SIMD-optimized RGB-to-grayscale conversion for MIPS DSPr2

git-svn-id: svn+ssh://svn.code.sf.net/p/libjpeg-turbo/code/trunk@1045 632fc199-4ca6-4c93-a231-07263d6284db

diff --git a/ChangeLog.txt b/ChangeLog.txt
index fde8c4917aa90d228c7d0e94ed265f00880ee36d..c8103e0baf663a8681ca70724401a4922c3cd6e4 100644 (file)
--- a/ChangeLog.txt
+++ b/ChangeLog.txt
@@ -8,7 +8,7 @@ extended to support image scaling.
 
 [2] Added SIMD acceleration for performing color conversion, downsampling,
 and upsampling on DSPr2-capable MIPS platforms.  This speeds up the compression
-of full-color JPEGs by 6-18% on such platforms and decompression by 3-12%.
+of full-color JPEGs by 6-21% on such platforms and decompression by 6-17%.
 
 [3] Added support for 4:1:1 subsampling to the TurboJPEG API.  This is mainly
 included for compatibility, since 4:1:1 is not fully accelerated in

diff --git a/simd/jsimd.h b/simd/jsimd.h
index 907e852a45af8d056b8f38425b2d44ab5d7486ed..807fd7fb4ca065095df37d0ff6335190806c6105 100644 (file)
--- a/simd/jsimd.h
+++ b/simd/jsimd.h
@@ -4,7 +4,7 @@
  * Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
  * Copyright 2011 D. R. Commander
  * Copyright (C) 2013, MIPS Technologies, Inc., California
- * 
+ *
  * Based on the x86 SIMD extension for IJG JPEG library,
  * Copyright (C) 1999-2006, MIYASAKA Masaru.
  * For conditions of distribution and use, see copyright notice in jsimdext.inc
@@ -417,6 +417,35 @@ EXTERN(void) jsimd_extxrgb_ycc_convert_mips_dspr2
              JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
              JDIMENSION output_row, int num_rows));
 
+EXTERN(void) jsimd_rgb_gray_convert_mips_dspr2
+        JPP((JDIMENSION img_width,
+             JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+             JDIMENSION output_row, int num_rows));
+EXTERN(void) jsimd_extrgb_gray_convert_mips_dspr2
+        JPP((JDIMENSION img_width,
+             JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+             JDIMENSION output_row, int num_rows));
+EXTERN(void) jsimd_extrgbx_gray_convert_mips_dspr2
+        JPP((JDIMENSION img_width,
+             JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+             JDIMENSION output_row, int num_rows));
+EXTERN(void) jsimd_extbgr_gray_convert_mips_dspr2
+        JPP((JDIMENSION img_width,
+             JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+             JDIMENSION output_row, int num_rows));
+EXTERN(void) jsimd_extbgrx_gray_convert_mips_dspr2
+        JPP((JDIMENSION img_width,
+             JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+             JDIMENSION output_row, int num_rows));
+EXTERN(void) jsimd_extxbgr_gray_convert_mips_dspr2
+        JPP((JDIMENSION img_width,
+             JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+             JDIMENSION output_row, int num_rows));
+EXTERN(void) jsimd_extxrgb_gray_convert_mips_dspr2
+        JPP((JDIMENSION img_width,
+             JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+             JDIMENSION output_row, int num_rows));
+
 EXTERN (void) jsimd_ycc_rgb_convert_mips_dspr2
         JPP((JDIMENSION img_width,
              JSAMPIMAGE input_buf, JDIMENSION input_row,

diff --git a/simd/jsimd_mips.c b/simd/jsimd_mips.c
index e0cefb0f6a67eb36affce654efd9c4804b625c6a..63edcf6817e49459073c0360d785049c3b148b6d 100644 (file)
--- a/simd/jsimd_mips.c
+++ b/simd/jsimd_mips.c
@@ -4,7 +4,7 @@
  * Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
  * Copyright 2009-2011 D. R. Commander
  * Copyright (C) 2013, MIPS Technologies, Inc., California
- * 
+ *
  * Based on the x86 SIMD extension for IJG JPEG library,
  * Copyright (C) 1999-2006, MIYASAKA Masaru.
  * For conditions of distribution and use, see copyright notice in jsimdext.inc
@@ -100,6 +100,18 @@ jsimd_can_rgb_ycc (void)
 GLOBAL(int)
 jsimd_can_rgb_gray (void)
 {
+  init_simd();
+
+  /* The code is optimised for these values only */
+  if (BITS_IN_JSAMPLE != 8)
+    return 0;
+  if (sizeof(JDIMENSION) != 4)
+    return 0;
+  if ((RGB_PIXELSIZE != 3) && (RGB_PIXELSIZE != 4))
+    return 0;
+  if (simd_support & JSIMD_MIPS_DSPR2)
+    return 1;
+
   return 0;
 }
 
