From mboxrd@z Thu Jan 1 00:00:00 1970 From: "Antonino A. Daplas" Subject: [PATCH][FBCON]: Optimization for accel_putcs() Date: Fri, 2 Jul 2004 11:31:10 +0800 Sender: linux-fbdev-devel-admin@lists.sourceforge.net Message-ID: <200407021056.27194.adaplas@hotpop.com> Reply-To: adaplas@pol.net Mime-Version: 1.0 Content-Transfer-Encoding: 7bit Return-path: Received: from sc8-sf-mx2-b.sourceforge.net ([10.3.1.12] helo=sc8-sf-mx2.sourceforge.net) by sc8-sf-list1.sourceforge.net with esmtp (Exim 4.30) id 1BgElJ-0006DQ-Iw for linux-fbdev-devel@lists.sourceforge.net; Thu, 01 Jul 2004 20:30:53 -0700 Received: from twix.hotpop.com ([38.113.3.71]) by sc8-sf-mx2.sourceforge.net with esmtp (Exim 4.34) id 1BgElI-0005Pp-BF for linux-fbdev-devel@lists.sourceforge.net; Thu, 01 Jul 2004 20:30:53 -0700 Received: from hotpop.com (kubrick.hotpop.com [38.113.3.103]) by twix.hotpop.com (Postfix) with SMTP id F3D81FD5AC4 for ; Fri, 2 Jul 2004 03:12:16 +0000 (UTC) Content-Disposition: inline Errors-To: linux-fbdev-devel-admin@lists.sourceforge.net List-Unsubscribe: , List-Id: List-Post: List-Help: List-Subscribe: , List-Archive: Content-Type: text/plain; charset="us-ascii" To: Andrew Morton , jsimmons@pentafluge.infradead.org Cc: Linux Fbdev development list Hi all, I did some simple benchmarking (time cat linux-2.6.7-mm5/MAINTAINERS) between 2.4 and 2.6 and I am not satisfied with what I see (It's claimed that fbdev-2.6 is faster than 2.4). The reason for the claim: 2.4 putcs - draw small amounts of data a lot of times 2.6 putcs - draw larger amounts of data a fewer times The way characters are drawn in 2.6 is optimal for accelerated drivers but should also give a speed boost for drivers that rely on software drawing. However the penaly incurred when preparing a large bitmap from a number of small bitmaps is currently very high. This is because of the following reasons: 1. fb_move_buf_{aligned|unaligned} uses pixmap->{out|in}buf. This is very expensive since outbuf and inbuf methods process only a byte or 2 of data at a time. 2. fb_sys_outbuf (the default method for pixmap->outbuf) uses memcpy(). Not a good choice if moving only a few bytes. 2. fb_move_buf_unaligned (used for fonts such as 12x22) also involves a lot of bit operations + a lot of calls to outbuf/inbuf which proportionately increases the penaly. So, I thought of separating fb_move_buf_* to fb_iomove_buf_* and fb_sysmove_buf_*. fb_iomove_buf_* - used if drivers specified outbuf and inbuf methods fb_sysmove_buf_* - used if drivers have no outbuf or inbuf methods *Most, if not all drivers fall in the second category. Below is a table that show differences between 2.4, 2.6 and 2.6 + abovementioned changes. To reduce the effect of panning and fillrect/copyarea, the scrollmode is forced to redraw. ================================================================= Test Hardware: P4 2G nVidia GeForce2 MX 64 Scrollmode: redraw time cat linux-2.6.7-mm5/MAINTAINERS 1024x768-8 1024x768-16 1024x768-32 ================================================================= 8x16 noaccel (2.4) real 0m5.490s real 0m8.535s real 0m15.388s user 0m0.001s user 0m0.000s user 0m0.001s sys 0m5.487s sys 0m8.535s sys 0m15.386s 8x16 noaccel (2.6) real 0m5.166s real 0m7.195s real 0m12.177s user 0m0.001s user 0m0.000s user 0m0.000s sys 0m5.164s sys 0m7.192s sys 0m12.176s 8x16 noaccel+patch (2.6) real 0m3.474s real 0m5.496s real 0m10.460s user 0m0.001s user 0m0.001s user 0m0.001s sys 0m5.492s sys 0m5.492s sys 0m10.454s ----------------------------------------------------------------- 8x16 accel (2.4) real 0m4.368s real 0m9.420s real 0m22.415s user 0m0.001s user 0m0.001s user 0m0.001s sys 0m4.019s sys 0m9.384s sys 0m22.312s 8x16 accel (2.6) real 0m4.296s real 0m4.339s real 0m4.391s user 0m0.001s user 0m0.001s user 0m0.000s sys 0m4.280s sys 0m4.336s sys 0m4.389s 8x16 accel+patch (2.6) real 0m2.536s real 0m2.649s real 0m2.799s user 0m0.000s user 0m0.000s user 0m0.001s sys 0m2.536s sys 0m2.645s sys 0m2.798s ----------------------------------------------------------------- 1024x768-8 1024x768-16 1024x768-32 ================================================================= 12x22 noaccel (2.4) real 0m7.883s real 0m12.175s real 0m21.134s user 0m0.000s user 0m0.000s user 0m0.001s sys 0m7.882s sys 0m12.174s sys 0m21.129s 12x22 noaccel (2.6) real 0m10.651s real 0m13.550s real 0m21.009s user 0m0.001s user 0m0.001s user 0m0.000s sys 0m10.617s sys 0m13.545s sys 0m21.008s 12x22 noaccel+patch (2.6) real 0m4.794s real 0m7.718s real 0m15.173s user 0m0.002s user 0m0.001s user 0m0.000s sys 0m4.792s sys 0m7.715s sys 0m15.170s ----------------------------------------------------------------- 12x22 accel (2.4) real 0m3.971s real 0m9.030s real 0m21.711s user 0m0.000s user 0m0.000s user 0m0.000s sys 0m3.950s sys 0m8.983s sys 0m21.602s 12x22 accel (2.6) real 0m9.392s real 0m9.486s real 0m9.508s user 0m0.000s user 0m0.000s user 0m0.001s sys 0m9.392s sys 0m9.484s sys 0m9.484s 12x22 accel+patch (2.6) real 0m3.570s real 0m3.603s real 0m3.848s user 0m0.001s user 0m0.000s user 0m0.000s sys 0m3.567s sys 0m3.600s sys 0m3.844s ----------------------------------------------------------------- Summary: 1. 2.6 unaccelerated is a bit faster than 2.4 when handling 8x16 fonts, with a higher speed differential at high color depths. 2. 2.4 unaccelerated is a bit faster than 2.6 when handling 12x22 fonts, with a smaller speed difference at high color depths (2.6 is actually a bit faster than 2.4 at 32bpp). 3. 2.4 rivafb accelerated suffers at high color depths, even becoming slower than unaccelerated, possibly because of the 'draw few bytes many times' method. 4. 2.6 rivafb accelerated has similar performance at any color depth, possibly because of 'draw lots of bytes a fewer times' method. 5. With the changes, there is a speed gain of ~1.7 seconds and ~5.7 seconds with 8x16 and 12x22 fonts respectively indepependent of the color depth or acceleration used. The speed gain is constant but significant. Below is a patch against 2.6.7-mm5. The effects will be very noticeable with drivers that uses SCROLL_REDRAW, but one should still see some speed gain even if SCROLL_YPAN/YWRAP is used. Tony Signed-off-by: Antonino Daplas Separated fb_sys_move_* into fb_iosys_move_* and fb_sysmove_* to reduce penalty when constructing fb_image->data from character maps. In my testcase (1024x768 SCROLL_REDRAW), I get a ~1.7 second advantage with 'time cat MAINTAINERS' using 8x16 fonts and ~5.7 seconds with 12x22 fonts. The speed gain is independent of acceleration or color depth. diff -Naur linux-2.6.7-mm5-orig/drivers/video/console/fbcon.c linux-2.6.7-mm5/drivers/video/console/fbcon.c --- linux-2.6.7-mm5-orig/drivers/video/console/fbcon.c 