diff options
| -rw-r--r-- | src/post/goom/xine_goom.c | 78 |
1 files changed, 77 insertions, 1 deletions
diff --git a/src/post/goom/xine_goom.c b/src/post/goom/xine_goom.c index 7cf1a9325..2dd42ce1f 100644 --- a/src/post/goom/xine_goom.c +++ b/src/post/goom/xine_goom.c @@ -34,6 +34,15 @@ #define LOG */ +/* TJ. this is what my 32bit Linux running on a AMD Athlon II X2 240e (2 * 2800Mhz) + performed @ 1024x576, 19 fps: + goom: csc_method 0 min 5355 us avg 5398 us + goom: csc_method 1 min 12718 us avg 12855 us + goom: csc_method 2 min 4199 us avg 4289 us + Seems gcc 4.5 has a very nice 64bit math emulation :-) +*/ +//#define BENCHMARK 1 + #include "config.h" #include <xine/xine_internal.h> #include <xine/xineutils.h> @@ -51,6 +60,7 @@ static const char* goom_csc_methods[]={ "Fast but not photorealistic", "Slow but looks better", + "Mostly fast and good quality", NULL }; @@ -99,9 +109,14 @@ struct post_plugin_goom_s { yuv_planes_t yuv; + void *rgb2yuy2; /* frame skipping */ int skip_frame; + +#ifdef BENCHMARK + int benchmark_frames, benchmark_min, benchmark_time; +#endif }; @@ -302,6 +317,17 @@ static post_plugin_t *goom_open_plugin(post_class_t *class_gen, int inputs, this->post.dispose = goom_dispose; +#ifdef __BIG_ENDIAN__ + this->rgb2yuy2 = rgb2yuy2_alloc (10, "argb"); +#else + this->rgb2yuy2 = rgb2yuy2_alloc (10, "bgra"); +#endif +#ifdef BENCHMARK + this->benchmark_frames = 0; + this->benchmark_min = 10000000; + this->benchmark_time = 0; +#endif + return &this->post; } @@ -326,6 +352,8 @@ static void goom_dispose(post_plugin_t *this_gen) { post_plugin_goom_t *this = (post_plugin_goom_t *)this_gen; + rgb2yuy2_free (this->rgb2yuy2); + if (_x_post_dispose(this_gen)) { this->class->ip = NULL; @@ -404,6 +432,14 @@ static void goom_port_close(xine_audio_port_t *port_gen, xine_stream_t *stream ) _x_post_dec_usage(port); } +#ifdef BENCHMARK +static int now (void) { + struct timeval tv; + gettimeofday (&tv, NULL); + return tv.tv_sec * 1000000 + tv.tv_usec; +} +#endif + static void goom_port_put_buffer (xine_audio_port_t *port_gen, audio_buffer_t *buf, xine_stream_t *stream) { @@ -503,13 +539,42 @@ static void goom_port_put_buffer (xine_audio_port_t *port_gen, this->metronom->got_video_frame(this->metronom, frame); if (!this->skip_frame) { +#ifdef BENCHMARK + int elapsed; +#endif /* Try to be fast */ goom_frame = (uint8_t *)goom_update (this->goom, this->data, 0, 0, NULL, NULL); dest_ptr = frame -> base[0]; goom_frame_end = goom_frame + 4 * (this->width_back * this->height_back); - if ((this->csc_method == 1) && +#ifdef BENCHMARK + elapsed = -now (); +#endif + + if (this->csc_method == 2) { + + if (!frame->proc_slice || (frame->height & 15)) { + /* do all at once */ + rgb2yuy2_slice (this->rgb2yuy2, goom_frame, 4 * this->width_back, + frame->base[0], frame->pitches[0], this->width_back, this->height_back); + } else { + /* sliced. This is a double edged sword. + On one hand, it is faster by using cache more effective. + On the flipside, when vo loop drops this frame, all this goes in vain */ + uint8_t *sptr[1]; + int y, h = 16, p = 4 * this->width_back; + for (y = 0; y < this->height_back; y += 16) { + if (y + 16 > this->height_back) + h = this->height_back & 15; + sptr[0] = frame->base[0] + y * frame->pitches[0]; + rgb2yuy2_slice (this->rgb2yuy2, goom_frame + y * p, p, + sptr[0], frame->pitches[0], this->width_back, h); + frame->proc_slice (frame, sptr); + } + } + + } else if ((this->csc_method == 1) && (xine_mm_accel() & MM_ACCEL_X86_MMX)) { int plane_ptr = 0; @@ -563,6 +628,17 @@ static void goom_port_put_buffer (xine_audio_port_t *port_gen, } } +#ifdef BENCHMARK + elapsed += now (); + if (this->benchmark_frames < 200) { + this->benchmark_time += elapsed; + if (elapsed < this->benchmark_min) this->benchmark_min = elapsed; + this->benchmark_frames++; + if (this->benchmark_frames == 200) printf ("goom: csc_method %d min %d us avg %d us\n", + this->csc_method, this->benchmark_min, this->benchmark_time / 200); + } +#endif + this->skip_frame = frame->draw(frame, XINE_ANON_STREAM); } else { frame->bad_frame = 1; |