epeg 1.0

data/epeg/vendor/libjpeg-turbo-2.0.4/simd/powerpc/jcgryext-altivec.c

@@ -0,0 +1,228 @@
+/*
+ * AltiVec optimizations for libjpeg-turbo
+ *
+ * Copyright (C) 2014-2015, D. R. Commander.  All Rights Reserved.
+ * Copyright (C) 2014, Jay Foad.  All Rights Reserved.
+ *
+ * This software is provided 'as-is', without any express or implied
+ * warranty.  In no event will the authors be held liable for any damages
+ * arising from the use of this software.
+ *
+ * Permission is granted to anyone to use this software for any purpose,
+ * including commercial applications, and to alter it and redistribute it
+ * freely, subject to the following restrictions:
+ *
+ * 1. The origin of this software must not be misrepresented; you must not
+ *    claim that you wrote the original software. If you use this software
+ *    in a product, an acknowledgment in the product documentation would be
+ *    appreciated but is not required.
+ * 2. Altered source versions must be plainly marked as such, and must not be
+ *    misrepresented as being the original software.
+ * 3. This notice may not be removed or altered from any source distribution.
+ */
+
+/* This file is included by jcgray-altivec.c */
+
+
+void jsimd_rgb_gray_convert_altivec(JDIMENSION img_width, JSAMPARRAY input_buf,
+                                    JSAMPIMAGE output_buf,
+                                    JDIMENSION output_row, int num_rows)
+{
+  JSAMPROW inptr, outptr;
+  int pitch = img_width * RGB_PIXELSIZE, num_cols;
+#if __BIG_ENDIAN__
+  int offset;
+  unsigned char __attribute__((aligned(16))) tmpbuf[RGB_PIXELSIZE * 16];
+#endif
+
+  __vector unsigned char rgb0, rgb1 = { 0 }, rgb2 = { 0 },
+    rgbg0, rgbg1, rgbg2, rgbg3, y;
+#if __BIG_ENDIAN__ || RGB_PIXELSIZE == 4
+  __vector unsigned char rgb3 = { 0 };
+#endif
+#if __BIG_ENDIAN__ && RGB_PIXELSIZE == 4
+  __vector unsigned char rgb4 = { 0 };
+#endif
+  __vector short rg0, rg1, rg2, rg3, bg0, bg1, bg2, bg3;
+  __vector unsigned short yl, yh;
+  __vector int y0, y1, y2, y3;
+
+  /* Constants */
+  __vector short pw_f0299_f0337 = { __4X2(F_0_299, F_0_337) },
+    pw_f0114_f0250 = { __4X2(F_0_114, F_0_250) };
+  __vector int pd_onehalf = { __4X(ONE_HALF) };
+  __vector unsigned char pb_zero = { __16X(0) },
+#if __BIG_ENDIAN__
+    shift_pack_index =
+      { 0, 1, 4, 5,  8,  9, 12, 13, 16, 17, 20, 21, 24, 25, 28, 29 };
+#else
+    shift_pack_index =
+      { 2, 3, 6, 7, 10, 11, 14, 15, 18, 19, 22, 23, 26, 27, 30, 31 };
+#endif
+
+  while (--num_rows >= 0) {
+    inptr = *input_buf++;
+    outptr = output_buf[0][output_row];
+    output_row++;
+
+    for (num_cols = pitch; num_cols > 0;
+         num_cols -= RGB_PIXELSIZE * 16, inptr += RGB_PIXELSIZE * 16,
+         outptr += 16) {
+
+#if __BIG_ENDIAN__
+      /* Load 16 pixels == 48 or 64 bytes */
+      offset = (size_t)inptr & 15;
+      if (offset) {
+        __vector unsigned char unaligned_shift_index;
+        int bytes = num_cols + offset;
+
+        if (bytes < (RGB_PIXELSIZE + 1) * 16 && (bytes & 15)) {
+          /* Slow path to prevent buffer overread.  Since there is no way to
+           * read a partial AltiVec register, overread would occur on the last
+           * chunk of the last image row if the right edge is not on a 16-byte
+           * boundary.  It could also occur on other rows if the bytes per row
+           * is low enough.  Since we can't determine whether we're on the last
+           * image row, we have to assume every row is the last.
