ARM NEON optimized version of 'ycc_rgb_convert'

Is approximately 6x faster than original C variant.

1 files changed, 175 insertions, 0 deletions

diff --git a/jdcolor.c b/jdcolor.c
index e02ea4f..079240c 100644
--- a/jdcolor.c
+++ b/jdcolor.c
@@ -4,6 +4,11 @@
  * Copyright (C) 1991-1997, Thomas G. Lane.
  * Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
  * Copyright (C) 2009, D. R. Commander.
+ *
+ * ARM NEON optimizations
+ * Copyright (C) 2010 Nokia Corporation and/or its subsidiary(-ies). All rights reserved.
+ * Contact: Alexander Bokovoy <alexander.bokovoy@nokia.com>
+ *
  * This file is part of the Independent JPEG Group's software.
  * For conditions of distribution and use, see the accompanying README file.
  *
@@ -119,6 +124,169 @@ build_ycc_rgb_table (j_decompress_ptr cinfo)
  * offset required on that side.
  */
 
+#if defined(__ARM_NEON__) && defined(WITH_SIMD) && \
+    (RGB_PIXELSIZE == 3) && (BITS_IN_JSAMPLE == 8)
+
+LOCAL(void)
+yuv444_to_rgb24_neon(unsigned char *rgb,
+		     unsigned char *y, unsigned char *u, unsigned char *v,
+		     int n)
+{
+    const static signed short c[12] = {
+        22971, -11277, -23401, 29033,
+        -128, -128, -128, -128, -128, -128, -128, -128
+    };
+    asm volatile (
+        ".fpu neon\n"
+        ".macro do_yuv_to_rgb\n"
+        "   vaddw.u8    q3, q1, d4\n"     /* q3 = u - 128 */
+        "   vaddw.u8    q4, q1, d5\n"     /* q2 = v - 128 */
+        "   vmull.s16   q10, d6, d1[1]\n" /* multiply by -11277 */
+        "   vmull.s16   q11, d7, d1[1]\n" /* multiply by -11277 */
+        "   vmlal.s16   q10, d8, d1[2]\n" /* multiply by -23401 */
+        "   vmlal.s16   q11, d9, d1[2]\n" /* multiply by -23401 */
+        "   vmull.s16   q12, d8, d1[0]\n" /* multiply by 22971 */
+        "   vmull.s16   q13, d9, d1[0]\n" /* multiply by 22971 */
+        "   vmull.s16   q14, d6, d1[3]\n" /* multiply by 29033 */
+        "   vmull.s16   q15, d7, d1[3]\n" /* multiply by 29033 */
+        "   vrshrn.s32  d20, q10, #15\n"
+        "   vrshrn.s32  d21, q11, #15\n"
+        "   vrshrn.s32  d24, q12, #14\n"
+        "   vrshrn.s32  d25, q13, #14\n"
+        "   vrshrn.s32  d28, q14, #14\n"
+        "   vrshrn.s32  d29, q15, #14\n"
+        "   vaddw.u8    q10, q10, d0\n"
+        "   vaddw.u8    q12, q12, d0\n"
+        "   vaddw.u8    q14, q14, d0\n"
+        "   vqmovun.s16 d11, q10\n"
+        "   vqmovun.s16 d10, q12\n"
+        "   vqmovun.s16 d12, q14\n"
+        ".endm\n"
+        ".macro do_load size\n"
+            ".if \\size == 8\n"
+                "vld1.8  {d4}, [%[u]]!\n"
+                "vld1.8  {d5}, [%[v]]!\n"
+                "vld1.8  {d0}, [%[y]]!\n"
+                "pld [%[y], #64]\n"
+                "pld [%[u], #64]\n"
+                "pld [%[v], #64]\n"
+            ".elseif \\size == 4\n"
+                "vld1.8  {d4[0]}, [%[u]]!\n"
+                "vld1.8  {d4[1]}, [%[u]]!\n"
+                "vld1.8  {d4[2]}, [%[u]]!\n"
+                "vld1.8  {d4[3]}, [%[u]]!\n"
+                "vld1.8  {d5[0]}, [%[v]]!\n"
+                "vld1.8  {d5[1]}, [%[v]]!\n"
+                "vld1.8  {d5[2]}, [%[v]]!\n"
+                "vld1.8  {d5[3]}, [%[v]]!\n"
+                "vld1.8  {d0[0]}, [%[y]]!\n"
+                "vld1.8  {d0[1]}, [%[y]]!\n"
+                "vld1.8  {d0[2]}, [%[y]]!\n"
+                "vld1.8  {d0[3]}, [%[y]]!\n"
+            ".elseif \\size == 2\n"
+                "vld1.8  {d4[4]}, [%[u]]!\n"
+                "vld1.8  {d4[5]}, [%[u]]!\n"
+                "vld1.8  {d5[4]}, [%[v]]!\n"
+                "vld1.8  {d5[5]}, [%[v]]!\n"
+                "vld1.8  {d0[4]}, [%[y]]!\n"