@@ -167,6 +179,40 @@ jsimd_rgb_gray_convert (j_compress_ptr cinfo,
                         JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
                         JDIMENSION output_row, int num_rows)
 {
+  void (*mipsdspr2fct)(JDIMENSION, JSAMPARRAY, JSAMPIMAGE, JDIMENSION, int);
+  switch(cinfo->in_color_space)
+  {
+    case JCS_EXT_RGB:
+      mipsdspr2fct=jsimd_extrgb_gray_convert_mips_dspr2;
+      break;
+    case JCS_EXT_RGBX:
+    case JCS_EXT_RGBA:
+      mipsdspr2fct=jsimd_extrgbx_gray_convert_mips_dspr2;
+      break;
+    case JCS_EXT_BGR:
+      mipsdspr2fct=jsimd_extbgr_gray_convert_mips_dspr2;
+      break;
+    case JCS_EXT_BGRX:
+    case JCS_EXT_BGRA:
+      mipsdspr2fct=jsimd_extbgrx_gray_convert_mips_dspr2;
+      break;
+    case JCS_EXT_XBGR:
+    case JCS_EXT_ABGR:
+      mipsdspr2fct=jsimd_extxbgr_gray_convert_mips_dspr2;
+      break;
+    case JCS_EXT_XRGB:
+    case JCS_EXT_ARGB:
+      mipsdspr2fct=jsimd_extxrgb_gray_convert_mips_dspr2;
+      break;
+    default:
+      mipsdspr2fct=jsimd_extrgb_gray_convert_mips_dspr2;
+      break;
+  }
+
+  if (simd_support & JSIMD_MIPS_DSPR2)
+    mipsdspr2fct(cinfo->image_width, input_buf,
+        output_buf, output_row, num_rows);
+
 }
 
 GLOBAL(void)

diff --git a/simd/jsimd_mips_dspr2.S b/simd/jsimd_mips_dspr2.S
index b7136b52fc74c245946f9ae47f548a0af3de3954..55635b5dc6ab3db93514281dfbd9a14253cfd6f4 100644 (file)
--- a/simd/jsimd_mips_dspr2.S
+++ b/simd/jsimd_mips_dspr2.S
@@ -247,6 +247,134 @@ GENERATE_JSIMD_YCC_RGB_CONVERT_MIPS_DSPR2 extbgrx, 4, 2, 1, 0, 3
 GENERATE_JSIMD_YCC_RGB_CONVERT_MIPS_DSPR2 extxbgr, 4, 3, 2, 1, 0
 GENERATE_JSIMD_YCC_RGB_CONVERT_MIPS_DSPR2 extxrgb, 4, 1, 2, 3, 0
 