+           */
+          memcpy(tmpbuf, inptr, min(num_cols, RGB_PIXELSIZE * 16));
+          rgb0 = vec_ld(0, tmpbuf);
+          rgb1 = vec_ld(16, tmpbuf);
+          rgb2 = vec_ld(32, tmpbuf);
+#if RGB_PIXELSIZE == 4
+          rgb3 = vec_ld(48, tmpbuf);
+#endif
+        } else {
+          /* Fast path */
+          rgb0 = vec_ld(0, inptr);
+          if (bytes > 16)
+            rgb1 = vec_ld(16, inptr);
+          if (bytes > 32)
+            rgb2 = vec_ld(32, inptr);
+          if (bytes > 48)
+            rgb3 = vec_ld(48, inptr);
+#if RGB_PIXELSIZE == 4
+          if (bytes > 64)
+            rgb4 = vec_ld(64, inptr);
+#endif
+          unaligned_shift_index = vec_lvsl(0, inptr);
+          rgb0 = vec_perm(rgb0, rgb1, unaligned_shift_index);
+          rgb1 = vec_perm(rgb1, rgb2, unaligned_shift_index);
+          rgb2 = vec_perm(rgb2, rgb3, unaligned_shift_index);
+#if RGB_PIXELSIZE == 4
+          rgb3 = vec_perm(rgb3, rgb4, unaligned_shift_index);
+#endif
+        }
+      } else {
+        if (num_cols < RGB_PIXELSIZE * 16 && (num_cols & 15)) {
+          /* Slow path */
+          memcpy(tmpbuf, inptr, min(num_cols, RGB_PIXELSIZE * 16));
+          rgb0 = vec_ld(0, tmpbuf);
+          rgb1 = vec_ld(16, tmpbuf);
+          rgb2 = vec_ld(32, tmpbuf);
+#if RGB_PIXELSIZE == 4
+          rgb3 = vec_ld(48, tmpbuf);
+#endif
+        } else {
+          /* Fast path */
+          rgb0 = vec_ld(0, inptr);
+          if (num_cols > 16)
+            rgb1 = vec_ld(16, inptr);
+          if (num_cols > 32)
+            rgb2 = vec_ld(32, inptr);
+#if RGB_PIXELSIZE == 4
+          if (num_cols > 48)
+            rgb3 = vec_ld(48, inptr);
+#endif
+        }
+      }
+#else
+      /* Little endian */
+      rgb0 = vec_vsx_ld(0, inptr);
+      if (num_cols > 16)
+        rgb1 = vec_vsx_ld(16, inptr);
+      if (num_cols > 32)
+        rgb2 = vec_vsx_ld(32, inptr);
+#if RGB_PIXELSIZE == 4
+      if (num_cols > 48)
+        rgb3 = vec_vsx_ld(48, inptr);
+#endif
+#endif
+
+#if RGB_PIXELSIZE == 3
+      /* rgb0 = R0 G0 B0 R1 G1 B1 R2 G2 B2 R3 G3 B3 R4 G4 B4 R5
+       * rgb1 = G5 B5 R6 G6 B6 R7 G7 B7 R8 G8 B8 R9 G9 B9 Ra Ga
+       * rgb2 = Ba Rb Gb Bb Rc Gc Bc Rd Gd Bd Re Ge Be Rf Gf Bf
+       *
+       * rgbg0 = R0 G0 R1 G1 R2 G2 R3 G3 B0 G0 B1 G1 B2 G2 B3 G3
+       * rgbg1 = R4 G4 R5 G5 R6 G6 R7 G7 B4 G4 B5 G5 B6 G6 B7 G7
+       * rgbg2 = R8 G8 R9 G9 Ra Ga Rb Gb B8 G8 B9 G9 Ba Ga Bb Gb
+       * rgbg3 = Rc Gc Rd Gd Re Ge Rf Gf Bc Gc Bd Gd Be Ge Bf Gf
+       */
+      rgbg0 = vec_perm(rgb0, rgb0, (__vector unsigned char)RGBG_INDEX0);
+      rgbg1 = vec_perm(rgb0, rgb1, (__vector unsigned char)RGBG_INDEX1);
+      rgbg2 = vec_perm(rgb1, rgb2, (__vector unsigned char)RGBG_INDEX2);
+      rgbg3 = vec_perm(rgb2, rgb2, (__vector unsigned char)RGBG_INDEX3);
+#else
+      /* rgb0 = R0 G0 B0 X0 R1 G1 B1 X1 R2 G2 B2 X2 R3 G3 B3 X3
+       * rgb1 = R4 G4 B4 X4 R5 G5 B5 X5 R6 G6 B6 X6 R7 G7 B7 X7
+       * rgb2 = R8 G8 B8 X8 R9 G9 B9 X9 Ra Ga Ba Xa Rb Gb Bb Xb
+       * rgb3 = Rc Gc Bc Xc Rd Gd Bd Xd Re Ge Be Xe Rf Gf Bf Xf
+       *
+       * rgbg0 = R0 G0 R1 G1 R2 G2 R3 G3 B0 G0 B1 G1 B2 G2 B3 G3
+       * rgbg1 = R4 G4 R5 G5 R6 G6 R7 G7 B4 G4 B5 G5 B6 G6 B7 G7
+       * rgbg2 = R8 G8 R9 G9 Ra Ga Rb Gb B8 G8 B9 G9 Ba Ga Bb Gb
+       * rgbg3 = Rc Gc Rd Gd Re Ge Rf Gf Bc Gc Bd Gd Be Ge Bf Gf
+       */
+      rgbg0 = vec_perm(rgb0, rgb0, (__vector unsigned char)RGBG_INDEX);
+      rgbg1 = vec_perm(rgb1, rgb1, (__vector unsigned char)RGBG_INDEX);
+      rgbg2 = vec_perm(rgb2, rgb2, (__vector unsigned char)RGBG_INDEX);
+      rgbg3 = vec_perm(rgb3, rgb3, (__vector unsigned char)RGBG_INDEX);
+#endif
+
+      /* rg0 = R0 G0 R1 G1 R2 G2 R3 G3
+       * bg0 = B0 G0 B1 G1 B2 G2 B3 G3
+       * ...
+       *
+       * NOTE: We have to use vec_merge*() here because vec_unpack*() doesn't
+       * support unsigned vectors.
+       */
+      rg0 = (__vector signed short)VEC_UNPACKHU(rgbg0);
+      bg0 = (__vector signed short)VEC_UNPACKLU(rgbg0);
+      rg1 = (__vector signed short)VEC_UNPACKHU(rgbg1);
+      bg1 = (__vector signed short)VEC_UNPACKLU(rgbg1);
+      rg2 = (__vector signed short)VEC_UNPACKHU(rgbg2);
+      bg2 = (__vector signed short)VEC_UNPACKLU(rgbg2);
+      rg3 = (__vector signed short)VEC_UNPACKHU(rgbg3);
+      bg3 = (__vector signed short)VEC_UNPACKLU(rgbg3);
+
+      /* (Original)
+       * Y  =  0.29900 * R + 0.58700 * G + 0.11400 * B
+       *
+       * (This implementation)
+       * Y  =  0.29900 * R + 0.33700 * G + 0.11400 * B + 0.25000 * G
+       */
+
+      /* Calculate Y values */
+
+      y0 = vec_msums(rg0, pw_f0299_f0337, pd_onehalf);
+      y1 = vec_msums(rg1, pw_f0299_f0337, pd_onehalf);
+      y2 = vec_msums(rg2, pw_f0299_f0337, pd_onehalf);
+      y3 = vec_msums(rg3, pw_f0299_f0337, pd_onehalf);
+      y0 = vec_msums(bg0, pw_f0114_f0250, y0);
+      y1 = vec_msums(bg1, pw_f0114_f0250, y1);
+      y2 = vec_msums(bg2, pw_f0114_f0250, y2);
+      y3 = vec_msums(bg3, pw_f0114_f0250, y3);
+      /* Clever way to avoid 4 shifts + 2 packs.  This packs the high word from
+       * each dword into a new 16-bit vector, which is the equivalent of
+       * descaling the 32-bit results (right-shifting by 16 bits) and then
+       * packing them.
+       */
+      yl = vec_perm((__vector unsigned short)y0, (__vector unsigned short)y1,
+                    shift_pack_index);
+      yh = vec_perm((__vector unsigned short)y2, (__vector unsigned short)y3,
+                    shift_pack_index);
+      y = vec_pack(yl, yh);
+      vec_st(y, 0, outptr);
+    }
+  }
+}

data/epeg/vendor/libjpeg-turbo-2.0.4/simd/powerpc/jcsample-altivec.c

@@ -0,0 +1,159 @@
+/*
+ * AltiVec optimizations for libjpeg-turbo
+ *
+ * Copyright (C) 2015, D. R. Commander.  All Rights Reserved.
+ *
+ * This software is provided 'as-is', without any express or implied
+ * warranty.  In no event will the authors be held liable for any damages
+ * arising from the use of this software.
+ *
+ * Permission is granted to anyone to use this software for any purpose,
+ * including commercial applications, and to alter it and redistribute it
+ * freely, subject to the following restrictions:
+ *
+ * 1. The origin of this software must not be misrepresented; you must not
+ *    claim that you wrote the original software. If you use this software
+ *    in a product, an acknowledgment in the product documentation would be
+ *    appreciated but is not required.
+ * 2. Altered source versions must be plainly marked as such, and must not be
+ *    misrepresented as being the original software.
+ * 3. This notice may not be removed or altered from any source distribution.
+ */
+
+/* CHROMA DOWNSAMPLING */
+
+#include "jsimd_altivec.h"
+#include "jcsample.h"
+
+
+void jsimd_h2v1_downsample_altivec(JDIMENSION image_width,
+                                   int max_v_samp_factor,
+                                   JDIMENSION v_samp_factor,
+                                   JDIMENSION width_in_blocks,
+                                   JSAMPARRAY input_data,
+                                   JSAMPARRAY output_data)
+{
+  int outrow, outcol;
+  JDIMENSION output_cols = width_in_blocks * DCTSIZE;
+  JSAMPROW inptr, outptr;
+
+  __vector unsigned char this0, next0, out;
+  __vector unsigned short this0e, this0o, next0e, next0o, outl, outh;
+
+  /* Constants */
+  __vector unsigned short pw_bias = { __4X2(0, 1) },
+    pw_one = { __8X(1) };
+  __vector unsigned char even_odd_index =
+    {  0,  2,  4,  6,  8, 10, 12, 14,  1,  3,  5,  7,  9, 11, 13, 15 },
+    pb_zero = { __16X(0) };
+
+  expand_right_edge(input_data, max_v_samp_factor, image_width,
+                    output_cols * 2);
+
+  for (outrow = 0; outrow < v_samp_factor; outrow++) {
+    outptr = output_data[outrow];
+    inptr = input_data[outrow];
+
+    for (outcol = output_cols; outcol > 0;
+         outcol -= 16, inptr += 32, outptr += 16) {
+
+      this0 = vec_ld(0, inptr);
+      this0 = vec_perm(this0, this0, even_odd_index);
+      this0e = (__vector unsigned short)VEC_UNPACKHU(this0);
+      this0o = (__vector unsigned short)VEC_UNPACKLU(this0);
+      outl = vec_add(this0e, this0o);
+      outl = vec_add(outl, pw_bias);
+      outl = vec_sr(outl, pw_one);
+
+      if (outcol > 8) {
+        next0 = vec_ld(16, inptr);
+        next0 = vec_perm(next0, next0, even_odd_index);
+        next0e = (__vector unsigned short)VEC_UNPACKHU(next0);
+        next0o = (__vector unsigned short)VEC_UNPACKLU(next0);
+        outh = vec_add(next0e, next0o);
+        outh = vec_add(outh, pw_bias);
+        outh = vec_sr(outh, pw_one);
+      } else
+        outh = vec_splat_u16(0);
+
+      out = vec_pack(outl, outh);
+      vec_st(out, 0, outptr);
+    }
+  }
+}
+
+
+void
+jsimd_h2v2_downsample_altivec(JDIMENSION image_width, int max_v_samp_factor,
+                              JDIMENSION v_samp_factor,
+                              JDIMENSION width_in_blocks,
+                              JSAMPARRAY input_data, JSAMPARRAY output_data)
+{
+  int inrow, outrow, outcol;
+  JDIMENSION output_cols = width_in_blocks * DCTSIZE;
+  JSAMPROW inptr0, inptr1, outptr;
+
+  __vector unsigned char this0, next0, this1, next1, out;
+  __vector unsigned short this0e, this0o, next0e, next0o, this1e, this1o,
+    next1e, next1o, out0l, out0h, out1l, out1h, outl, outh;
+
+  /* Constants */
+  __vector unsigned short pw_bias = { __4X2(1, 2) },
+    pw_two = { __8X(2) };
+  __vector unsigned char even_odd_index =
+    {  0,  2,  4,  6,  8, 10, 12, 14,  1,  3,  5,  7,  9, 11, 13, 15 },
+    pb_zero = { __16X(0) };
+
+  expand_right_edge(input_data, max_v_samp_factor, image_width,
+                    output_cols * 2);
+
+  for (inrow = 0, outrow = 0; outrow < v_samp_factor;
+       inrow += 2, outrow++) {
+
+    inptr0 = input_data[inrow];
+    inptr1 = input_data[inrow + 1];
+    outptr = output_data[outrow];
+