+                "vld1.8  {d0[5]}, [%[y]]!\n"
+            ".elseif \\size == 1\n"
+                "vld1.8  {d4[6]}, [%[u]]!\n"
+                "vld1.8  {d5[6]}, [%[v]]!\n"
+                "vld1.8  {d0[6]}, [%[y]]!\n"
+            ".else\n"
+                ".error \"unsupported macroblock size\"\n"
+            ".endif\n"
+        ".endm\n"
+        ".macro do_store size\n"
+            ".if \\size == 8\n"
+                "vst3.8 {d10, d11, d12}, [%[rgb]]!\n"
+            ".elseif \\size == 4\n"
+                "vst3.8 {d10[0], d11[0], d12[0]}, [%[rgb]]!\n"
+                "vst3.8 {d10[1], d11[1], d12[1]}, [%[rgb]]!\n"
+                "vst3.8 {d10[2], d11[2], d12[2]}, [%[rgb]]!\n"
+                "vst3.8 {d10[3], d11[3], d12[3]}, [%[rgb]]!\n"
+            ".elseif \\size == 2\n"
+                "vst3.8 {d10[4], d11[4], d12[4]}, [%[rgb]]!\n"
+                "vst3.8 {d10[5], d11[5], d12[5]}, [%[rgb]]!\n"
+            ".elseif \\size == 1\n"
+                "vst3.8 {d10[6], d11[6], d12[6]}, [%[rgb]]!\n"
+            ".else\n"
+                ".error \"unsupported macroblock size\"\n"
+            ".endif\n"
+        ".endm\n"
+        "vld1.16 {d1, d2, d3}, [%[c]]\n" /* load constants */
+        "subs        %[n], %[n], #8\n"
+        "blt         2f\n"
+    "1:\n"
+        "do_load 8\n"
+        "do_yuv_to_rgb\n"
+        "do_store 8\n"
+        "subs        %[n], %[n], #8\n"
+        "bge         1b\n"
+        "tst         %[n], #7\n"
+        "beq         8f\n"
+    "2:\n"
+        "tst         %[n], #4\n"
+        "beq         3f\n"
+        "do_load 4\n"
+    "3:\n"
+        "tst         %[n], #2\n"
+        "beq         4f\n"
+        "do_load 2\n"
+    "4:\n"
+        "tst         %[n], #1\n"
+        "beq         5f\n"
+        "do_load 1\n"
+    "5:\n"
+        "do_yuv_to_rgb\n"
+        "tst         %[n], #4\n"
+        "beq         6f\n"
+        "do_store 4\n"
+    "6:\n"
+        "tst         %[n], #2\n"
+        "beq         7f\n"
+        "do_store 2\n"
+    "7:\n"
+        "tst         %[n], #1\n"
+        "beq         8f\n"
+        "do_store 1\n"
+    "8:\n"
+    ".purgem do_load\n"
+    ".purgem do_yuv_to_rgb\n"
+    ".purgem do_store\n"
+    : [rgb] "+&r" (rgb), [y] "+&r" (y), [u] "+&r" (u), [v] "+&r" (v),
+      [n] "+&r" (n)
+    : [c] "r" (&c[0])
+    : "cc", "memory",
+      "d0",  "d1",  "d2",  "d3",  "d4",  "d5",  "d6",  "d7",
+      "d8",  "d9",  "d10", "d11", "d12", "d13", "d14", "d15",
+      "d16", "d17", "d18", "d19", "d20", "d21", "d22", "d23",
+      "d24", "d25", "d26", "d27", "d28", "d29", "d30", "d31");
+}
+
+METHODDEF(void)
+ycc_rgb_convert_neon (j_decompress_ptr cinfo,
+		      JSAMPIMAGE input_buf, JDIMENSION input_row,
+		      JSAMPARRAY output_buf, int num_rows)
+{
+  register JSAMPROW outptr;
+  register JSAMPROW inptr0, inptr1, inptr2;
+  JDIMENSION num_cols = cinfo->output_width;
+
+  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++;
+    yuv444_to_rgb24_neon(outptr, inptr0, inptr1, inptr2, num_cols);
+    outptr += 3 * num_cols;
+  }
+}
+
+#endif
+
 METHODDEF(void)
 ycc_rgb_convert (j_decompress_ptr cinfo,
 		 JSAMPIMAGE input_buf, JDIMENSION input_row,
@@ -365,6 +533,13 @@ jinit_color_deconverter (j_decompress_ptr cinfo)
 
   case JCS_RGB:
   case JCS_EXT_RGB:
+#if defined(__ARM_NEON__) && defined(WITH_SIMD) && \
+    (RGB_PIXELSIZE == 3) && (BITS_IN_JSAMPLE == 8)
+    if (cinfo->jpeg_color_space == JCS_YCbCr) {
+        cconvert->pub.color_convert = ycc_rgb_convert_neon;
+        break;
+    }
+#endif
   case JCS_EXT_RGBX:
   case JCS_EXT_BGR:
   case JCS_EXT_BGRX: