On Sat, 2009-08-22 at 04:18 -0700, Dave Airlie wrote: > src/radeon_exa.c | 3 +-- > src/radeon_exa_funcs.c | 35 +++++++++++++++++++++++++++++++++++ > 2 files changed, 36 insertions(+), 2 deletions(-) > > New commits: > commit bac224912c750dc1c85ff2d9b8526dad6c23b572 > Author: Dave Airlie <[email protected]> > Date: Sat Aug 22 21:17:59 2009 +1000 > > radeon: don't spec any initial placement for pixmaps. > > allow the first use to decide placement. > > diff --git a/src/radeon_exa.c b/src/radeon_exa.c > index ec65722..3f3c9ba 100644 > --- a/src/radeon_exa.c > +++ b/src/radeon_exa.c > @@ -391,8 +391,7 @@ void *RADEONEXACreatePixmap(ScreenPtr pScreen, int size, > int align) > return new_priv; > > new_priv->bo = radeon_bo_open(info->bufmgr, 0, size, > - align, RADEON_GEM_DOMAIN_VRAM | > - RADEON_GEM_DOMAIN_GTT, 0); > + align, 0, 0); > if (!new_priv->bo) { > xfree(new_priv); > ErrorF("Failed to alloc memory\n");
IME this has a significant impact on 2D performance (e.g. x11perf -aa10text drops from about 325K/s to about 225K/s, and this is noticeable in real apps), until the kernel side avoids unnecessary BO copies. Also, current EXA (with EXA_MIXED_PIXMAPS) actually only calls the driver CreatePixmap(2) hook just before the first accelerated operation on the pixmap. So I'd suggest setting RADEON_GEM_DOMAIN_VRAM | RADEON_GEM_DOMAIN_GTT in both hooks. > commit 77f98717d825162da106c6898cdbcbdf5c984ae6 > Author: Dave Airlie <[email protected]> > Date: Sat Aug 22 21:16:25 2009 +1000 > > exa/cs: add DFS from GTT optimisation > > This uses the new libdrm busy interface, once I had this in place > I added a error if this happened and it does on my desktop here, > so may as well add the optimisation that used to be in my old KMS tree. > > Signed-off-by: Dave Airlie <[email protected]> > > diff --git a/src/radeon_exa_funcs.c b/src/radeon_exa_funcs.c > index e48317f..f937c4e 100644 > --- a/src/radeon_exa_funcs.c > +++ b/src/radeon_exa_funcs.c > @@ -507,6 +507,33 @@ out: > } > > static Bool > +RADEONDownloadFromScreenGTT(PixmapPtr pSrc, int x, int y, int w, > + int h, char *dst, int dst_pitch) > +{ > + struct radeon_exa_pixmap_priv *driver_priv; > + int src_pitch = exaGetPixmapPitch(pSrc); > + int bpp = pSrc->drawable.bitsPerPixel; > + int src_offset; > + int r; > + > + driver_priv = exaGetPixmapDriverPrivate(pSrc); > + r = radeon_bo_map(driver_priv->bo, 0); > + if (r) > + return FALSE; > + > + src_offset = (x * bpp / 8) + (y * src_pitch); > + w *= bpp / 8; > + > + while (h--) { > + memcpy(dst, driver_priv->bo->ptr + src_offset, w); > + src_offset += src_pitch; > + dst += dst_pitch; > + } > + radeon_bo_unmap(driver_priv->bo); > + return TRUE; > +} > + > +static Bool > RADEONDownloadFromScreenCS(PixmapPtr pSrc, int x, int y, int w, > int h, char *dst, int dst_pitch) > { > @@ -519,11 +546,19 @@ RADEONDownloadFromScreenCS(PixmapPtr pSrc, int x, int > y, int w, > unsigned bpp = pSrc->drawable.bitsPerPixel; > uint32_t scratch_pitch = (w * bpp / 8 + 63) & ~63; > Bool r; > + uint32_t src_domain; > + int busy; > > if (bpp < 8) > return FALSE; > > driver_priv = exaGetPixmapDriverPrivate(pSrc); > + > + busy = radeon_bo_is_busy(driver_priv->bo, &src_domain); > + > + if (src_domain == RADEON_GEM_DOMAIN_GTT) > + return RADEONDownloadFromScreenGTT(pSrc, x, y, w, h, > + dst, dst_pitch); RADEONDownloadFromScreenGTT is superfluous, just return FALSE from RADEONDownloadFromScreenCS. Also, this doesn't take into account any unflushed operations involving the BO. Below is what I've been testing for DFS / UTS. diff --git a/src/radeon_exa_funcs.c b/src/radeon_exa_funcs.c index bada4fb..abb8be3 100644 --- a/src/radeon_exa_funcs.c +++ b/src/radeon_exa_funcs.c @@ -459,6 +459,7 @@ RADEONUploadToScreenCS(PixmapPtr pDst, int x, int y, int w, int h, struct radeon_bo *scratch; unsigned size; uint32_t datatype = 0; + uint32_t dst_domain; uint32_t dst_pitch_offset; unsigned bpp = pDst->drawable.bitsPerPixel; uint32_t scratch_pitch = (w * bpp / 8 + 63) & ~63; @@ -470,6 +471,18 @@ RADEONUploadToScreenCS(PixmapPtr pDst, int x, int y, int w, int h, driver_priv = exaGetPixmapDriverPrivate(pDst); + /* If we know the BO won't be busy, don't bother */ + if (driver_priv->bo->cref == 1 && + !radeon_bo_is_busy(driver_priv->bo, &dst_domain)) +{ + static Bool printed; + if (!printed) { + ErrorF("%s: Bailing because BO is idle\n", __func__); + printed = TRUE; + } + return FALSE; +} + size = scratch_pitch * h; scratch = radeon_bo_open(info->bufmgr, 0, size, 0, RADEON_GEM_DOMAIN_GTT, 0); if (scratch == NULL) { @@ -519,6 +532,7 @@ RADEONDownloadFromScreenCS(PixmapPtr pSrc, int x, int y, int w, struct radeon_bo *scratch; unsigned size; uint32_t datatype = 0; + uint32_t src_domain = 0; uint32_t src_pitch_offset; unsigned bpp = pSrc->drawable.bitsPerPixel; uint32_t scratch_pitch = (w * bpp / 8 + 63) & ~63; @@ -529,6 +543,30 @@ RADEONDownloadFromScreenCS(PixmapPtr pSrc, int x, int y, int w, driver_priv = exaGetPixmapDriverPrivate(pSrc); + /* If we know the BO will end up in GTT anyway, don't bother */ + if (driver_priv->bo->cref > 1) { + src_domain = driver_priv->bo->space_accounted & 0xffff; + if (!src_domain) + src_domain = driver_priv->bo->space_accounted >> 16; + + if ((src_domain & (RADEON_GEM_DOMAIN_GTT | RADEON_GEM_DOMAIN_VRAM)) == + (RADEON_GEM_DOMAIN_GTT | RADEON_GEM_DOMAIN_VRAM)) + src_domain = 0; + } + + if (!src_domain) + radeon_bo_is_busy(driver_priv->bo, &src_domain); + + if (src_domain == RADEON_GEM_DOMAIN_GTT) +{ + static Bool printed; + if (!printed) { + ErrorF("%s: Bailing because BO is in GTT\n", __func__); + printed = TRUE; + } + return FALSE; +} + size = scratch_pitch * h; scratch = radeon_bo_open(info->bufmgr, 0, size, 0, RADEON_GEM_DOMAIN_GTT, 0); if (scratch == NULL) { -- Earthling Michel Dänzer | http://www.vmware.com Libre software enthusiast | Debian, X and DRI developer _______________________________________________ xorg-driver-ati mailing list [email protected] http://lists.x.org/mailman/listinfo/xorg-driver-ati