2004-07-02 01:43:40.477020360 +0000 +++ linux-2.6.7-mm5/drivers/video/console/fbcon.c 2004-07-02 01:42:12.935328720 +0000 @@ -436,8 +436,15 @@ } void accel_putcs(struct vc_data *vc, struct fb_info *info, - const unsigned short *s, int count, int yy, int xx) + const unsigned short *s, int count, int yy, int xx) { + void (*move_unaligned)(struct fb_info *info, struct fb_pixmap *buf, + u8 *dst, u32 d_pitch, u8 *src, u32 idx, + u32 height, u32 shift_high, u32 shift_low, + u32 mod); + void (*move_aligned)(struct fb_info *info, struct fb_pixmap *buf, + u8 *dst, u32 d_pitch, u8 *src, u32 s_pitch, + u32 height); unsigned short charmask = vc->vc_hi_font_mask ? 0x1ff : 0xff; unsigned int width = (vc->vc_font.width + 7) >> 3; unsigned int cellsize = vc->vc_font.height * width; @@ -446,20 +453,26 @@ unsigned int buf_align = info->pixmap.buf_align - 1; unsigned int shift_low = 0, mod = vc->vc_font.width % 8; unsigned int shift_high = 8, pitch, cnt, size, k; - int bgshift = (vc->vc_hi_font_mask) ? 13 : 12; - int fgshift = (vc->vc_hi_font_mask) ? 9 : 8; unsigned int idx = vc->vc_font.width >> 3; struct fb_image image; - u16 c = scr_readw(s); - u8 *src, *dst, *dst0; + u8 *src, *dst; - image.fg_color = attr_fgcol(fgshift, c); - image.bg_color = attr_bgcol(bgshift, c); + image.fg_color = attr_fgcol((vc->vc_hi_font_mask) ? 9 : 8, + scr_readw(s)); + image.bg_color = attr_bgcol((vc->vc_hi_font_mask) ? 13 : 12, + scr_readw(s)); image.dx = xx * vc->vc_font.width; image.dy = yy * vc->vc_font.height; image.height = vc->vc_font.height; image.depth = 1; + if (info->pixmap.outbuf && info->pixmap.inbuf) { + move_aligned = fb_iomove_buf_aligned; + move_unaligned = fb_iomove_buf_unaligned; + } else { + move_aligned = fb_sysmove_buf_aligned; + move_unaligned = fb_sysmove_buf_unaligned; + } while (count) { if (count > maxcnt) cnt = k = maxcnt; @@ -471,24 +484,27 @@ pitch &= ~scan_align; size = pitch * image.height + buf_align; size &= ~buf_align; - dst0 = fb_get_buffer_offset(info, &info->pixmap, size); - image.data = dst0; - while (k--) { - src = vc->vc_font.data + (scr_readw(s++) & charmask)*cellsize; - dst = dst0; - - if (mod) { - fb_move_buf_unaligned(info, &info->pixmap, dst, pitch, - src, idx, image.height, shift_high, - shift_low, mod); + dst = fb_get_buffer_offset(info, &info->pixmap, size); + image.data = dst; + if (mod) { + while (k--) { + src = vc->vc_font.data + (scr_readw(s++)& + charmask)*cellsize; + move_unaligned(info, &info->pixmap, dst, pitch, + src, idx, image.height, + shift_high, shift_low, mod); shift_low += mod; - dst0 += (shift_low >= 8) ? width : width - 1; + dst += (shift_low >= 8) ? width : width - 1; shift_low &= 7; shift_high = 8 - shift_low; - } else { - fb_move_buf_aligned(info, &info->pixmap, dst, pitch, - src, idx, image.height); - dst0 += width; + } + } else { + while (k--) { + src = vc->vc_font.data + (scr_readw(s++)& + charmask)*cellsize; + move_aligned(info, &info->pixmap, dst, pitch, + src, idx, image.height); + dst += width; } } info->fbops->fb_imageblit(info, &image); @@ -950,8 +966,11 @@ dst = fb_get_buffer_offset(info, &info->pixmap, size); image.data = dst; - fb_move_buf_aligned(info, &info->pixmap, dst, pitch, src, width, image.height); - + if (info->pixmap.outbuf) + fb_iomove_buf_aligned(info, &info->pixmap, dst, pitch, src, width, image.height); + else + fb_sysmove_buf_aligned(info, &info->pixmap, dst, pitch, src, width, image.height); + info->fbops->fb_imageblit(info, &image); } diff -Naur linux-2.6.7-mm5-orig/drivers/video/fbmem.c linux-2.6.7-mm5/drivers/video/fbmem.c --- linux-2.6.7-mm5-orig/drivers/video/fbmem.c 2004-07-02 01:43:07.470038184 +0000 +++ linux-2.6.7-mm5/drivers/video/fbmem.c 2004-07-02 01:41:36.126924448 +0000 @@ -430,34 +430,37 @@ /* * Drawing helpers. */ -static u8 fb_sys_inbuf(struct fb_info *info, u8 *src) +void fb_iomove_buf_aligned(struct fb_info *info, struct fb_pixmap *buf, + u8 *dst, u32 d_pitch, u8 *src, u32 s_pitch, + u32 height) { - return *src; -} + int i; -static void fb_sys_outbuf(struct fb_info *info, u8 *dst, - u8 *src, unsigned int size) -{ - memcpy(dst, src, size); -} + for (i = height; i--; ) { + buf->outbuf(info, dst, src, s_pitch); + src += s_pitch; + dst += d_pitch; + } +} -void fb_move_buf_aligned(struct fb_info *info, struct fb_pixmap *buf, - u8 *dst, u32 d_pitch, u8 *src, u32 s_pitch, - u32 height) +void fb_sysmove_buf_aligned(struct fb_info *info, struct fb_pixmap *buf, + u8 *dst, u32 d_pitch, u8 *src, u32 s_pitch, + u32 height) { - int i; + int i, j; for (i = height; i--; ) { - buf->outbuf(info, dst, src, s_pitch); + for (j = 0; j < s_pitch; j++) + dst[j] = src[j]; src += s_pitch; dst += d_pitch; } } -void fb_move_buf_unaligned(struct fb_info *info, struct fb_pixmap *buf, - u8 *dst, u32 d_pitch, u8 *src, u32 idx, - u32 height, u32 shift_high, u32 shift_low, - u32 mod) +void fb_iomove_buf_unaligned(struct fb_info *info, struct fb_pixmap *buf, + u8 *dst, u32 d_pitch, u8 *src, u32 idx, + u32 height, u32 shift_high, u32 shift_low, + u32 mod) { u8 mask = (u8) (0xfff << shift_high), tmp; int i, j; @@ -485,6 +488,37 @@ } } +void fb_sysmove_buf_unaligned(struct fb_info *info, struct fb_pixmap *buf, + u8 *dst, u32 d_pitch, u8 *src, u32 idx, + u32 height, u32 shift_high, u32 shift_low, + u32 mod) +{ + u8 mask = (u8) (0xfff << shift_high), tmp; + int i, j; + + for (i = height; i--; ) { + for (j = 0; j < idx; j++) { + tmp = dst[j]; + tmp &= mask; + tmp |= *src >> shift_low; + dst[j] = tmp; + tmp = *src << shift_high; + dst[j+1] = tmp; + src++; + } + tmp = dst[idx]; + tmp &= mask; + tmp |= *src >> shift_low; + dst[idx] = tmp; + if (shift_high < mod) { + tmp = *src << shift_high; + dst[idx+1] = tmp; + } + src++; + dst += d_pitch; + } +} + /* * we need to lock this section since fb_cursor * may use fb_imageblit() @@ -897,7 +931,10 @@ unsigned int width = (info->cursor.image.width + 7) >> 3; u8 *data = (u8 *) info->cursor.image.data; - info->sprite.outbuf(info, info->sprite.addr, data, width); + if (info->sprite.outbuf) + info->sprite.outbuf(info, info->sprite.addr, data, width); + else + memcpy(info->sprite.addr, data, width); } int @@ -1319,10 +1356,6 @@ } } fb_info->pixmap.offset = 0; - if (fb_info->pixmap.outbuf == NULL) - fb_info->pixmap.outbuf = fb_sys_outbuf; - if (fb_info->pixmap.inbuf == NULL) - fb_info->pixmap.inbuf = fb_sys_inbuf; if (fb_info->sprite.addr == NULL) { fb_info->sprite.addr = kmalloc(FBPIXMAPSIZE, GFP_KERNEL); @@ -1335,10 +1368,6 @@ } } fb_info->sprite.offset = 0; - if (fb_info->sprite.outbuf == NULL) - fb_info->sprite.outbuf = fb_sys_outbuf; - if (fb_info->sprite.inbuf == NULL) - fb_info->sprite.inbuf = fb_sys_inbuf; registered_fb[i] = fb_info; @@ -1533,8 +1562,10 @@ EXPORT_SYMBOL(fb_blank); EXPORT_SYMBOL(fb_pan_display); EXPORT_SYMBOL(fb_get_buffer_offset); -EXPORT_SYMBOL(fb_move_buf_unaligned); -EXPORT_SYMBOL(fb_move_buf_aligned); +EXPORT_SYMBOL(fb_iomove_buf_unaligned); +EXPORT_SYMBOL(fb_iomove_buf_aligned); +EXPORT_SYMBOL(fb_sysmove_buf_unaligned); +EXPORT_SYMBOL(fb_sysmove_buf_aligned); EXPORT_SYMBOL(fb_load_cursor_image); EXPORT_SYMBOL(fb_set_suspend); EXPORT_SYMBOL(fb_register_client); diff -Naur linux-2.6.7-mm5-orig/drivers/video/riva/fbdev.c linux-2.6.7-mm5/drivers/video/riva/fbdev.c --- linux-2.6.7-mm5-orig/drivers/video/riva/fbdev.c 2004-07-02 01:43:29.172738872 +0000 +++ linux-2.6.7-mm5/drivers/video/riva/fbdev.c 2004-07-02 01:45:41.549614528 +0000 @@ -1619,8 +1619,8 @@ break; } - fb_move_buf_aligned(info, &info->sprite, data, d_pitch, src, - s_pitch, info->cursor.image.height); + fb_sysmove_buf_aligned(info, &info->sprite, data, d_pitch, src, + s_pitch, info->cursor.image.height); bg = ((info->cmap.red[bg_idx] & 0xf8) << 7) | ((info->cmap.green[bg_idx] & 0xf8) << 2) | diff -Naur linux-2.6.7-mm5-orig/drivers/video/softcursor.c linux-2.6.7-mm5/drivers/video/softcursor.c --- linux-2.6.7-mm5-orig/drivers/video/softcursor.c 2004-07-02 01:43:15.299847872 +0000 +++ linux-2.6.7-mm5/drivers/video/softcursor.c 2004-07-02 01:41:41.226149248 +0000 @@ -73,7 +73,12 @@ } else memcpy(src, cursor->image.data, dsize); - fb_move_buf_aligned(info, &info->sprite, dst, d_pitch, src, s_pitch, info->cursor.image.height); + if (info->sprite.outbuf) + fb_iomove_buf_aligned(info, &info->sprite, dst, d_pitch, src, + s_pitch, info->cursor.image.height); + else + fb_sysmove_buf_aligned(info, &info->sprite, dst, d_pitch, src, + s_pitch, info->cursor.image.height); info->cursor.image.data = dst; info->fbops->fb_imageblit(info, &info->cursor.image); diff -Naur linux-2.6.7-mm5-orig/include/linux/fb.h linux-2.6.7-mm5/include/linux/fb.h --- linux-2.6.7-mm5-orig/include/linux/fb.h 2004-07-02 01:43:54.523884912 +0000 +++ linux-2.6.7-mm5/include/linux/fb.h 2004-07-02 01:42:42.077898376 +0000 @@ -638,10 +638,16 @@ extern int fb_prepare_logo(struct fb_info *fb_info); extern int fb_show_logo(struct fb_info *fb_info); extern char* fb_get_buffer_offset(struct fb_info *info, struct fb_pixmap *buf, u32 size); -extern void fb_move_buf_unaligned(struct fb_info *info, struct fb_pixmap *buf, +extern void fb_iomove_buf_unaligned(struct fb_info *info, struct fb_pixmap *buf, u8 *dst, u32 d_pitch, u8 *src, u32 idx, u32 height, u32 shift_high, u32 shift_low, u32 mod); -extern void fb_move_buf_aligned(struct fb_info *info, struct fb_pixmap *buf, +extern void fb_iomove_buf_aligned(struct fb_info *info, struct fb_pixmap *buf, + u8 *dst, u32 d_pitch, u8 *src, u32 s_pitch, + u32 height); +extern void fb_sysmove_buf_unaligned(struct fb_info *info, struct fb_pixmap *buf, + u8 *dst, u32 d_pitch, u8 *src, u32 idx, + u32 height, u32 shift_high, u32 shift_low, u32 mod); +extern void fb_sysmove_buf_aligned(struct fb_info *info, struct fb_pixmap *buf, u8 *dst, u32 d_pitch, u8 *src, u32 s_pitch, u32 height); extern void fb_load_cursor_image(struct fb_info *); ------------------------------------------------------- This SF.Net email sponsored by Black Hat Briefings & Training. Attend Black Hat Briefings & Training, Las Vegas July 24-29 - digital self defense, top technical experts, no vendor pitches, unmatched networking opportunities. Visit www.blackhat.com