFTSHIFTED4_3);
+      rp_vec = vec_add(rp_vec,part_vec);
+
+      vec_st(rp_vec,0,rp);
+      rp += 15;
+      istop -= 16;
+
+      /* Since 16 % bpp = 16 % 3 = 1, last element of array must
+       * be proceeded manually
+       */
+      *rp = (png_byte)(((int)(*rp) + (int)(*(rp-bpp))) & 0xff);
+      rp++;
+   }
+
+   if(istop > 0)
+      for (i = 0; i < istop % 16; i++)
+      {
+         *rp = (png_byte)(((int)(*rp) + (int)(*(rp-bpp))) & 0xff);
+         rp++;
+      }
+}
+
+void png_read_filter_row_avg4_vsx(png_row_infop row_info, png_bytep row,
+                                  png_const_bytep prev_row)
+{
+   png_byte bpp = 4;
+
+   vector unsigned char rp_vec;
+   vector unsigned char pp_vec;
+   vector unsigned char pp_part_vec;
+   vector unsigned char rp_part_vec;
+   vector unsigned char avg_vec;
+
+   vsx_declare_common_vars(row_info,row,prev_row,bpp)
+   rp -= bpp;
+   if(istop >= bpp)
+      istop -= bpp;
+
+   for (i = 0; i < bpp; i++)
+   {
+      *rp = (png_byte)(((int)(*rp) +
+         ((int)(*pp++) / 2 )) & 0xff);
+
+      rp++;
+   }
+
+   /* Altivec operations require 16-byte aligned data
+    * but input can be unaligned. So we calculate
+    * unaligned part as usual.
+    */
+   for (i = 0; i < unaligned_top; i++)
+   {
+      *rp = (png_byte)(((int)(*rp) +
+         (int)(*pp++ + *(rp-bpp)) / 2 ) & 0xff);
+
+      rp++;
+   }
+
+   /* Using SIMD while we can */
+   while( istop >= 16 )
+   {
+      for(i=0;i < bpp ; i++)
+      {
+         *rp = (png_byte)(((int)(*rp) +
+            (int)(*pp++ + *(rp-bpp)) / 2 ) & 0xff);
+
+         rp++;
+      }
+      rp -= bpp;
+      pp -= bpp;
+
+      vec_ld_unaligned(pp_vec,pp);
+      rp_vec = vec_ld(0,rp);
+
+      rp_part_vec = vec_perm(rp_vec,VSX_CHAR_ZERO,VSX_LEFTSHIFTED1_4);
+      pp_part_vec = vec_perm(pp_vec,VSX_CHAR_ZERO,VSX_NOT_SHIFTED1_4);
+      avg_vec = vec_avg(rp_part_vec,pp_part_vec);
+      avg_vec = vec_sub(avg_vec, vec_and(vec_xor(rp_part_vec,pp_part_vec),vec_splat_u8(1)));
+      rp_vec = vec_add(rp_vec,avg_vec);
+
+      rp_part_vec = vec_perm(rp_vec,VSX_CHAR_ZERO,VSX_LEFTSHIFTED2_4);
+      pp_part_vec = vec_perm(pp_vec,VSX_CHAR_ZERO,VSX_NOT_SHIFTED2_4);
+      avg_vec = vec_avg(rp_part_vec,pp_part_vec);
+      avg_vec = vec_sub(avg_vec, vec_and(vec_xor(rp_part_vec,pp_part_vec),vec_splat_u8(1)));
+      rp_vec = vec_add(rp_vec,avg_vec);
+
+      rp_part_vec = vec_perm(rp_vec,VSX_CHAR_ZERO,VSX_LEFTSHIFTED3_4);
+      pp_part_vec = vec_perm(pp_vec,VSX_CHAR_ZERO,VSX_NOT_SHIFTED3_4);
+      avg_vec = vec_avg(rp_part_vec,pp_part_vec);
+      avg_vec = vec_sub(avg_vec, vec_and(vec_xor(rp_part_vec,pp_part_vec),vec_splat_u8(1)));
+      rp_vec = vec_add(rp_vec,avg_vec);
+
+      vec_st(rp_vec,0,rp);
+
+      rp += 16;
+      pp += 16;
+      istop -= 16;
+   }
+
+   if(istop  > 0)
+      for (i = 0; i < istop % 16; i++)
+      {
+         *rp = (png_byte)(((int)(*rp) +
+            (int)(*pp++ + *(rp-bpp)) / 2 ) & 0xff);
+
+         rp++;
+      }
+}
+
+void png_read_filter_row_avg3_vsx(png_row_infop row_info, png_bytep row,
+                                  png_const_bytep prev_row)
+{
+  png_byte bpp = 3;
+
+  vector unsigned char rp_vec;
+  vector unsigned char pp_vec;
+  vector unsigned char pp_part_vec;
+  vector unsigned char rp_part_vec;
+  vector unsigned char avg_vec;
+
+  vsx_declare_common_vars(row_info,row,prev_row,bpp)
+  rp -= bpp;
+  if(istop >= bpp)
+     istop -= bpp;
+
+  for (i = 0; i < bpp; i++)
+  {
+     *rp = (png_byte)(((int)(*rp) +
+        ((int)(*pp++) / 2 )) & 0xff);
+
+     rp++;
+  }
+
+  /* Altivec operations require 16-byte aligned data
+   * but input can be unaligned. So we calculate
+   * unaligned part as usual.