+    for (outcol = output_cols; outcol > 0;
+         outcol -= 16, inptr0 += 32, inptr1 += 32, outptr += 16) {
+
+      this0 = vec_ld(0, inptr0);
+      this0 = vec_perm(this0, this0, even_odd_index);
+      this0e = (__vector unsigned short)VEC_UNPACKHU(this0);
+      this0o = (__vector unsigned short)VEC_UNPACKLU(this0);
+      out0l = vec_add(this0e, this0o);
+
+      this1 = vec_ld(0, inptr1);
+      this1 = vec_perm(this1, this1, even_odd_index);
+      this1e = (__vector unsigned short)VEC_UNPACKHU(this1);
+      this1o = (__vector unsigned short)VEC_UNPACKLU(this1);
+      out1l = vec_add(this1e, this1o);
+
+      outl = vec_add(out0l, out1l);
+      outl = vec_add(outl, pw_bias);
+      outl = vec_sr(outl, pw_two);
+
+      if (outcol > 8) {
+        next0 = vec_ld(16, inptr0);
+        next0 = vec_perm(next0, next0, even_odd_index);
+        next0e = (__vector unsigned short)VEC_UNPACKHU(next0);
+        next0o = (__vector unsigned short)VEC_UNPACKLU(next0);
+        out0h = vec_add(next0e, next0o);
+
+        next1 = vec_ld(16, inptr1);
+        next1 = vec_perm(next1, next1, even_odd_index);
+        next1e = (__vector unsigned short)VEC_UNPACKHU(next1);
+        next1o = (__vector unsigned short)VEC_UNPACKLU(next1);
+        out1h = vec_add(next1e, next1o);
+
+        outh = vec_add(out0h, out1h);
+        outh = vec_add(outh, pw_bias);
+        outh = vec_sr(outh, pw_two);
+      } else
+        outh = vec_splat_u16(0);
+
+      out = vec_pack(outl, outh);
+      vec_st(out, 0, outptr);
+    }
+  }
+}

data/epeg/vendor/libjpeg-turbo-2.0.4/simd/powerpc/jcsample.h

@@ -0,0 +1,28 @@
+/*
+ * jcsample.h
+ *
+ * This file was part of the Independent JPEG Group's software:
+ * Copyright (C) 1991-1996, Thomas G. Lane.
+ * For conditions of distribution and use, see the accompanying README.ijg
+ * file.
+ */
+
+LOCAL(void)
+expand_right_edge(JSAMPARRAY image_data, int num_rows, JDIMENSION input_cols,
+                  JDIMENSION output_cols)
+{
+  register JSAMPROW ptr;
+  register JSAMPLE pixval;
+  register int count;
+  int row;
+  int numcols = (int)(output_cols - input_cols);
+
+  if (numcols > 0) {
+    for (row = 0; row < num_rows; row++) {
+      ptr = image_data[row] + input_cols;
+      pixval = ptr[-1];         /* don't need GETJSAMPLE() here */
+      for (count = numcols; count > 0; count--)
+        *ptr++ = pixval;
+    }
+  }
+}

data/epeg/vendor/libjpeg-turbo-2.0.4/simd/powerpc/jdcolext-altivec.c

@@ -0,0 +1,276 @@
+/*
+ * AltiVec optimizations for libjpeg-turbo
+ *
+ * Copyright (C) 2015, D. R. Commander.  All Rights Reserved.
+ *
+ * This software is provided 'as-is', without any express or implied
+ * warranty.  In no event will the authors be held liable for any damages
+ * arising from the use of this software.
+ *
+ * Permission is granted to anyone to use this software for any purpose,
+ * including commercial applications, and to alter it and redistribute it
+ * freely, subject to the following restrictions:
+ *
+ * 1. The origin of this software must not be misrepresented; you must not
+ *    claim that you wrote the original software. If you use this software
+ *    in a product, an acknowledgment in the product documentation would be
+ *    appreciated but is not required.
+ * 2. Altered source versions must be plainly marked as such, and must not be
+ *    misrepresented as being the original software.
+ * 3. This notice may not be removed or altered from any source distribution.