+/*****************************************************************************/
+/*
+ * jsimd_extrgb_gray_convert_mips_dspr2
+ * jsimd_extbgr_gray_convert_mips_dspr2
+ * jsimd_extrgbx_gray_convert_mips_dspr2
+ * jsimd_extbgrx_gray_convert_mips_dspr2
+ * jsimd_extxbgr_gray_convert_mips_dspr2
+ * jsimd_extxrgb_gray_convert_mips_dspr2
+ *
+ * Colorspace conversion RGB -> GRAY
+ */
+
+.macro GENERATE_JSIMD_RGB_GRAY_CONVERT_MIPS_DSPR2 colorid, pixel_size, r_offs, g_offs, b_offs
+
+.macro DO_RGB_TO_GRAY r,    \
+                      g,    \
+                      b,    \
+                      inptr
+    lbu     \r, \r_offs(\inptr)
+    lbu     \g, \g_offs(\inptr)
+    lbu     \b, \b_offs(\inptr)
+    addiu   \inptr, \pixel_size
+.endm
+
+LEAF_MIPS_DSPR2(jsimd_\colorid\()_gray_convert_mips_dspr2)
+/*
+ * a0     - cinfo->image_width
+ * a1     - input_buf
+ * a2     - output_buf
+ * a3     - output_row
+ * 16(sp) - num_rows
+ */
+
+    SAVE_REGS_ON_STACK 32, s0, s1, s2, s3, s4, s5, s6, s7
+
+    li      s0, 0x4c8b             // s0 = FIX(0.29900)
+    li      s1, 0x9646             // s1 = FIX(0.58700)
+    li      s2, 0x1d2f             // s2 = FIX(0.11400)
+    li      s7, 0x8000             // s7 = FIX(0.50000)
+    lw      s6, 48(sp)
+    andi    t7, a0, 3
+
+0:
+    addiu   s6, -1                 // s6 = num_rows
+    lw      t0, 0(a1)
+    lw      t1, 0(a2)
+    sll     t3, a3, 2
+    lwx     t1, t3(t1)
+    addiu   a3, 1
+    addu    t9, t1, a0
+    subu    t8, t9, t7
+    beq     t1, t8, 2f
+     nop
+
+1:
+    DO_RGB_TO_GRAY t3, t4, t5, t0
+    DO_RGB_TO_GRAY s3, s4, s5, t0
+
+    mtlo    s7, $ac0
+    maddu   $ac0, s2, t5
+    maddu   $ac0, s1, t4
+    maddu   $ac0, s0, t3
+    mtlo    s7, $ac1
+    maddu   $ac1, s2, s5
+    maddu   $ac1, s1, s4
+    maddu   $ac1, s0, s3
+    extr.w  t6, $ac0, 16
+
+    DO_RGB_TO_GRAY t3, t4, t5, t0
+    DO_RGB_TO_GRAY s3, s4, s5, t0
+
+    mtlo    s7, $ac0
+    maddu   $ac0, s2, t5
+    maddu   $ac0, s1, t4
+    extr.w  t2, $ac1, 16
+    maddu   $ac0, s0, t3
+    mtlo    s7, $ac1
+    maddu   $ac1, s2, s5
+    maddu   $ac1, s1, s4
+    maddu   $ac1, s0, s3
+    extr.w  t5, $ac0, 16
+    sb      t6, 0(t1)
+    sb      t2, 1(t1)
+    extr.w  t3, $ac1, 16
+    addiu   t1, 4
+    sb      t5, -2(t1)
+    sb      t3, -1(t1)
+    bne     t1, t8, 1b
+     nop
+
+2:
+    beqz    t7, 4f
+     nop
+
+3:
+    DO_RGB_TO_GRAY t3, t4, t5, t0
+
+    mtlo    s7, $ac0
+    maddu   $ac0, s2, t5
+    maddu   $ac0, s1, t4
+    maddu   $ac0, s0, t3
+    extr.w  t6, $ac0, 16
+    sb      t6, 0(t1)
+    addiu   t1, 1
+    bne     t1, t9, 3b
+     nop
+
+4:
+    bgtz    s6, 0b
+     addiu  a1, 4
+
+    RESTORE_REGS_FROM_STACK 32, s0, s1, s2, s3, s4, s5, s6, s7
+
+    j ra
+     nop
+END(jsimd_\colorid\()_gray_convert_mips_dspr2)
+
+.purgem DO_RGB_TO_GRAY
+
+.endm
+
+/*------------------------------------------id --  pix R  G  B */
+GENERATE_JSIMD_RGB_GRAY_CONVERT_MIPS_DSPR2 extrgb,  3, 0, 1, 2
+GENERATE_JSIMD_RGB_GRAY_CONVERT_MIPS_DSPR2 extbgr,  3, 2, 1, 0
+GENERATE_JSIMD_RGB_GRAY_CONVERT_MIPS_DSPR2 extrgbx, 4, 0, 1, 2
+GENERATE_JSIMD_RGB_GRAY_CONVERT_MIPS_DSPR2 extbgrx, 4, 2, 1, 0
+GENERATE_JSIMD_RGB_GRAY_CONVERT_MIPS_DSPR2 extxbgr, 4, 3, 2, 1
+GENERATE_JSIMD_RGB_GRAY_CONVERT_MIPS_DSPR2 extxrgb, 4, 1, 2, 3
 /*****************************************************************************/
 /*
  * jsimd_h2v2_fancy_upsample_mips_dspr2