+   */
+  for (i = 0; i < unaligned_top; i++)
+  {
+     *rp = (png_byte)(((int)(*rp) +
+        (int)(*pp++ + *(rp-bpp)) / 2 ) & 0xff);
+
+     rp++;
+  }
+
+  /* Using SIMD while we can */
+  while( istop >= 16 )
+  {
+     for(i=0;i < bpp ; i++)
+     {
+        *rp = (png_byte)(((int)(*rp) +
+           (int)(*pp++ + *(rp-bpp)) / 2 ) & 0xff);
+
+        rp++;
+     }
+     rp -= bpp;
+     pp -= bpp;
+
+     vec_ld_unaligned(pp_vec,pp);
+     rp_vec = vec_ld(0,rp);
+
+     rp_part_vec = vec_perm(rp_vec,VSX_CHAR_ZERO,VSX_LEFTSHIFTED1_3);
+     pp_part_vec = vec_perm(pp_vec,VSX_CHAR_ZERO,VSX_NOT_SHIFTED1_3);
+     avg_vec = vec_avg(rp_part_vec,pp_part_vec);
+     avg_vec = vec_sub(avg_vec, vec_and(vec_xor(rp_part_vec,pp_part_vec),vec_splat_u8(1)));
+     rp_vec = vec_add(rp_vec,avg_vec);
+
+     rp_part_vec = vec_perm(rp_vec,VSX_CHAR_ZERO,VSX_LEFTSHIFTED2_3);
+     pp_part_vec = vec_perm(pp_vec,VSX_CHAR_ZERO,VSX_NOT_SHIFTED2_3);
+     avg_vec = vec_avg(rp_part_vec,pp_part_vec);
+     avg_vec = vec_sub(avg_vec, vec_and(vec_xor(rp_part_vec,pp_part_vec),vec_splat_u8(1)));
+     rp_vec = vec_add(rp_vec,avg_vec);
+
+     rp_part_vec = vec_perm(rp_vec,VSX_CHAR_ZERO,VSX_LEFTSHIFTED3_3);
+     pp_part_vec = vec_perm(pp_vec,VSX_CHAR_ZERO,VSX_NOT_SHIFTED3_3);
+     avg_vec = vec_avg(rp_part_vec,pp_part_vec);
+     avg_vec = vec_sub(avg_vec, vec_and(vec_xor(rp_part_vec,pp_part_vec),vec_splat_u8(1)));
+     rp_vec = vec_add(rp_vec,avg_vec);
+
+     rp_part_vec = vec_perm(rp_vec,VSX_CHAR_ZERO,VSX_LEFTSHIFTED4_3);
+     pp_part_vec = vec_perm(pp_vec,VSX_CHAR_ZERO,VSX_NOT_SHIFTED4_3);
+     avg_vec = vec_avg(rp_part_vec,pp_part_vec);
+     avg_vec = vec_sub(avg_vec, vec_and(vec_xor(rp_part_vec,pp_part_vec),vec_splat_u8(1)));
+     rp_vec = vec_add(rp_vec,avg_vec);
+
+     vec_st(rp_vec,0,rp);
+
+     rp += 15;
+     pp += 15;
+     istop -= 16;
+
+     /* Since 16 % bpp = 16 % 3 = 1, last element of array must
+      * be proceeded manually
+      */
+     *rp = (png_byte)(((int)(*rp) +
+        (int)(*pp++ + *(rp-bpp)) / 2 ) & 0xff);
+     rp++;
+  }
+
+  if(istop  > 0)
+     for (i = 0; i < istop % 16; i++)
+     {
+        *rp = (png_byte)(((int)(*rp) +
+           (int)(*pp++ + *(rp-bpp)) / 2 ) & 0xff);
+
+        rp++;
+     }
+}
+
+/* Bytewise c ? t : e. */
+#define if_then_else(c,t,e) vec_sel(e,t,c)
+
+#define vsx_paeth_process(rp,pp,a,b,c,pa,pb,pc,bpp) {\
+      c = *(pp - bpp);\
+      a = *(rp - bpp);\
+      b = *pp++;\
+      p = b - c;\
+      pc = a - c;\
+      pa = vsx_abs(p);\
+      pb = vsx_abs(pc);\
+      pc = vsx_abs(p + pc);\
+      if (pb < pa) pa = pb, a = b;\
+      if (pc < pa) a = c;\
+      a += *rp;\
+      *rp++ = (png_byte)a;\
+      }
+
+void png_read_filter_row_paeth4_vsx(png_row_infop row_info, png_bytep row,
+   png_const_bytep prev_row)
+{
+   png_byte bpp = 4;
+
+   int a, b, c, pa, pb, pc, p;
+   vector unsigned char rp_vec;
+   vector unsigned char pp_vec;
+   vector unsigned short a_vec,b_vec,c_vec,nearest_vec;
+   vector signed short pa_vec,pb_vec,pc_vec,smallest_vec;
+
+   vsx_declare_common_vars(row_info,row,prev_row,bpp)
+   rp -= bpp;
+   if(istop >= bpp)
+      istop -= bpp;
+
+   /* Process the first pixel in the row completely (this is the same as 'up'
+    * because there is only one candidate predictor for the first row).
+    */
+   for(i = 0; i < bpp ; i++)
+   {
+      *rp = (png_byte)( *rp + *pp);
+      rp++;
+      pp++;
+   }
+
+   for(i = 0; i < unaligned_top ; i++)
+   {
+      vsx_paeth_process(rp,pp,a,b,c,pa,pb,pc,bpp)
+   }
+
+   while( istop >= 16)
+   {
+      for(i = 0; i < bpp ; i++)
+      {
+         vsx_paeth_process(rp,pp,a,b,c,pa,pb,pc,bpp)
+      }
+
+      rp -= bpp;
+      pp -= bpp;
+      rp_vec = vec_ld(0,rp);
+      vec_ld_unaligned(pp_vec,pp);
+
+      a_vec = vsx_char_to_short(vec_perm(rp_vec , VSX_CHAR_ZERO , VSX_LEFTSHIFTED1_4),1,4);
+      b_vec = vsx_char_to_short(vec_perm(pp_vec , VSX_CHAR_ZERO , VSX_NOT_SHIFTED1_4),1,4);
+      c_vec = vsx_char_to_short(vec_perm(pp_vec , VSX_CHAR_ZERO , VSX_LEFTSHIFTED1_4),1,4);
+      pa_vec = (vector signed short) vec_sub(b_vec,c_vec);
+      pb_vec = (vector signed short) vec_sub(a_vec , c_vec);
+      pc_vec = vec_add(pa_vec,pb_vec);
+      pa_vec = vec_abs(pa_vec);
+      pb_vec = vec_abs(pb_vec);
+      pc_vec = vec_abs(pc_vec);
+      smallest_vec = vec_min(pc_vec, vec_min(pa_vec,pb_vec));
+      nearest_vec =  if_then_else(
+            vec_cmpeq(pa_vec,smallest_vec),
+            a_vec,
+            if_then_else(
+              vec_cmpeq(pb_vec,smallest_vec),
+              b_vec,
+              c_vec
+              )
+            );
+      rp_vec = vec_add(rp_vec,(vsx_short_to_char(nearest_vec,1,4)));
+
+      a_vec = vsx_char_to_short(vec_perm(rp_vec , VSX_CHAR_ZERO , VSX_LEFTSHIFTED2_4),2,4);
+      b_vec = vsx_char_to_short(vec_perm(pp_vec , VSX_CHAR_ZERO , VSX_NOT_SHIFTED2_4),2,4);
+      c_vec = vsx_char_to_short(vec_perm(pp_vec , VSX_CHAR_ZERO , VSX_LEFTSHIFTED2_4),2,4);
+      pa_vec = (vector signed short) vec_sub(b_vec,c_vec);
+      pb_vec = (vector signed short) vec_sub(a_vec , c_vec);
+      pc_vec = vec_add(pa_vec,pb_vec);
+      pa_vec = vec_abs(pa_vec);
+      pb_vec = vec_abs(pb_vec);
+      pc_vec = vec_abs(pc_vec);
+      smallest_vec = vec_min(pc_vec, vec_min(pa_vec,pb_vec));
+      nearest_vec =  if_then_else(
+            vec_cmpeq(pa_vec,smallest_vec),
+            a_vec,
+            if_then_else(
+              vec_cmpeq(pb_vec,smallest_vec),
+              b_vec,
+              c_vec
+              )
+            );
+      rp_vec = vec_add(rp_vec,(vsx_short_to_char(nearest_vec,2,4)));
+
+      a_vec = vsx_char_to_short(vec_perm(rp_vec , VSX_CHAR_ZERO , VSX_LEFTSHIFTED3_4),3,4);
+      b_vec = vsx_char_to_short(vec_perm(pp_vec , VSX_CHAR_ZERO , VSX_NOT_SHIFTED3_4),3,4);
+      c_vec = vsx_char_to_short(vec_perm(pp_vec , VSX_CHAR_ZERO , VSX_LEFTSHIFTED3_4),3,4);
+      pa_vec = (vector signed short) vec_sub(b_vec,c_vec);
+      pb_vec = (vector signed short) vec_sub(a_vec , c_vec);
+      pc_vec = vec_add(pa_vec,pb_vec);
+      pa_vec = vec_abs(pa_vec);
+      pb_vec = vec_abs(pb_vec);
+      pc_vec = vec_abs(pc_vec);
+      smallest_vec = vec_min(pc_vec, vec_min(pa_vec,pb_vec));
+      nearest_vec =  if_then_else(
+            vec_cmpeq(pa_vec,smallest_vec),
+            a_vec,
+            if_then_else(
+              vec_cmpeq(pb_vec,smallest_vec),
+              b_vec,
+              c_vec
+              )
+            );
+      rp_vec = vec_add(rp_vec,(vsx_short_to_char(nearest_vec,3,4)));
+
+      vec_st(rp_vec,0,rp);
+
+      rp += 16;
+      pp += 16;
+      istop -= 16;
+   }
+
+   if(istop > 0)
+      for (i = 0; i < istop % 16; i++)
+      {
+         vsx_paeth_process(rp,pp,a,b,c,pa,pb,pc,bpp)
+      }
+}
+
+void png_read_filter_row_paeth3_vsx(png_row_infop row_info, png_bytep row,
+   png_const_bytep prev_row)
+{
+  png_byte bpp = 3;
+
+  int a, b, c, pa, pb, pc, p;
+  vector unsigned char rp_vec;
+  vector unsigned char pp_vec;
+  vector unsigned short a_vec,b_vec,c_vec,nearest_vec;
+  vector signed short pa_vec,pb_vec,pc_vec,smallest_vec;
+
+  vsx_declare_common_vars(row_info,row,prev_row,bpp)
+  rp -= bpp;
+  if(istop >= bpp)
+     istop -= bpp;
+
+  /* Process the first pixel in the row completely (this is the same as 'up'
+   * because there is only one candidate predictor for the first row).
+   */
+  for(i = 0; i < bpp ; i++)
+  {
+     *rp = (png_byte)( *rp + *pp);
+     rp++;
+     pp++;
+  }
+
+  for(i = 0; i < unaligned_top ; i++)
+  {
+     vsx_paeth_process(rp,pp,a,b,c,pa,pb,pc,bpp)
+  }
+
+  while( istop >= 16)
+  {
+     for(i = 0; i < bpp ; i++)
+     {
+        vsx_paeth_process(rp,pp,a,b,c,pa,pb,pc,bpp)
+     }
+
+     rp -= bpp;
+     pp -= bpp;
+     rp_vec = vec_ld(0,rp);
+     vec_ld_unaligned(pp_vec,pp);
+
+     a_vec = vsx_char_to_short(vec_perm(rp_vec , VSX_CHAR_ZERO , VSX_LEFTSHIFTED1_3),1,3);
+     b_vec = vsx_char_to_short(vec_perm(pp_vec , VSX_CHAR_ZERO , VSX_NOT_SHIFTED1_3),1,3);
+     c_vec = vsx_char_to_short(vec_perm(pp_vec , VSX_CHAR_ZERO , VSX_LEFTSHIFTED1_3),1,3);
+     pa_vec = (vector signed short) vec_sub(b_vec,c_vec);
+     pb_vec = (vector signed short) vec_sub(a_vec , c_vec);
+     pc_vec = vec_add(pa_vec,pb_vec);
+     pa_vec = vec_abs(pa_vec);
+     pb_vec = vec_abs(pb_vec);
+     pc_vec = vec_abs(pc_vec);
+     smallest_vec = vec_min(pc_vec, vec_min(pa_vec,pb_vec));
+     nearest_vec =  if_then_else(
+           vec_cmpeq(pa_vec,smallest_vec),
+           a_vec,
+           if_then_else(
+             vec_cmpeq(pb_vec,smallest_vec),
+             b_vec,
+             c_vec
+             )
+           );
+     rp_vec = vec_add(rp_vec,(vsx_short_to_char(nearest_vec,1,3)));
+
+     a_vec = vsx_char_to_short(vec_perm(rp_vec , VSX_CHAR_ZERO , VSX_LEFTSHIFTED2_3),2,3);
+     b_vec = vsx_char_to_short(vec_perm(pp_vec , VSX_CHAR_ZERO , VSX_NOT_SHIFTED2_3),2,3);
+     c_vec = vsx_char_to_short(vec_perm(pp_vec , VSX_CHAR_ZERO , VSX_LEFTSHIFTED2_3),2,3);
+     pa_vec = (vector signed short) vec_sub(b_vec,c_vec);
+     pb_vec = (vector signed short) vec_sub(a_vec , c_vec);
+     pc_vec = vec_add(pa_vec,pb_vec);
+     pa_vec = vec_abs(pa_vec);
+     pb_vec = vec_abs(pb_vec);
+     pc_vec = vec_abs(pc_vec);
+     smallest_vec = vec_min(pc_vec, vec_min(pa_vec,pb_vec));
+     nearest_vec =  if_then_else(
+           vec_cmpeq(pa_vec,smallest_vec),
+           a_vec,
+           if_then_else(
+             vec_cmpeq(pb_vec,smallest_vec),
+             b_vec,
+             c_vec
+             )
+           );
+     rp_vec = vec_add(rp_vec,(vsx_short_to_char(nearest_vec,2,3)));
+
+     a_vec = vsx_char_to_short(vec_perm(rp_vec , VSX_CHAR_ZERO , VSX_LEFTSHIFTED3_3),3,3);
+     b_vec = vsx_char_to_short(vec_perm(pp_vec , VSX_CHAR_ZERO , VSX_NOT_SHIFTED3_3),3,3);
+     c_vec = vsx_char_to_short(vec_perm(pp_vec , VSX_CHAR_ZERO , VSX_LEFTSHIFTED3_3),3,3);
+     pa_vec = (vector signed short) vec_sub(b_vec,c_vec);
+     pb_vec = (vector signed short) vec_sub(a_vec , c_vec);
+     pc_vec = vec_add(pa_vec,pb_vec);
+     pa_vec = vec_abs(pa_vec);
+     pb_vec = vec_abs(pb_vec);
+     pc_vec = vec_abs(pc_vec);
+     smallest_vec = vec_min(pc_vec, vec_min(pa_vec,pb_vec));
+     nearest_vec =  if_then_else(
+           vec_cmpeq(pa_vec,smallest_vec),
+           a_vec,
+           if_then_else(
+             vec_cmpeq(pb_vec,smallest_vec),
+             b_vec,
+             c_vec
+             )
+           );
+     rp_vec = vec_add(rp_vec,(vsx_short_to_char(nearest_vec,3,3)));
+
+     a_vec = vsx_char_to_short(vec_perm(rp_vec , VSX_CHAR_ZERO , VSX_LEFTSHIFTED4_3),4,3);
+     b_vec = vsx_char_to_short(vec_perm(pp_vec , VSX_CHAR_ZERO , VSX_NOT_SHIFTED4_3),4,3);
+     c_vec = vsx_char_to_short(vec_perm(pp_vec , VSX_CHAR_ZERO , VSX_LEFTSHIFTED4_3),4,3);
+     pa_vec = (vector signed short) vec_sub(b_vec,c_vec);
+     pb_vec = (vector signed short) vec_sub(a_vec , c_vec);
+     pc_vec = vec_add(pa_vec,pb_vec);
+     pa_vec = vec_abs(pa_vec);
+     pb_vec = vec_abs(pb_vec);
+     pc_vec = vec_abs(pc_vec);
+     smallest_vec = vec_min(pc_vec, vec_min(pa_vec,pb_vec));
+     nearest_vec =  if_then_else(
+           vec_cmpeq(pa_vec,smallest_vec),
+           a_vec,
+           if_then_else(
+             vec_cmpeq(pb_vec,smallest_vec),
+             b_vec,
+             c_vec
+             )
+           );
+     rp_vec = vec_add(rp_vec,(vsx_short_to_char(nearest_vec,4,3)));
+
+     vec_st(rp_vec,0,rp);
+
+     rp += 15;
+     pp += 15;
+     istop -= 16;
+
+     /* Since 16 % bpp = 16 % 3 = 1, last element of array must
+      * be proceeded manually
+      */
+     vsx_paeth_process(rp,pp,a,b,c,pa,pb,pc,bpp)
+  }
+
+  if(istop > 0)
+     for (i = 0; i < istop % 16; i++)
+     {
+        vsx_paeth_process(rp,pp,a,b,c,pa,pb,pc,bpp)
+     }
+}
+
+#endif /* PNG_POWERPC_VSX_OPT > 0 */
+#endif /* PNG_POWERPC_VSX_IMPLEMENTATION == 1 (intrinsics) */
+#endif /* READ */
diff --git a/thirdparty/libpng/powerpc/powerpc_init.c b/thirdparty/libpng/powerpc/powerpc_init.c
new file mode 100644
index 0000000000..54426c558e
--- /dev/null
+++ b/thirdparty/libpng/powerpc/powerpc_init.c
@@ -0,0 +1,126 @@
+
+/* powerpc_init.c - POWERPC optimised filter functions
+ *
+ * Copyright (c) 2018 Cosmin Truta
+ * Copyright (c) 2017 Glenn Randers-Pehrson
+ * Written by Vadim Barkov, 2017.
+ *
+ * This code is released under the libpng license.
+ * For conditions of distribution and use, see the disclaimer
+ * and license in png.h
+ */
+
+/* Below, after checking __linux__, various non-C90 POSIX 1003.1 functions are
+ * called.
+ */
+#define _POSIX_SOURCE 1
+
+#include <stdio.h>
+#include "../pngpriv.h"
+
+#ifdef PNG_READ_SUPPORTED
+
+#if PNG_POWERPC_VSX_OPT > 0
+#ifdef PNG_POWERPC_VSX_CHECK_SUPPORTED /* Do run-time checks */
+/* WARNING: it is strongly recommended that you do not build libpng with
+ * run-time checks for CPU features if at all possible.  In the case of the PowerPC
+ * VSX instructions there is no processor-specific way of detecting the
+ * presence of the required support, therefore run-time detection is extremely
+ * OS specific.
+ *
+ * You may set the macro PNG_POWERPC_VSX_FILE to the file name of file containing
+ * a fragment of C source code which defines the png_have_vsx function.  There
+ * are a number of implementations in contrib/powerpc-vsx, but the only one that
+ * has partial support is contrib/powerpc-vsx/linux.c - a generic Linux
+ * implementation which reads /proc/cpufino.
+ */
+#ifndef PNG_POWERPC_VSX_FILE
+#  ifdef __linux__
+#     define  PNG_POWERPC_VSX_FILE "contrib/powerpc-vsx/linux_aux.c"
+#  endif
+#endif
+
+#ifdef PNG_POWERPC_VSX_FILE
+
+#include <signal.h> /* for sig_atomic_t */
+static int png_have_vsx(png_structp png_ptr);
+#include PNG_POWERPC_VSX_FILE
+
+#else  /* PNG_POWERPC_VSX_FILE */
+#  error "PNG_POWERPC_VSX_FILE undefined: no support for run-time POWERPC VSX checks"
+#endif /* PNG_POWERPC_VSX_FILE */
+#endif /* PNG_POWERPC_VSX_CHECK_SUPPORTED */
+
+void
+png_init_filter_functions_vsx(png_structp pp, unsigned int bpp)
+{
+   /* The switch statement is compiled in for POWERPC_VSX_API, the call to
+    * png_have_vsx is compiled in for POWERPC_VSX_CHECK. If both are defined
+    * the check is only performed if the API has not set the PowerPC option on
+    * or off explicitly. In this case the check controls what happens.
+    */
+
+#ifdef PNG_POWERPC_VSX_API_SUPPORTED
+   switch ((pp->options >> PNG_POWERPC_VSX) & 3)
+   {
+      case PNG_OPTION_UNSET:
+         /* Allow the run-time check to execute if it has been enabled -
+          * thus both API and CHECK can be turned on.  If it isn't supported
+          * this case will fall through to the 'default' below, which just
+          * returns.
+          */
+#endif /* PNG_POWERPC_VSX_API_SUPPORTED */
+#ifdef PNG_POWERPC_VSX_CHECK_SUPPORTED
+         {
+            static volatile sig_atomic_t no_vsx = -1; /* not checked */
+
+            if (no_vsx < 0)
+               no_vsx = !png_have_vsx(pp);
+
+            if (no_vsx)
+               return;
+         }
+#ifdef PNG_POWERPC_VSX_API_SUPPORTED
+         break;
+#endif
+#endif /* PNG_POWERPC_VSX_CHECK_SUPPORTED */
+
+#ifdef PNG_POWERPC_VSX_API_SUPPORTED
+      default: /* OFF or INVALID */
+         return;
+
+      case PNG_OPTION_ON:
+         /* Option turned on */
+         break;
+   }
+#endif
+
+   /* IMPORTANT: any new internal functions used here must be declared using
+    * PNG_INTERNAL_FUNCTION in ../pngpriv.h.  This is required so that the
+    * 'prefix' option to configure works:
+    *
+    *    ./configure --with-libpng-prefix=foobar_
+    *
+    * Verify you have got this right by running the above command, doing a build
+    * and examining pngprefix.h; it must contain a #define for every external
+    * function you add.  (Notice that this happens automatically for the
+    * initialization function.)
+    */
+   pp->read_filter[PNG_FILTER_VALUE_UP-1] = png_read_filter_row_up_vsx;
+
+   if (bpp == 3)
+   {
+      pp->read_filter[PNG_FILTER_VALUE_SUB-1] = png_read_filter_row_sub3_vsx;
+      pp->read_filter[PNG_FILTER_VALUE_AVG-1] = png_read_filter_row_avg3_vsx;
+      pp->read_filter[PNG_FILTER_VALUE_PAETH-1] = png_read_filter_row_paeth3_vsx;
+   }
+
+   else if (bpp == 4)
+   {
+      pp->read_filter[PNG_FILTER_VALUE_SUB-1] = png_read_filter_row_sub4_vsx;
+      pp->read_filter[PNG_FILTER_VALUE_AVG-1] = png_read_filter_row_avg4_vsx;
+      pp->read_filter[PNG_FILTER_VALUE_PAETH-1] = png_read_filter_row_paeth4_vsx;
+   }
+}
+#endif /* PNG_POWERPC_VSX_OPT > 0 */
+#endif /* READ */