Skip to contentSkip to conversation

uxa/glamor: performance tuning.

PATCH In review [email protected] 2 participants 7 messages Dec 27, 2011 – Dec 29, 2011

lore
Share & tools
lore
#0 lore raw

cover uxa/glamor: performance tuning.

From: Zhigang Gong <[email protected]>

According to your advice, split the previous 1 patch to 4. And slightly changed the patch of create GC. The previous verison will override the GC ops to glamor's GC ops which is really not what I expected. If there are still some other unexpected side effects, please help to point out thanks. This patchset need to work with the latest glamor. I will move glamor to official repository this week and gave an initial version 0.1.0 and will bump the version number each time the API changed.

-

#1 lore raw

PATCH 1/4 uxa/glamor: Remove dead code.

From: Zhigang Gong <[email protected]>

---
 src/intel_glamor.c |   13 -------------
 src/intel_glamor.h |    2 --
 2 files changed, 0 insertions(+), 15 deletions(-)
src/intel_glamor.c +0 -13
index 8daa4b1..2b56ae5 100644
@@ -192,19 +192,6 @@ intel_glamor_flush(intel_screen_private * intel)
 }
 
 Bool
-intel_glamor_create_screen_image(ScreenPtr screen, int handle, int stride)
-{
-	ScrnInfoPtr scrn = xf86Screens[screen->myNum];
-	intel_screen_private *intel;
-
-	intel = intel_get_screen_private(scrn);
-	if ((intel->uxa_flags & UXA_USE_GLAMOR) == 0)
-		return TRUE;
-
-	return glamor_egl_create_textured_screen(screen, handle, stride);
-}
-
-Bool
 intel_glamor_close_screen(ScreenPtr screen)
 {
 	ScrnInfoPtr scrn = xf86Screens[screen->myNum];
src/intel_glamor.h +0 -2
index 1374588..3065132 100644
@@ -35,7 +35,6 @@
 Bool intel_glamor_pre_init(ScrnInfoPtr scrn);
 Bool intel_glamor_init(ScreenPtr screen);
 Bool intel_glamor_create_screen_resources(ScreenPtr screen);
-Bool intel_glamor_create_screen_image(ScreenPtr screen, int handle, int stride);
 Bool intel_glamor_close_screen(ScreenPtr screen);
 void intel_glamor_free_screen(int scrnIndex, int flags);
 
@@ -51,7 +50,6 @@ PixmapPtr intel_glamor_create_pixmap(ScreenPtr screen, int w, int h,
 static inline Bool intel_glamor_pre_init(ScrnInfoPtr scrn) { return TRUE; }
 static inline Bool intel_glamor_init(ScreenPtr screen) { return TRUE; }
 static inline Bool intel_glamor_create_screen_resources(ScreenPtr screen) { return TRUE; }
-static inline Bool intel_glamor_create_screen_image(ScreenPtr screen, int handle, int stride) { return TRUE; }
 static inline Bool intel_glamor_close_screen(ScreenPtr screen) { return TRUE; }
 static inline void intel_glamor_free_screen(int scrnIndex, int flags) { }
 
Signature · 1 line
1.7.4.4
#2 lore raw

PATCH 2/4 uxa/glamor: Remove extraneous flush.

In reply to [email protected]

From: Zhigang Gong <[email protected]>

When glamor is enabled, a pixmap will not be accessed by the UXA's accelerated function. Only un-accelerated funtion may access those pixmap, and before each un-accelerated rendering, it calls uxa_prepare_access which will do glFlush there.

---
 src/intel_glamor.c |    1 -
 1 files changed, 0 insertions(+), 1 deletions(-)
src/intel_glamor.c +0 -1
index 2b56ae5..669b1b8 100644
@@ -135,7 +135,6 @@ intel_glamor_finish_access(PixmapPtr pixmap, uxa_access_t access)
 		break;
 	case UXA_GLAMOR_ACCESS_RW:
 		intel_glamor_need_flush(&pixmap->drawable);
-		glamor_block_handler(pixmap->drawable.pScreen);
 		break;
 	default:
 		ErrorF("Invalid access mode %d\n", access);
Signature · 1 line
1.7.4.4
#3 lore raw

PATCH 3/4 uxa/glamor: Let glamor do the GC validation.

In reply to [email protected]

From: Zhigang Gong <[email protected]>

If we are using GLAMOR, then a tile pixmap or stipple pixmap may be pure glamor pixmap and thus UXA doesn't know how to render to them, we should let glamor to do the validation.

---
 uxa/uxa-glamor.h |    1 +
 uxa/uxa.c        |   12 ++++++++++++
 2 files changed, 13 insertions(+), 0 deletions(-)
uxa/uxa-glamor.h +1 -0
index 71a9c29..2b4b452 100644
@@ -51,6 +51,7 @@
 #define glamor_trapezoids_nf(...)	FALSE
 #define glamor_triangles_nf(...)	FALSE
 #define glamor_add_traps_nf(...)	FALSE
+#define glamor_create_gc(...)		FALSE
 #endif
 
 #endif /* UXA_GLAMOR_H */
uxa/uxa.c +12 -0
index 5b3d709..56bc233 100644
@@ -38,6 +38,7 @@
 #include <X11/fonts/fontstruct.h>
 #include "dixfontstr.h"
 #include "uxa.h"
+#include "uxa-glamor.h"
 
 #if HAS_DEVPRIVATEKEYREC
 DevPrivateKeyRec uxa_screen_index;
@@ -183,11 +184,21 @@ void uxa_finish_access(DrawablePtr pDrawable, uxa_access_t access)
 static void
 uxa_validate_gc(GCPtr pGC, unsigned long changes, DrawablePtr pDrawable)
 {
+	uxa_screen_t *uxa_screen = uxa_get_screen(pGC->pScreen);
 	/* fbValidateGC will do direct access to pixmaps if the tiling has
 	 * changed.
 	 * Preempt fbValidateGC by doing its work and masking the change out, so
 	 * that we can do the Prepare/finish_access.
 	 */
+
+	/* If we are using GLAMOR, then the tile or stipple pixmap
+	 * may be pure GLAMOR pixmap, then we should let the glamor
+	 * to do the validation.
+	 */
+	if (uxa_screen->info->flags & UXA_USE_GLAMOR) {
+		glamor_validate_gc(pGC, changes, pDrawable);
+		goto set_ops;
+	}
 #ifdef FB_24_32BIT
 	if ((changes & GCTile) && fbGetRotatedPixmap(pGC)) {
 		(*pGC->pScreen->DestroyPixmap) (fbGetRotatedPixmap(pGC));
@@ -256,6 +267,7 @@ uxa_validate_gc(GCPtr pGC, unsigned long changes, DrawablePtr pDrawable)
 		fbValidateGC(pGC, changes, pDrawable);
 	}
 
+set_ops:
 	pGC->ops = (GCOps *) & uxa_ops;
 }
 
Signature · 1 line
1.7.4.4
#4 lore raw

PATCH 4/4 uxa/glamor: Create glamor pixmap by default.

In reply to [email protected]

From: Zhigang Gong <[email protected]>

As a pure glamor pixmap has a local texture rather than bind a pixmap to a external BO. This can avoid some unecessary flush, and can achieve better performance. The testing on my machine shows that aa10text/rgb10text get about 20-30% performance improvement.

---
 src/intel_uxa.c |    5 +++++
 1 files changed, 5 insertions(+), 0 deletions(-)
src/intel_uxa.c +5 -0
index 9d74554..ca3220a 100644
@@ -1026,6 +1026,11 @@ intel_uxa_create_pixmap(ScreenPtr screen, int w, int h, int depth,
 	struct intel_pixmap *priv;
 	PixmapPtr pixmap, new_pixmap = NULL;
 
+	if (!(usage & INTEL_CREATE_PIXMAP_DRI2)) {
+		pixmap = intel_glamor_create_pixmap(screen, w, h, depth, usage);
+		if (pixmap)
+			return pixmap;
+	}
 	if (w > 32767 || h > 32767)
 		return NullPixmap;
 
Signature · 1 line
1.7.4.4
#5 lore raw

Re: [PATCH 4/4] uxa/glamor: Create glamor pixmap by default.

Quoted 7 lines · On Tue, 27 Dec 2011 17:09:18 +0800, [email protected] wrote:

From: Zhigang Gong <[email protected]>

As a pure glamor pixmap has a local texture rather than bind a pixmap to a external BO. This can avoid some unecessary flush, and can achieve better performance. The testing on my machine shows that aa10text/rgb10text get about 20-30% performance improvement.

How do we get the name back for a pixmap created by glamor and then attached to a DRI2Drawable? And then once that name is exposed, how do you prevent GL from recreating the bo attached to the pixmap? -Chris

Signature · 1 line
Chris Wilson, Intel Open Source Technology Centre
#6 lore raw

Re: [PATCH 4/4] uxa/glamor: Create glamor pixmap by default.

Quoted 18 lines

-----Original Message-----
From: Chris Wilson [mailto:[email protected]]
Sent: Wednesday, December 28, 2011 10:27 PM
To: [email protected]
Cc: [email protected]; [email protected]; [email protected]
Subject: Re: [PATCH 4/4] uxa/glamor: Create glamor pixmap by default.

Quoted 7 lines · On Tue, 27 Dec 2011 17:09:18 +0800, [email protected] wrote:

From: Zhigang Gong <[email protected]>

As a pure glamor pixmap has a local texture rather than bind a pixmap to a external BO. This can avoid some unecessary flush, and can achieve better performance.
The testing on my machine shows that aa10text/rgb10text get about 20-30% performance improvement.

Quoted 3 lines

How do we get the name back for a pixmap created by glamor and then attached to a DRI2Drawable? And then once that name is exposed, how do you prevent GL from recreating the bo attached to the pixmap?

You are right, this patch will break the DRI2 usage. And currently, one solution come to my mind:

At I830DRI2Create buffer, After we get a non-null pixmap at get_front_buffer

we check whether The pixmap has a valid bo, if it doesn't, we will think this is
a pure glamor pixmap (maybe we can check more fields, such as zero devPrivate.ptr)
Then we will create new a valid DRI2 pixmap, then we call fixup_glamor to unbind
the bo from the new DRI2 pixmap and then bind it to the old pure glamor pixmap
and call intel_create_textured_pixmap to create a coherent glamor-drm pixmap, the old pixmap's pointer will not change, only the type will be changed
to a glamor-drm and has a valid bo attached to it.

Does this way work? And will it bring too much overhead.

Base on my current testing, change to use pure glamor pixmap for non-dri2 pixmap really get much better performance then use coherent glamor-drm pixmap.

Any comments?

Quoted 4 lines

-Chris

--
Chris Wilson, Intel Open Source Technology Centre

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx