[PATCH 07/20] glamor: Add simple upload/download functions in glamor_transfer
Eric Anholt
eric at anholt.net
Thu Mar 20 13:24:21 PDT 2014
Markus Wick <markus at selfnet.de> writes:
> Am 2014-03-19 06:09, schrieb Keith Packard:
>> These use glTexSubimage2D for upload and glReadPixels for
>> download. There are a variety of interfaces to the basic function as
>> needed by the callers.
>>
>> Signed-off-by: Keith Packard <keithp at keithp.com>
>> ---
>> glamor/Makefile.am | 2 +
>> glamor/glamor_transfer.c | 260
>> +++++++++++++++++++++++++++++++++++++++++++++++
>> glamor/glamor_transfer.h | 55 ++++++++++
>> 3 files changed, 317 insertions(+)
>> create mode 100644 glamor/glamor_transfer.c
>> create mode 100644 glamor/glamor_transfer.h
>>
>> diff --git a/glamor/Makefile.am b/glamor/Makefile.am
>> index dec6467..9544923 100644
>> --- a/glamor/Makefile.am
>> +++ b/glamor/Makefile.am
>> @@ -23,6 +23,8 @@ libglamor_la_SOURCES = \
>> glamor_render.c \
>> glamor_gradient.c \
>> glamor_program.c \
>> + glamor_transfer.c \
>> + glamor_transfer.h \
>> glamor_transform.c \
>> glamor_trapezoid.c \
>> glamor_tile.c \
>> diff --git a/glamor/glamor_transfer.c b/glamor/glamor_transfer.c
>> new file mode 100644
>> index 0000000..d25c659
>> --- /dev/null
>> +++ b/glamor/glamor_transfer.c
>> @@ -0,0 +1,260 @@
>> +/*
>> + * Copyright © 2014 Keith Packard
>> + *
>> + * Permission to use, copy, modify, distribute, and sell this software
>> and its
>> + * documentation for any purpose is hereby granted without fee,
>> provided that
>> + * the above copyright notice appear in all copies and that both that
>> copyright
>> + * notice and this permission notice appear in supporting
>> documentation, and
>> + * that the name of the copyright holders not be used in advertising
>> or
>> + * publicity pertaining to distribution of the software without
>> specific,
>> + * written prior permission. The copyright holders make no
>> representations
>> + * about the suitability of this software for any purpose. It is
>> provided "as
>> + * is" without express or implied warranty.
>> + *
>> + * THE COPYRIGHT HOLDERS DISCLAIM ALL WARRANTIES WITH REGARD TO THIS
>> SOFTWARE,
>> + * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS, IN
>> NO
>> + * EVENT SHALL THE COPYRIGHT HOLDERS BE LIABLE FOR ANY SPECIAL,
>> INDIRECT OR
>> + * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS
>> OF USE,
>> + * DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
>> OTHER
>> + * TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
>> PERFORMANCE
>> + * OF THIS SOFTWARE.
>> + */
>> +
>> +#include "glamor_priv.h"
>> +#include "glamor_transfer.h"
>> +
>> +/* XXX a kludge for now */
>> +void
>> +glamor_format_for_pixmap(PixmapPtr pixmap, GLenum *format, GLenum
>> *type)
>> +{
>> + switch (pixmap->drawable.depth) {
>> + case 24:
>> + case 32:
>> + *format = GL_BGRA;
>> + *type = GL_UNSIGNED_INT_8_8_8_8_REV;
>> + break;
>> + case 16:
>> + *format = GL_RGB;
>> + *type = GL_UNSIGNED_SHORT_5_6_5;
>> + break;
>> + case 15:
>> + *format = GL_BGRA;
>> + *type = GL_UNSIGNED_SHORT_1_5_5_5_REV;
>> + break;
>> + case 8:
>> + *format = GL_ALPHA;
>> + *type = GL_UNSIGNED_BYTE;
>> + break;
>> + default:
>> + FatalError("Invalid pixmap depth %d\n",
>> pixmap->drawable.depth);
>> + break;
>> + }
>> +}
>> +
>> +/*
>> + * Write a region of bits into a pixmap
>> + */
>> +void
>> +glamor_upload_boxes(PixmapPtr pixmap, BoxPtr in_boxes, int in_nbox,
>> + int dx_src, int dy_src,
>> + int dx_dst, int dy_dst,
>> + uint8_t *bits, uint32_t byte_stride)
>> +{
>> + ScreenPtr screen = pixmap->drawable.pScreen;
>> + glamor_screen_private *glamor_priv =
>> glamor_get_screen_private(screen);
>> + glamor_pixmap_private *priv =
>> glamor_get_pixmap_private(pixmap);
>> + int box_x, box_y;
>> + int bytes_per_pixel =
>> pixmap->drawable.bitsPerPixel >> 3;
>> + GLenum type;
>> + GLenum format;
>> +
>> + glamor_format_for_pixmap(pixmap, &format, &type);
>> +
>> + glamor_get_context(glamor_priv);
>> +
>> + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
>> + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
>> +
>> + glPixelStorei(GL_UNPACK_ALIGNMENT, 4);
>> + glPixelStorei(GL_UNPACK_ROW_LENGTH, byte_stride /
>> bytes_per_pixel);
>> +
>> + glamor_pixmap_loop(priv, box_x, box_y) {
>> + BoxPtr box = glamor_pixmap_box_at(priv,
>> box_x, box_y);
>> + glamor_pixmap_fbo *fbo = glamor_pixmap_fbo_at(priv,
>> box_x, box_y);
>> + BoxPtr boxes = in_boxes;
>> + int nbox = in_nbox;
>> +
>> + glActiveTexture(GL_TEXTURE0);
>> + glBindTexture(GL_TEXTURE_2D, fbo->tex);
>> +
>> + while (nbox--) {
>> +
>> + /* compute drawable coordinates */
>> + int x1 = boxes->x1 + dx_dst;
>> + int x2 = boxes->x2 + dx_dst;
>> + int y1 = boxes->y1 + dy_dst;
>> + int y2 = boxes->y2 + dy_dst;
>> +
>> + if (x1 < box->x1)
>> + x1 = box->x1;
>> + if (box->x2 < x2)
>> + x2 = box->x2;
>
> x1 = MAX(x1, box->x1);
> x2 = MIN(x2, box->x2);
>
>> +
>> + if (x2 <= x1)
>> + continue;
>> +
>> + if (y1 < box->y1)
>> + y1 = box->y1;
>> + if (box->y2 < y2)
>> + y2 = box->y2;
>> +
>> + if (y2 <= y1)
>> + continue;
>
> So a box coverage pre-check should be moved to glamor_utils. Maybe also
> the box loop?
>
>> +
>> + glPixelStorei(GL_UNPACK_SKIP_ROWS, y1 - dy_dst + dy_src);
>
> Isn't this the same as an offset in glTexSubImage2D?
>
>> + glPixelStorei(GL_UNPACK_SKIP_PIXELS, x1 - dx_dst +
>> dx_src);
>> +
>> + glTexSubImage2D(GL_TEXTURE_2D, 0,
>> + x1 - box->x1, y1 - box->y1,
>> + x2 - x1, y2 - y1,
>> + format, type,
>> + bits);
>
> I think in theory, this won't do any redundant uploads, but we shouldn't
> do so many uploading calls. As we pretty much always need a temporary
> staging buffer (in the driver) to avoid stalling, we could do this
> explicit with a PBO.
> But PBO shouldn't be used with uncommon pixel types as some of them are
> emulated on the CPU...
>
>> + }
>> + }
>> +
>> + glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
>> + glPixelStorei(GL_UNPACK_SKIP_ROWS, 0);
>> + glPixelStorei(GL_UNPACK_SKIP_PIXELS, 0);
>> + glamor_put_context(glamor_priv);
>> +}
>> +
>> +/*
>> + * Upload a region of data
>> + */
>> +
>> +void
>> +glamor_upload_region(PixmapPtr pixmap, RegionPtr region,
>> + int region_x, int region_y,
>> + uint8_t *bits, uint32_t byte_stride)
>> +{
>> + glamor_upload_boxes(pixmap, RegionRects(region),
>> RegionNumRects(region),
>> + -region_x, -region_y,
>> + 0, 0,
>> + bits, byte_stride);
>> +}
>> +
>> +/*
>> + * Take the data in the pixmap and stuff it back into the FBO
>> + */
>> +void
>> +glamor_upload_pixmap(PixmapPtr pixmap)
>> +{
>> + BoxRec box;
>> +
>> + box.x1 = 0;
>> + box.x2 = pixmap->drawable.width;
>> + box.y1 = 0;
>> + box.y2 = pixmap->drawable.height;
>> + glamor_upload_boxes(pixmap, &box, 1, 0, 0, 0, 0,
>> + pixmap->devPrivate.ptr, pixmap->devKind);
>> +}
>> +
>> +/*
>> + * Read stuff from the pixmap FBOs and write to memory
>> + */
>> +void
>> +glamor_download_boxes(PixmapPtr pixmap, BoxPtr in_boxes, int in_nbox,
>> + int dx_src, int dy_src,
>> + int dx_dst, int dy_dst,
>> + uint8_t *bits, uint32_t byte_stride)
>> +{
>> + ScreenPtr screen = pixmap->drawable.pScreen;
>> + glamor_screen_private *glamor_priv =
>> glamor_get_screen_private(screen);
>> + glamor_pixmap_private *priv = glamor_get_pixmap_private(pixmap);
>> + int box_x, box_y;
>> + int bytes_per_pixel = pixmap->drawable.bitsPerPixel >> 3;
>> + GLenum type;
>> + GLenum format;
>> +
>> + glamor_format_for_pixmap(pixmap, &format, &type);
>> +
>> + glamor_get_context(glamor_priv);
>> +
>> + glPixelStorei(GL_PACK_ALIGNMENT, 4);
>> + glPixelStorei(GL_PACK_ROW_LENGTH, byte_stride / bytes_per_pixel);
>> +
>> + glamor_pixmap_loop(priv, box_x, box_y) {
>> + BoxPtr box = glamor_pixmap_box_at(priv,
>> box_x, box_y);
>> + glamor_pixmap_fbo *fbo = glamor_pixmap_fbo_at(priv,
>> box_x, box_y);
>> + BoxPtr boxes = in_boxes;
>> + int nbox = in_nbox;
>> +
>> + glBindFramebuffer(GL_READ_FRAMEBUFFER, fbo->fb);
>> +
>> + while (nbox--) {
>> +
>> + /* compute drawable coordinates */
>> + int x1 = boxes->x1 + dx_src;
>> + int x2 = boxes->x2 + dx_src;
>> + int y1 = boxes->y1 + dy_src;
>> + int y2 = boxes->y2 + dy_src;
>> +
>> + if (x1 < box->x1)
>> + x1 = box->x1;
>> + if (box->x2 < x2)
>> + x2 = box->x2;
>> +
>> + if (y1 < box->y1)
>> + y1 = box->y1;
>> + if (box->y2 < y2)
>> + y2 = box->y2;
>> +
>> + if (x2 <= x1)
>> + continue;
>> + if (y2 <= y1)
>> + continue;
>> +
>> + glPixelStorei(GL_PACK_SKIP_PIXELS, x1 - dx_src + dx_dst);
>> + glPixelStorei(GL_PACK_SKIP_ROWS, y1 - dy_src + dy_dst);
>> + glReadPixels(x1 - box->x1, y1 - box->y1, x2 - x1, y2 -
>> y1, format, type, bits);
>
> Here, a PBO is more important to not stall once per box. i965 may not
> stall, but on dedicated GPUs, the framebuffer can't be accessed without
> a round trip time.
That's assuming, though, that nbox > 1 happens in the real world. The
advantage to skipping the PBO is that if your implementation can map
user memory as BOs, you save a copy.
i965 currently doesn't, but it could soon, if the kernel team figures
out how to land that code. It's also something I expect the discrete
drivers should be interested in doing, since managing GL BO migration
between address spaces is hard.
> So in the end, I suggest to use PBO all the time. It's likely an
> overhead compared to the internal staging buffer of the driver, but imo
> this shouldn't matter compared with the memcpy time.
> But on shared memory systems, this is a huge overhead as it has to
> memcpy the texture twice. The driver doesn't have to do this when either
> the texture isn't in use (unlikely) or we update the complete texture
> (not that unlikely).
PBOs are so hard for a driver to get right. You've got these hints
associated with them, which application authors emit random noise into.
Because of that, you don't know when you're asked to copy data into a BO
if the user is going to be mapping it with the CPU next, or using it as
the source for some GL operation on the GPU. What that means is that I
expecct on discrete you'll often end up doing *another* copy to get the
data mappable, which X can then memcpy out of.
Basically, PBOs tend to require performance data indicating that
they're a win before I can get excited about seeing them used.
> How should we deal with plattforms without ARB_pbo? I think glamor
> should work, but performance doesn't matter that hard. I don't want to
> fork this function to work fast with and without PBO.
I don't think we should bother. PBOs are an ancient feature.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 818 bytes
Desc: not available
URL: <http://lists.x.org/archives/xorg-devel/attachments/20140320/5385b83b/attachment.sig>
More information about the xorg-devel
mailing list