Re: [PATCH} Optimized rasops32 putchar
On Tue, Jul 14, 2020 at 12:15:37PM +0200, Frederic Cambus wrote: > The diff makes sense to me, I will commit it this week with some minor > style(9) fixes for the switch statement (don't indent the case), unless > I hear objections. Committed, thanks again!
Re: [PATCH} Optimized rasops32 putchar
On Sun, Jul 12, 2020 at 07:16:13PM +0200, Frederic Cambus wrote: > On Fri, Jun 26, 2020 at 07:42:50AM -0700, jo...@armadilloaerospace.com wrote: > > Optimized 32 bit character rendering with unrolled rows and pairwise > > foreground / background pixel rendering. > > > > If it weren't for the 5x8 font, I would have just assumed everything > > was an even width and made the fallback path also pairwise. > > > > In isolation, the 16x32 character case got 2x faster, but that wasn't > > a huge real world speedup where the space rendering that was already > > at memory bandwidth limits accounted for most of the character > > rendering time. However, in combination with the previous fast > > conditional console scrolling that removes most of the space rendering, > > it becomes significant. > > On my Ryzen desktop with radeondrm, I don't see any improvements, the > rasops_vcons_copyrows() optimizations seems to have made character > plotting fast enough so that it's not a bottleneck anymore, which is > definitely great. > > cpu0: AMD Ryzen 7 2700 Eight-Core Processor, 3394.18 MHz, 17-08-02 > radeondrm0 at pci8 dev 0 function 0 "ATI Radeon HD 6450" rev 0x00 > radeondrm0: 1920x1080, 32bpp > > On my T450 however, this diff makes cat'ing my usual test file [1] up > to 20% faster with the default 12x24 font on the built-in 1600x900 > screen, which I think is significant enough for the diff to go in. > > cpu0: Intel(R) Core(TM) i5-5200U CPU @ 2.20GHz, 2095.47 MHz, 06-3d-04 > inteldrm0 at pci0 dev 2 function 0 "Intel HD Graphics 5500" rev 0x09 > drm0 at inteldrm0 > inteldrm0: 1600x900, 32bpp > > On my Cubieboard2 (armv7) I didn't notice any meaningful difference, > which I assume is to be expected on a 32-bit platform. I suppose it's > also reasonable to assume other 32-bit platforms (i386, hppa, macppc) > will not see any regression beyond noise level? > > Anyone willing to OK this diff? > > [1] https://norvig.com/big.txt So I tested on the other 32-bit machine I have, and didn't notice any regression on my i386 machine with inteldrm, it is actually up to 1.5% faster. It seems we can remove the 'q' variable and drop this assignement, as it is not used: q = u.q[0]; The diff makes sense to me, I will commit it this week with some minor style(9) fixes for the switch statement (don't indent the case), unless I hear objections.
Re: [PATCH} Optimized rasops32 putchar
On Fri, Jun 26, 2020 at 07:42:50AM -0700, jo...@armadilloaerospace.com wrote: > Optimized 32 bit character rendering with unrolled rows and pairwise > foreground / background pixel rendering. > > If it weren't for the 5x8 font, I would have just assumed everything > was an even width and made the fallback path also pairwise. > > In isolation, the 16x32 character case got 2x faster, but that wasn't > a huge real world speedup where the space rendering that was already > at memory bandwidth limits accounted for most of the character > rendering time. However, in combination with the previous fast > conditional console scrolling that removes most of the space rendering, > it becomes significant. On my Ryzen desktop with radeondrm, I don't see any improvements, the rasops_vcons_copyrows() optimizations seems to have made character plotting fast enough so that it's not a bottleneck anymore, which is definitely great. cpu0: AMD Ryzen 7 2700 Eight-Core Processor, 3394.18 MHz, 17-08-02 radeondrm0 at pci8 dev 0 function 0 "ATI Radeon HD 6450" rev 0x00 radeondrm0: 1920x1080, 32bpp On my T450 however, this diff makes cat'ing my usual test file [1] up to 20% faster with the default 12x24 font on the built-in 1600x900 screen, which I think is significant enough for the diff to go in. cpu0: Intel(R) Core(TM) i5-5200U CPU @ 2.20GHz, 2095.47 MHz, 06-3d-04 inteldrm0 at pci0 dev 2 function 0 "Intel HD Graphics 5500" rev 0x09 drm0 at inteldrm0 inteldrm0: 1600x900, 32bpp On my Cubieboard2 (armv7) I didn't notice any meaningful difference, which I assume is to be expected on a 32-bit platform. I suppose it's also reasonable to assume other 32-bit platforms (i386, hppa, macppc) will not see any regression beyond noise level? Anyone willing to OK this diff? [1] https://norvig.com/big.txt
Re: [PATCH} Optimized rasops32 putchar
I did some more tests, and I think the odd performance I am seeing may be due to TLB thrash on the 32x64 characters with 4k pages, since writing each character will require 64 data TLB. Are huge page mappings supported in OpenBSD? Original Message Subject: Re: [PATCH} Optimized rasops32 putchar From: Mark Kettenis Date: Sat, June 27, 2020 1:30 pm To: Cc: tech@openbsd.org > Content-Type: text/plain; charset="utf-8" > From: > > I was doing my timings with a user mode program after mmmaping the > efifb display, so the mapping might be different in the kernel. That should still give you a write-combining mapping as efifb_mmap() adds the PMAP_WC flag to the physical address. Cachable on x86 means write-back cachable. And using a write-back cachable mapping for a framebuffer often leads to interesting "damage" where pixels in certain cache lines show up "late" on the display. Not sure if you'd see that on recent Intel graphics hardware as the current hardware designs are much more coherent than what was produced in the past. > Related to that, I was going to add mmap / WSDISPLAYIO_LINEBYTES / > WSDISPLAYIO_SMODE to the drm drivers by consolidating code into > rasops. While the point of the DRM drivers is to get fully hardware > accelerated drawing in X, there isn't any reason why they can't > support dumb framebuffer mappings as well. True. Although there are DRM interfaces that give you a dumb framebuffer as well. Using those interfaces is a bit more complicated though. Centralising the code would be good. That code probably should use bus_space_mmap(4) as the PMAP_WC flag is amd64-specific. Unfortunately the amd64 implementation of bus_space_mmap(4) is incomplete and doesn't actually implement write-combining for mappings with the BUS_SPACE_MAP_PREFETCHABLE flag set. So that has to be fixed as well. > ---- Original Message > Subject: RE: [PATCH} Optimized rasops32 putchar > From: > Date: Sat, June 27, 2020 11:13 am > To: "Mark Kettenis" > Cc: "tech@openbsd.org" > > I believe it is mapped as normally cached right now, rather than > uncached or write combining. > > Reads aren't ultra-slow, and the timings of 48 byte writes appear to > involve a cacheline read. > > 128 byte writes are actually slower than 64 byte writes, which I > guessed might be because of automatic prefetching kicking in and > reading the following cacheline. > > > Original Message > Subject: Re: [PATCH} Optimized rasops32 putchar > From: Mark Kettenis > Date: Sat, June 27, 2020 7:56 am > To: > Cc: tech@openbsd.org > > > From: > > Date: Fri, 26 Jun 2020 07:42:50 -0700 > > > > Optimized 32 bit character rendering with unrolled rows and pairwise > > foreground / background pixel rendering. > > > > If it weren't for the 5x8 font, I would have just assumed everything > > was an even width and made the fallback path also pairwise. > > > > In isolation, the 16x32 character case got 2x faster, but that wasn't > > a huge real world speedup where the space rendering that was already > > at memory bandwidth limits accounted for most of the character > > rendering time. However, in combination with the previous fast > > conditional console scrolling that removes most of the space rendering, > > it becomes significant. > > > > I also found that at least the efi and intel framebuffers are not > > currently mapped write combining, which makes this much slower than > > it should be. > > Hi John, > > The framebuffer should be mapped write-combining. In OpenBSD this is > requested by specifying the BUS_SPACE_MAP_PREFETCHABLE flag to > bbus_space_map(9) when mapping the framebuffer. > > I'm fairly confident since until last January the initial mapping of > the framebuffer that we used wasn't write-combining. And things were > really, really slow. > > Cheers, > > Mark > > > Index: rasops32.c > > === > > RCS file: /cvs/src/sys/dev/rasops/rasops32.c,v > > retrieving revision 1.10 > > diff -u -p -r1.10 rasops32.c > > --- rasops32.c 25 May 2020 09:55:49 - 1.10 > > +++ rasops32.c 26 Jun 2020 14:34:06 - > > @@ -65,9 +65,14 @@ rasops32_init(struct rasops_info *ri) > > int > > rasops32_putchar(void *cookie, int row, int col, u_int uc, uint32_t > > attr) > > { > > - int width, height, cnt, fs, fb, clr[2]; > > + int width, height, step, cnt, fs, b, f; > > + uint32_t fb, clr[2]; > > struct rasops_info *ri; > > - int32_t *dp, *rp; > > + int64_t *rp, q; >
Re: [PATCH} Optimized rasops32 putchar
> Content-Type: text/plain; charset="utf-8" > From: > > I was doing my timings with a user mode program after mmmaping the > efifb display, so the mapping might be different in the kernel. That should still give you a write-combining mapping as efifb_mmap() adds the PMAP_WC flag to the physical address. Cachable on x86 means write-back cachable. And using a write-back cachable mapping for a framebuffer often leads to interesting "damage" where pixels in certain cache lines show up "late" on the display. Not sure if you'd see that on recent Intel graphics hardware as the current hardware designs are much more coherent than what was produced in the past. > Related to that, I was going to add mmap / WSDISPLAYIO_LINEBYTES / > WSDISPLAYIO_SMODE to the drm drivers by consolidating code into > rasops. While the point of the DRM drivers is to get fully hardware > accelerated drawing in X, there isn't any reason why they can't > support dumb framebuffer mappings as well. True. Although there are DRM interfaces that give you a dumb framebuffer as well. Using those interfaces is a bit more complicated though. Centralising the code would be good. That code probably should use bus_space_mmap(4) as the PMAP_WC flag is amd64-specific. Unfortunately the amd64 implementation of bus_space_mmap(4) is incomplete and doesn't actually implement write-combining for mappings with the BUS_SPACE_MAP_PREFETCHABLE flag set. So that has to be fixed as well. > ---- Original Message > Subject: RE: [PATCH} Optimized rasops32 putchar > From: > Date: Sat, June 27, 2020 11:13 am > To: "Mark Kettenis" > Cc: "tech@openbsd.org" > > I believe it is mapped as normally cached right now, rather than > uncached or write combining. > > Reads aren't ultra-slow, and the timings of 48 byte writes appear to > involve a cacheline read. > > 128 byte writes are actually slower than 64 byte writes, which I > guessed might be because of automatic prefetching kicking in and > reading the following cacheline. > > > Original Message > Subject: Re: [PATCH} Optimized rasops32 putchar > From: Mark Kettenis > Date: Sat, June 27, 2020 7:56 am > To: > Cc: tech@openbsd.org > > > From: > > Date: Fri, 26 Jun 2020 07:42:50 -0700 > > > > Optimized 32 bit character rendering with unrolled rows and pairwise > > foreground / background pixel rendering. > > > > If it weren't for the 5x8 font, I would have just assumed everything > > was an even width and made the fallback path also pairwise. > > > > In isolation, the 16x32 character case got 2x faster, but that wasn't > > a huge real world speedup where the space rendering that was already > > at memory bandwidth limits accounted for most of the character > > rendering time. However, in combination with the previous fast > > conditional console scrolling that removes most of the space rendering, > > it becomes significant. > > > > I also found that at least the efi and intel framebuffers are not > > currently mapped write combining, which makes this much slower than > > it should be. > > Hi John, > > The framebuffer should be mapped write-combining. In OpenBSD this is > requested by specifying the BUS_SPACE_MAP_PREFETCHABLE flag to > bbus_space_map(9) when mapping the framebuffer. > > I'm fairly confident since until last January the initial mapping of > the framebuffer that we used wasn't write-combining. And things were > really, really slow. > > Cheers, > > Mark > > > Index: rasops32.c > > === > > RCS file: /cvs/src/sys/dev/rasops/rasops32.c,v > > retrieving revision 1.10 > > diff -u -p -r1.10 rasops32.c > > --- rasops32.c 25 May 2020 09:55:49 - 1.10 > > +++ rasops32.c 26 Jun 2020 14:34:06 - > > @@ -65,9 +65,14 @@ rasops32_init(struct rasops_info *ri) > > int > > rasops32_putchar(void *cookie, int row, int col, u_int uc, uint32_t > > attr) > > { > > - int width, height, cnt, fs, fb, clr[2]; > > + int width, height, step, cnt, fs, b, f; > > + uint32_t fb, clr[2]; > > struct rasops_info *ri; > > - int32_t *dp, *rp; > > + int64_t *rp, q; > > + union { > > + int64_t q[4]; > > + int32_t d[4][2]; > > + } u; > > u_char *fr; > > > > ri = (struct rasops_info *)cookie; > > @@ -81,48 +86,128 @@ rasops32_putchar(void *cookie, int row, > > return 0; > > #endif > > > > - rp = (int32_t *)(ri->ri_bits + row*ri->ri_yscale + col*ri->ri_xscale); > > + rp =
Re: [PATCH} Optimized rasops32 putchar
I was doing my timings with a user mode program after mmmaping the efifb display, so the mapping might be different in the kernel. Related to that, I was going to add mmap / WSDISPLAYIO_LINEBYTES / WSDISPLAYIO_SMODE to the drm drivers by consolidating code into rasops. While the point of the DRM drivers is to get fully hardware accelerated drawing in X, there isn't any reason why they can't support dumb framebuffer mappings as well. Original Message Subject: RE: [PATCH} Optimized rasops32 putchar From: Date: Sat, June 27, 2020 11:13 am To: "Mark Kettenis" Cc: "tech@openbsd.org" I believe it is mapped as normally cached right now, rather than uncached or write combining. Reads aren't ultra-slow, and the timings of 48 byte writes appear to involve a cacheline read. 128 byte writes are actually slower than 64 byte writes, which I guessed might be because of automatic prefetching kicking in and reading the following cacheline. Original Message ---- Subject: Re: [PATCH} Optimized rasops32 putchar From: Mark Kettenis Date: Sat, June 27, 2020 7:56 am To: Cc: tech@openbsd.org > From: > Date: Fri, 26 Jun 2020 07:42:50 -0700 > > Optimized 32 bit character rendering with unrolled rows and pairwise > foreground / background pixel rendering. > > If it weren't for the 5x8 font, I would have just assumed everything > was an even width and made the fallback path also pairwise. > > In isolation, the 16x32 character case got 2x faster, but that wasn't > a huge real world speedup where the space rendering that was already > at memory bandwidth limits accounted for most of the character > rendering time. However, in combination with the previous fast > conditional console scrolling that removes most of the space rendering, > it becomes significant. > > I also found that at least the efi and intel framebuffers are not > currently mapped write combining, which makes this much slower than > it should be. Hi John, The framebuffer should be mapped write-combining. In OpenBSD this is requested by specifying the BUS_SPACE_MAP_PREFETCHABLE flag to bbus_space_map(9) when mapping the framebuffer. I'm fairly confident since until last January the initial mapping of the framebuffer that we used wasn't write-combining. And things were really, really slow. Cheers, Mark > Index: rasops32.c > === > RCS file: /cvs/src/sys/dev/rasops/rasops32.c,v > retrieving revision 1.10 > diff -u -p -r1.10 rasops32.c > --- rasops32.c 25 May 2020 09:55:49 - 1.10 > +++ rasops32.c 26 Jun 2020 14:34:06 - > @@ -65,9 +65,14 @@ rasops32_init(struct rasops_info *ri) > int > rasops32_putchar(void *cookie, int row, int col, u_int uc, uint32_t > attr) > { > - int width, height, cnt, fs, fb, clr[2]; > + int width, height, step, cnt, fs, b, f; > + uint32_t fb, clr[2]; > struct rasops_info *ri; > - int32_t *dp, *rp; > + int64_t *rp, q; > + union { > + int64_t q[4]; > + int32_t d[4][2]; > + } u; > u_char *fr; > > ri = (struct rasops_info *)cookie; > @@ -81,48 +86,128 @@ rasops32_putchar(void *cookie, int row, > return 0; > #endif > > - rp = (int32_t *)(ri->ri_bits + row*ri->ri_yscale + col*ri->ri_xscale); > + rp = (int64_t *)(ri->ri_bits + row*ri->ri_yscale + col*ri->ri_xscale); > > height = ri->ri_font->fontheight; > width = ri->ri_font->fontwidth; > + step = ri->ri_stride >> 3; > > - clr[0] = ri->ri_devcmap[(attr >> 16) & 0xf]; > - clr[1] = ri->ri_devcmap[(attr >> 24) & 0xf]; > + b = ri->ri_devcmap[(attr >> 16) & 0xf]; > + f = ri->ri_devcmap[(attr >> 24) & 0xf]; > + u.d[0][0] = b; u.d[0][1] = b; > + u.d[1][0] = b; u.d[1][1] = f; > + u.d[2][0] = f; u.d[2][1] = b; > + u.d[3][0] = f; u.d[3][1] = f; > > if (uc == ' ') { > + q = u.q[0]; > while (height--) { > - dp = rp; > - DELTA(rp, ri->ri_stride, int32_t *); > - > - for (cnt = width; cnt; cnt--) > - *dp++ = clr[0]; > + /* the general, pixel-at-a-time case is fast enough */ > + for (cnt = 0; cnt < width; cnt++) > + ((int *)rp)[cnt] = b; > + rp += step; > } > } else { > uc -= ri->ri_font->firstchar; > fr = (u_char *)ri->ri_font->data + uc * ri->ri_fontscale; > fs = ri->ri_font->stride; > - > - while (height--) { > - dp = rp; > - fb = fr[3] | (fr[2] << 8) | (fr[1] ><< 16) | > - (fr[0] << 24); > - fr += fs; > - DELTA(rp, ri->ri_stride, int32_t *); > - > - for (cnt = width; cnt; cnt--) { > - *dp++ = clr[(fb >> 31) & 1]; > - fb <<= 1; > - } > + /* double-pixel special cases for the common widths */ &
Re: [PATCH} Optimized rasops32 putchar
I believe it is mapped as normally cached right now, rather than uncached or write combining. Reads aren't ultra-slow, and the timings of 48 byte writes appear to involve a cacheline read. 128 byte writes are actually slower than 64 byte writes, which I guessed might be because of automatic prefetching kicking in and reading the following cacheline. Original Message Subject: Re: [PATCH} Optimized rasops32 putchar From: Mark Kettenis Date: Sat, June 27, 2020 7:56 am To: Cc: tech@openbsd.org > From: > Date: Fri, 26 Jun 2020 07:42:50 -0700 > > Optimized 32 bit character rendering with unrolled rows and pairwise > foreground / background pixel rendering. > > If it weren't for the 5x8 font, I would have just assumed everything > was an even width and made the fallback path also pairwise. > > In isolation, the 16x32 character case got 2x faster, but that wasn't > a huge real world speedup where the space rendering that was already > at memory bandwidth limits accounted for most of the character > rendering time. However, in combination with the previous fast > conditional console scrolling that removes most of the space rendering, > it becomes significant. > > I also found that at least the efi and intel framebuffers are not > currently mapped write combining, which makes this much slower than > it should be. Hi John, The framebuffer should be mapped write-combining. In OpenBSD this is requested by specifying the BUS_SPACE_MAP_PREFETCHABLE flag to bbus_space_map(9) when mapping the framebuffer. I'm fairly confident since until last January the initial mapping of the framebuffer that we used wasn't write-combining. And things were really, really slow. Cheers, Mark > Index: rasops32.c > === > RCS file: /cvs/src/sys/dev/rasops/rasops32.c,v > retrieving revision 1.10 > diff -u -p -r1.10 rasops32.c > --- rasops32.c 25 May 2020 09:55:49 - 1.10 > +++ rasops32.c 26 Jun 2020 14:34:06 - > @@ -65,9 +65,14 @@ rasops32_init(struct rasops_info *ri) > int > rasops32_putchar(void *cookie, int row, int col, u_int uc, uint32_t > attr) > { > - int width, height, cnt, fs, fb, clr[2]; > + int width, height, step, cnt, fs, b, f; > + uint32_t fb, clr[2]; > struct rasops_info *ri; > - int32_t *dp, *rp; > + int64_t *rp, q; > + union { > + int64_t q[4]; > + int32_t d[4][2]; > + } u; > u_char *fr; > > ri = (struct rasops_info *)cookie; > @@ -81,48 +86,128 @@ rasops32_putchar(void *cookie, int row, > return 0; > #endif > > - rp = (int32_t *)(ri->ri_bits + row*ri->ri_yscale + col*ri->ri_xscale); > + rp = (int64_t *)(ri->ri_bits + row*ri->ri_yscale + col*ri->ri_xscale); > > height = ri->ri_font->fontheight; > width = ri->ri_font->fontwidth; > + step = ri->ri_stride >> 3; > > - clr[0] = ri->ri_devcmap[(attr >> 16) & 0xf]; > - clr[1] = ri->ri_devcmap[(attr >> 24) & 0xf]; > + b = ri->ri_devcmap[(attr >> 16) & 0xf]; > + f = ri->ri_devcmap[(attr >> 24) & 0xf]; > + u.d[0][0] = b; u.d[0][1] = b; > + u.d[1][0] = b; u.d[1][1] = f; > + u.d[2][0] = f; u.d[2][1] = b; > + u.d[3][0] = f; u.d[3][1] = f; > > if (uc == ' ') { > + q = u.q[0]; > while (height--) { > - dp = rp; > - DELTA(rp, ri->ri_stride, int32_t *); > - > - for (cnt = width; cnt; cnt--) > - *dp++ = clr[0]; > + /* the general, pixel-at-a-time case is fast enough */ > + for (cnt = 0; cnt < width; cnt++) > + ((int *)rp)[cnt] = b; > + rp += step; > } > } else { > uc -= ri->ri_font->firstchar; > fr = (u_char *)ri->ri_font->data + uc * ri->ri_fontscale; > fs = ri->ri_font->stride; > - > - while (height--) { > - dp = rp; > - fb = fr[3] | (fr[2] << 8) | (fr[1] ><< 16) | > - (fr[0] << 24); > - fr += fs; > - DELTA(rp, ri->ri_stride, int32_t *); > - > - for (cnt = width; cnt; cnt--) { > - *dp++ = clr[(fb >> 31) & 1]; > - fb <<= 1; > - } > + /* double-pixel special cases for the common widths */ > + switch (width) { > + case 8: > + while (height--) { > + fb = fr[0]; > + rp[0] = u.q[fb >> 6]; > + rp[1] = u.q[(fb >> 4) & 3]; > + rp[2] = u.q[(fb >> 2) & 3]; > + rp[3] = u.q[fb & 3]; > + rp += step; > + fr += 1; > + } > + break; > + > + case 12: > + while (height--) { > + fb = fr[0]; > + rp[0] = u.q[fb >> 6]; > + rp[1] = u.q[(fb >> 4) & 3]; > + rp[2] = u.q[(fb >> 2) & 3]; > + rp[3] = u.q[fb & 3]; > + fb = fr[1]; > + rp[4] = u.q[fb >> 6]; > + rp[5] = u.q[(fb >> 4) & 3]; > + rp += s
Re: [PATCH} Optimized rasops32 putchar
> From: > Date: Fri, 26 Jun 2020 07:42:50 -0700 > > Optimized 32 bit character rendering with unrolled rows and pairwise > foreground / background pixel rendering. > > If it weren't for the 5x8 font, I would have just assumed everything > was an even width and made the fallback path also pairwise. > > In isolation, the 16x32 character case got 2x faster, but that wasn't > a huge real world speedup where the space rendering that was already > at memory bandwidth limits accounted for most of the character > rendering time. However, in combination with the previous fast > conditional console scrolling that removes most of the space rendering, > it becomes significant. > > I also found that at least the efi and intel framebuffers are not > currently mapped write combining, which makes this much slower than > it should be. Hi John, The framebuffer should be mapped write-combining. In OpenBSD this is requested by specifying the BUS_SPACE_MAP_PREFETCHABLE flag to bbus_space_map(9) when mapping the framebuffer. I'm fairly confident since until last January the initial mapping of the framebuffer that we used wasn't write-combining. And things were really, really slow. Cheers, Mark > Index: rasops32.c > === > RCS file: /cvs/src/sys/dev/rasops/rasops32.c,v > retrieving revision 1.10 > diff -u -p -r1.10 rasops32.c > --- rasops32.c25 May 2020 09:55:49 - 1.10 > +++ rasops32.c26 Jun 2020 14:34:06 - > @@ -65,9 +65,14 @@ rasops32_init(struct rasops_info *ri) > int > rasops32_putchar(void *cookie, int row, int col, u_int uc, uint32_t > attr) > { > - int width, height, cnt, fs, fb, clr[2]; > + int width, height, step, cnt, fs, b, f; > + uint32_t fb, clr[2]; > struct rasops_info *ri; > - int32_t *dp, *rp; > + int64_t *rp, q; > + union { > + int64_t q[4]; > + int32_t d[4][2]; > + } u; > u_char *fr; > > ri = (struct rasops_info *)cookie; > @@ -81,48 +86,128 @@ rasops32_putchar(void *cookie, int row, > return 0; > #endif > > - rp = (int32_t *)(ri->ri_bits + row*ri->ri_yscale + col*ri->ri_xscale); > + rp = (int64_t *)(ri->ri_bits + row*ri->ri_yscale + col*ri->ri_xscale); > > height = ri->ri_font->fontheight; > width = ri->ri_font->fontwidth; > + step = ri->ri_stride >> 3; > > - clr[0] = ri->ri_devcmap[(attr >> 16) & 0xf]; > - clr[1] = ri->ri_devcmap[(attr >> 24) & 0xf]; > + b = ri->ri_devcmap[(attr >> 16) & 0xf]; > + f = ri->ri_devcmap[(attr >> 24) & 0xf]; > + u.d[0][0] = b; u.d[0][1] = b; > + u.d[1][0] = b; u.d[1][1] = f; > + u.d[2][0] = f; u.d[2][1] = b; > + u.d[3][0] = f; u.d[3][1] = f; > > if (uc == ' ') { > + q = u.q[0]; > while (height--) { > - dp = rp; > - DELTA(rp, ri->ri_stride, int32_t *); > - > - for (cnt = width; cnt; cnt--) > - *dp++ = clr[0]; > + /* the general, pixel-at-a-time case is fast enough */ > + for (cnt = 0; cnt < width; cnt++) > + ((int *)rp)[cnt] = b; > + rp += step; > } > } else { > uc -= ri->ri_font->firstchar; > fr = (u_char *)ri->ri_font->data + uc * ri->ri_fontscale; > fs = ri->ri_font->stride; > - > - while (height--) { > - dp = rp; > - fb = fr[3] | (fr[2] << 8) | (fr[1] << 16) | > - (fr[0] << 24); > - fr += fs; > - DELTA(rp, ri->ri_stride, int32_t *); > - > - for (cnt = width; cnt; cnt--) { > - *dp++ = clr[(fb >> 31) & 1]; > - fb <<= 1; > - } > + /* double-pixel special cases for the common widths */ > + switch (width) { > + case 8: > + while (height--) { > + fb = fr[0]; > + rp[0] = u.q[fb >> 6]; > + rp[1] = u.q[(fb >> 4) & 3]; > + rp[2] = u.q[(fb >> 2) & 3]; > + rp[3] = u.q[fb & 3]; > + rp += step; > + fr += 1; > + } > + break; > + > + case 12: > + while (height--) { > + fb = fr[0]; > + rp[0] = u.q[fb >> 6]; > + rp[1] = u.q[(fb >> 4) & 3]; > + rp[2] = u.q[(fb >> 2) & 3]; > +
[PATCH} Optimized rasops32 putchar
Optimized 32 bit character rendering with unrolled rows and pairwise foreground / background pixel rendering. If it weren't for the 5x8 font, I would have just assumed everything was an even width and made the fallback path also pairwise. In isolation, the 16x32 character case got 2x faster, but that wasn't a huge real world speedup where the space rendering that was already at memory bandwidth limits accounted for most of the character rendering time. However, in combination with the previous fast conditional console scrolling that removes most of the space rendering, it becomes significant. I also found that at least the efi and intel framebuffers are not currently mapped write combining, which makes this much slower than it should be. Index: rasops32.c === RCS file: /cvs/src/sys/dev/rasops/rasops32.c,v retrieving revision 1.10 diff -u -p -r1.10 rasops32.c --- rasops32.c 25 May 2020 09:55:49 - 1.10 +++ rasops32.c 26 Jun 2020 14:34:06 - @@ -65,9 +65,14 @@ rasops32_init(struct rasops_info *ri) int rasops32_putchar(void *cookie, int row, int col, u_int uc, uint32_t attr) { - int width, height, cnt, fs, fb, clr[2]; + int width, height, step, cnt, fs, b, f; + uint32_t fb, clr[2]; struct rasops_info *ri; - int32_t *dp, *rp; + int64_t *rp, q; + union { + int64_t q[4]; + int32_t d[4][2]; + } u; u_char *fr; ri = (struct rasops_info *)cookie; @@ -81,48 +86,128 @@ rasops32_putchar(void *cookie, int row, return 0; #endif - rp = (int32_t *)(ri->ri_bits + row*ri->ri_yscale + col*ri->ri_xscale); + rp = (int64_t *)(ri->ri_bits + row*ri->ri_yscale + col*ri->ri_xscale); height = ri->ri_font->fontheight; width = ri->ri_font->fontwidth; + step = ri->ri_stride >> 3; - clr[0] = ri->ri_devcmap[(attr >> 16) & 0xf]; - clr[1] = ri->ri_devcmap[(attr >> 24) & 0xf]; + b = ri->ri_devcmap[(attr >> 16) & 0xf]; + f = ri->ri_devcmap[(attr >> 24) & 0xf]; + u.d[0][0] = b; u.d[0][1] = b; + u.d[1][0] = b; u.d[1][1] = f; + u.d[2][0] = f; u.d[2][1] = b; + u.d[3][0] = f; u.d[3][1] = f; if (uc == ' ') { + q = u.q[0]; while (height--) { - dp = rp; - DELTA(rp, ri->ri_stride, int32_t *); - - for (cnt = width; cnt; cnt--) - *dp++ = clr[0]; + /* the general, pixel-at-a-time case is fast enough */ + for (cnt = 0; cnt < width; cnt++) + ((int *)rp)[cnt] = b; + rp += step; } } else { uc -= ri->ri_font->firstchar; fr = (u_char *)ri->ri_font->data + uc * ri->ri_fontscale; fs = ri->ri_font->stride; - - while (height--) { - dp = rp; - fb = fr[3] | (fr[2] << 8) | (fr[1] << 16) | - (fr[0] << 24); - fr += fs; - DELTA(rp, ri->ri_stride, int32_t *); - - for (cnt = width; cnt; cnt--) { - *dp++ = clr[(fb >> 31) & 1]; - fb <<= 1; - } + /* double-pixel special cases for the common widths */ + switch (width) { + case 8: + while (height--) { + fb = fr[0]; + rp[0] = u.q[fb >> 6]; + rp[1] = u.q[(fb >> 4) & 3]; + rp[2] = u.q[(fb >> 2) & 3]; + rp[3] = u.q[fb & 3]; + rp += step; + fr += 1; + } + break; + + case 12: + while (height--) { + fb = fr[0]; + rp[0] = u.q[fb >> 6]; + rp[1] = u.q[(fb >> 4) & 3]; + rp[2] = u.q[(fb >> 2) & 3]; + rp[3] = u.q[fb & 3]; + fb = fr[1]; + rp[4] = u.q[fb >> 6]; + rp[5] = u.q[(fb >> 4) & 3]; + rp += step; + fr += 2; + } + break; + + case 16: + while (height--) { +