Re: [FFmpeg-devel] [PATCH 1/2] dnxhdenc: Optimize get_pixels_8x4_sym for 10-bit

[Timothy Gu]

On Tue, Oct 20, 2015 at 9:30 AM Michael Niedermayer 
wrote:

> On Sat, Oct 17, 2015 at 06:05:45PM -0700, Timothy Gu wrote:
> > This reverts commit 628e6d0164febc8e69b0f10dfa487e8a2dd1a28a and uses
> > a better fix.
> >
> > Before:
> > 4483 decicycles in get_pixels_8x4_sym,  131032 runs, 40 skips
> >
> > After:
> > 2569 decicycles in get_pixels_8x4_sym,  131054 runs, 18 skips
> > ---
> >  libavcodec/dnxhdenc.c | 24 
> >  1 file changed, 8 insertions(+), 16 deletions(-)
>
> LGTM
>

Applied.

Timothy
___
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
http://ffmpeg.org/mailman/listinfo/ffmpeg-devel

[FFmpeg-devel] [PATCH 1/2] dnxhdenc: Optimize get_pixels_8x4_sym for 10-bit

[Timothy Gu]

This reverts commit 628e6d0164febc8e69b0f10dfa487e8a2dd1a28a and uses
a better fix.

Before:
4483 decicycles in get_pixels_8x4_sym,  131032 runs, 40 skips

After:
2569 decicycles in get_pixels_8x4_sym,  131054 runs, 18 skips
---
 libavcodec/dnxhdenc.c | 24 
 1 file changed, 8 insertions(+), 16 deletions(-)

diff --git a/libavcodec/dnxhdenc.c b/libavcodec/dnxhdenc.c
index 36154ac..04eb5d5 100644
--- a/libavcodec/dnxhdenc.c
+++ b/libavcodec/dnxhdenc.c
@@ -87,22 +87,14 @@ void dnxhd_10bit_get_pixels_8x4_sym(int16_t *av_restrict 
block,
 const uint8_t *pixels,
 ptrdiff_t line_size)
 {
-int i;
-const uint16_t* pixels16 = (const uint16_t*)pixels;
-line_size >>= 1;
-
-for (i = 0; i < 4; i++) {
-block[0] = pixels16[0]; block[1] = pixels16[1];
-block[2] = pixels16[2]; block[3] = pixels16[3];
-block[4] = pixels16[4]; block[5] = pixels16[5];
-block[6] = pixels16[6]; block[7] = pixels16[7];
-pixels16 += line_size;
-block += 8;
-}
-memcpy(block,  block -  8, sizeof(*block) * 8);
-memcpy(block +  8, block - 16, sizeof(*block) * 8);
-memcpy(block + 16, block - 24, sizeof(*block) * 8);
-memcpy(block + 24, block - 32, sizeof(*block) * 8);
+memcpy(block + 0 * 8, pixels, 8 * sizeof(*block));
+memcpy(block + 7 * 8, pixels, 8 * sizeof(*block));
+memcpy(block + 1 * 8, pixels + 1 * line_size, 8 * sizeof(*block));
+memcpy(block + 6 * 8, pixels + 1 * line_size, 8 * sizeof(*block));
+memcpy(block + 2 * 8, pixels + 2 * line_size, 8 * sizeof(*block));
+memcpy(block + 5 * 8, pixels + 2 * line_size, 8 * sizeof(*block));
+memcpy(block + 3 * 8, pixels + 3 * line_size, 8 * sizeof(*block));
+memcpy(block + 4 * 8, pixels + 3 * line_size, 8 * sizeof(*block));
 }
 
 static int dnxhd_10bit_dct_quantize(MpegEncContext *ctx, int16_t *block,
-- 
1.9.1

___
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
http://ffmpeg.org/mailman/listinfo/ffmpeg-devel