+ */
+
+/* This file is included by jdcolor-altivec.c */
+
+
+void jsimd_ycc_rgb_convert_altivec(JDIMENSION out_width, JSAMPIMAGE input_buf,
+                                   JDIMENSION input_row, JSAMPARRAY output_buf,
+                                   int num_rows)
+{
+  JSAMPROW outptr, inptr0, inptr1, inptr2;
+  int pitch = out_width * RGB_PIXELSIZE, num_cols;
+#if __BIG_ENDIAN__
+  int offset;
+#endif
+  unsigned char __attribute__((aligned(16))) tmpbuf[RGB_PIXELSIZE * 16];
+
+  __vector unsigned char rgb0, rgb1, rgb2, rgbx0, rgbx1, rgbx2, rgbx3,
+    y, cb, cr;
+#if __BIG_ENDIAN__
+  __vector unsigned char edgel, edgeh, edges, out0, out1, out2, out3;
+#if RGB_PIXELSIZE == 4
+  __vector unsigned char out4;
+#endif
+#endif
+#if RGB_PIXELSIZE == 4
+  __vector unsigned char rgb3;
+#endif
+  __vector short rg0, rg1, rg2, rg3, bx0, bx1, bx2, bx3, yl, yh, cbl, cbh,
+    crl, crh, rl, rh, gl, gh, bl, bh, g0w, g1w, g2w, g3w;
+  __vector int g0, g1, g2, g3;
+
+  /* Constants
+   * NOTE: The >> 1 is to compensate for the fact that vec_madds() returns 17
+   * high-order bits, not 16.
+   */
+  __vector short pw_f0402 = { __8X(F_0_402 >> 1) },
+    pw_mf0228 = { __8X(-F_0_228 >> 1) },
+    pw_mf0344_f0285 = { __4X2(-F_0_344, F_0_285) },
+    pw_one = { __8X(1) }, pw_255 = { __8X(255) },
+    pw_cj = { __8X(CENTERJSAMPLE) };
+  __vector int pd_onehalf = { __4X(ONE_HALF) };
+  __vector unsigned char pb_zero = { __16X(0) },
+#if __BIG_ENDIAN__
+    shift_pack_index =
+      {  0,  1,  4,  5,  8,  9, 12, 13, 16, 17, 20, 21, 24, 25, 28, 29 };
+#else
+    shift_pack_index =
+      {  2,  3,  6,  7, 10, 11, 14, 15, 18, 19, 22, 23, 26, 27, 30, 31 };
+#endif
+
+  while (--num_rows >= 0) {
+    inptr0 = input_buf[0][input_row];
+    inptr1 = input_buf[1][input_row];
+    inptr2 = input_buf[2][input_row];
+    input_row++;
+    outptr = *output_buf++;
+
+    for (num_cols = pitch; num_cols > 0;
+         num_cols -= RGB_PIXELSIZE * 16, outptr += RGB_PIXELSIZE * 16,
+         inptr0 += 16, inptr1 += 16, inptr2 += 16) {
+
+      y = vec_ld(0, inptr0);
+      /* NOTE: We have to use vec_merge*() here because vec_unpack*() doesn't
+       * support unsigned vectors.
+       */
+      yl = (__vector signed short)VEC_UNPACKHU(y);
+      yh = (__vector signed short)VEC_UNPACKLU(y);
+
+      cb = vec_ld(0, inptr1);
+      cbl = (__vector signed short)VEC_UNPACKHU(cb);
+      cbh = (__vector signed short)VEC_UNPACKLU(cb);
+      cbl = vec_sub(cbl, pw_cj);
+      cbh = vec_sub(cbh, pw_cj);
+
+      cr = vec_ld(0, inptr2);
+      crl = (__vector signed short)VEC_UNPACKHU(cr);
+      crh = (__vector signed short)VEC_UNPACKLU(cr);
+      crl = vec_sub(crl, pw_cj);
+      crh = vec_sub(crh, pw_cj);
+
+      /* (Original)
+       * R = Y                + 1.40200 * Cr
+       * G = Y - 0.34414 * Cb - 0.71414 * Cr
+       * B = Y + 1.77200 * Cb
+       *
+       * (This implementation)
+       * R = Y                + 0.40200 * Cr + Cr
+       * G = Y - 0.34414 * Cb + 0.28586 * Cr - Cr
+       * B = Y - 0.22800 * Cb + Cb + Cb
+       */
+      bl = vec_add(cbl, cbl);
+      bh = vec_add(cbh, cbh);
+      bl = vec_madds(bl, pw_mf0228, pw_one);
+      bh = vec_madds(bh, pw_mf0228, pw_one);
+      bl = vec_sra(bl, (__vector unsigned short)pw_one);
+      bh = vec_sra(bh, (__vector unsigned short)pw_one);
+      bl = vec_add(bl, cbl);
+      bh = vec_add(bh, cbh);
+      bl = vec_add(bl, cbl);
+      bh = vec_add(bh, cbh);
+      bl = vec_add(bl, yl);
+      bh = vec_add(bh, yh);
+
+      rl = vec_add(crl, crl);
+      rh = vec_add(crh, crh);
+      rl = vec_madds(rl, pw_f0402, pw_one);
+      rh = vec_madds(rh, pw_f0402, pw_one);
+      rl = vec_sra(rl, (__vector unsigned short)pw_one);
+      rh = vec_sra(rh, (__vector unsigned short)pw_one);
+      rl = vec_add(rl, crl);
+      rh = vec_add(rh, crh);
+      rl = vec_add(rl, yl);
+      rh = vec_add(rh, yh);
+
+      g0w = vec_mergeh(cbl, crl);
+      g1w = vec_mergel(cbl, crl);
+      g0 = vec_msums(g0w, pw_mf0344_f0285, pd_onehalf);
+      g1 = vec_msums(g1w, pw_mf0344_f0285, pd_onehalf);
+      g2w = vec_mergeh(cbh, crh);
+      g3w = vec_mergel(cbh, crh);
+      g2 = vec_msums(g2w, pw_mf0344_f0285, pd_onehalf);
+      g3 = vec_msums(g3w, pw_mf0344_f0285, pd_onehalf);
+      /* Clever way to avoid 4 shifts + 2 packs.  This packs the high word from
+       * each dword into a new 16-bit vector, which is the equivalent of
+       * descaling the 32-bit results (right-shifting by 16 bits) and then
+       * packing them.
+       */
+      gl = vec_perm((__vector short)g0, (__vector short)g1, shift_pack_index);
+      gh = vec_perm((__vector short)g2, (__vector short)g3, shift_pack_index);
+      gl = vec_sub(gl, crl);
+      gh = vec_sub(gh, crh);
+      gl = vec_add(gl, yl);
+      gh = vec_add(gh, yh);
+
+      rg0 = vec_mergeh(rl, gl);
+      bx0 = vec_mergeh(bl, pw_255);
+      rg1 = vec_mergel(rl, gl);
+      bx1 = vec_mergel(bl, pw_255);
+      rg2 = vec_mergeh(rh, gh);
+      bx2 = vec_mergeh(bh, pw_255);
+      rg3 = vec_mergel(rh, gh);
+      bx3 = vec_mergel(bh, pw_255);
+
+      rgbx0 = vec_packsu(rg0, bx0);
+      rgbx1 = vec_packsu(rg1, bx1);
+      rgbx2 = vec_packsu(rg2, bx2);
+      rgbx3 = vec_packsu(rg3, bx3);
+
+#if RGB_PIXELSIZE == 3
+      /* rgbx0 = R0 G0 R1 G1 R2 G2 R3 G3 B0 X0 B1 X1 B2 X2 B3 X3
+       * rgbx1 = R4 G4 R5 G5 R6 G6 R7 G7 B4 X4 B5 X5 B6 X6 B7 X7
+       * rgbx2 = R8 G8 R9 G9 Ra Ga Rb Gb B8 X8 B9 X9 Ba Xa Bb Xb
+       * rgbx3 = Rc Gc Rd Gd Re Ge Rf Gf Bc Xc Bd Xd Be Xe Bf Xf
+       *
+       * rgb0 = R0 G0 B0 R1 G1 B1 R2 G2 B2 R3 G3 B3 R4 G4 B4 R5
+       * rgb1 = G5 B5 R6 G6 B6 R7 G7 B7 R8 G8 B8 R9 G9 B9 Ra Ga
+       * rgb2 = Ba Rb Gb Bb Rc Gc Bc Rd Gd Bd Re Ge Be Rf Gf Bf
+       */
+      rgb0 = vec_perm(rgbx0, rgbx1, (__vector unsigned char)RGB_INDEX0);
+      rgb1 = vec_perm(rgbx1, rgbx2, (__vector unsigned char)RGB_INDEX1);
+      rgb2 = vec_perm(rgbx2, rgbx3, (__vector unsigned char)RGB_INDEX2);
+#else
+      /* rgbx0 = R0 G0 R1 G1 R2 G2 R3 G3 B0 X0 B1 X1 B2 X2 B3 X3
+       * rgbx1 = R4 G4 R5 G5 R6 G6 R7 G7 B4 X4 B5 X5 B6 X6 B7 X7
+       * rgbx2 = R8 G8 R9 G9 Ra Ga Rb Gb B8 X8 B9 X9 Ba Xa Bb Xb
+       * rgbx3 = Rc Gc Rd Gd Re Ge Rf Gf Bc Xc Bd Xd Be Xe Bf Xf
+       *
+       * rgb0 = R0 G0 B0 X0 R1 G1 B1 X1 R2 G2 B2 X2 R3 G3 B3 X3
+       * rgb1 = R4 G4 B4 X4 R5 G5 B5 X5 R6 G6 B6 X6 R7 G7 B7 X7
+       * rgb2 = R8 G8 B8 X8 R9 G9 B9 X9 Ra Ga Ba Xa Rb Gb Bb Xb
+       * rgb3 = Rc Gc Bc Xc Rd Gd Bd Xd Re Ge Be Xe Rf Gf Bf Xf
+       */
+      rgb0 = vec_perm(rgbx0, rgbx0, (__vector unsigned char)RGB_INDEX);
+      rgb1 = vec_perm(rgbx1, rgbx1, (__vector unsigned char)RGB_INDEX);
+      rgb2 = vec_perm(rgbx2, rgbx2, (__vector unsigned char)RGB_INDEX);
+      rgb3 = vec_perm(rgbx3, rgbx3, (__vector unsigned char)RGB_INDEX);
+#endif
+
+#if __BIG_ENDIAN__
+      offset = (size_t)outptr & 15;
+      if (offset) {
+        __vector unsigned char unaligned_shift_index;
+        int bytes = num_cols + offset;
+
+        if (bytes < (RGB_PIXELSIZE + 1) * 16 && (bytes & 15)) {
+          /* Slow path to prevent buffer overwrite.  Since there is no way to
+           * write a partial AltiVec register, overwrite would occur on the
+           * last chunk of the last image row if the right edge is not on a
+           * 16-byte boundary.  It could also occur on other rows if the bytes
+           * per row is low enough.  Since we can't determine whether we're on
+           * the last image row, we have to assume every row is the last.
+           */
+          vec_st(rgb0, 0, tmpbuf);
+          vec_st(rgb1, 16, tmpbuf);
+          vec_st(rgb2, 32, tmpbuf);
+#if RGB_PIXELSIZE == 4
+          vec_st(rgb3, 48, tmpbuf);
+#endif
+          memcpy(outptr, tmpbuf, min(num_cols, RGB_PIXELSIZE * 16));
+        } else {
+          /* Fast path */
+          unaligned_shift_index = vec_lvsl(0, outptr);
+          edgel = vec_ld(0, outptr);
+          edgeh = vec_ld(min(num_cols - 1, RGB_PIXELSIZE * 16), outptr);
+          edges = vec_perm(edgeh, edgel, unaligned_shift_index);
+          unaligned_shift_index = vec_lvsr(0, outptr);
+          out0 = vec_perm(edges, rgb0, unaligned_shift_index);
+          out1 = vec_perm(rgb0, rgb1, unaligned_shift_index);
+          out2 = vec_perm(rgb1, rgb2, unaligned_shift_index);
+#if RGB_PIXELSIZE == 4
+          out3 = vec_perm(rgb2, rgb3, unaligned_shift_index);
+          out4 = vec_perm(rgb3, edges, unaligned_shift_index);
+#else
+          out3 = vec_perm(rgb2, edges, unaligned_shift_index);
+#endif
+          vec_st(out0, 0, outptr);
+          if (bytes > 16)
+            vec_st(out1, 16, outptr);
+          if (bytes > 32)
+            vec_st(out2, 32, outptr);
+          if (bytes > 48)
+            vec_st(out3, 48, outptr);
+#if RGB_PIXELSIZE == 4
+          if (bytes > 64)
+            vec_st(out4, 64, outptr);
+#endif
+        }
+      } else {
+#endif /* __BIG_ENDIAN__ */
+        if (num_cols < RGB_PIXELSIZE * 16 && (num_cols & 15)) {
+          /* Slow path */
+          VEC_ST(rgb0, 0, tmpbuf);
+          VEC_ST(rgb1, 16, tmpbuf);
+          VEC_ST(rgb2, 32, tmpbuf);
+#if RGB_PIXELSIZE == 4
+          VEC_ST(rgb3, 48, tmpbuf);
+#endif
+          memcpy(outptr, tmpbuf, min(num_cols, RGB_PIXELSIZE * 16));
+        } else {
+          /* Fast path */
+          VEC_ST(rgb0, 0, outptr);
+          if (num_cols > 16)
+            VEC_ST(rgb1, 16, outptr);
+          if (num_cols > 32)
+            VEC_ST(rgb2, 32, outptr);
+#if RGB_PIXELSIZE == 4
+          if (num_cols > 48)
+            VEC_ST(rgb3, 48, outptr);
+#endif
+        }
+#if __BIG_ENDIAN__
+      }
+#endif
+    }
+  }
+}