874369 - Use normal memory instead of Shmem when passing textures between threads on the same process

Status: RESOLVED FIXED

(Core :: Graphics: Layers, defect)

Description

[Matt Woodrow (:mattwoodrow)]

Attachment: Use raw memory

We spend significant amounts of time allocating Shmem with OMTC, and we are only sharing memory between threads in the same process.

This lets us just normal memory instead in this case, and makes everything happier.

Comment 1

[Matt Woodrow (:mattwoodrow)]

Attachment: Use raw memory v2

Forgot to qref and include a memory leak fix

Comment 2

[Bas Schouten (:bas.schouten)]

Comment on attachment 752090 [details] [diff] [review]
Use raw memory v2

Review of attachment 752090 [details] [diff] [review]:
-----------------------------------------------------------------

::: gfx/layers/ipc/ISurfaceAllocator.cpp
@@ +90,5 @@
> +    gfxImageFormat format = aContent == gfxASurface::CONTENT_COLOR_ALPHA ?
> +                            gfxASurface::ImageFormatARGB32 :
> +                            gfxASurface::ImageFormatRGB24;
> +    int32_t stride = gfxASurface::FormatStrideForWidth(format, aSize.width);
> +    unsigned char *data = new unsigned char[stride * aSize.height];

nit: For consistency replace unsigned char by uint8_t.

::: gfx/layers/ipc/ShadowLayerUtilsMac.cpp
@@ +38,5 @@
> +    const MemoryImage& image = aSurface.get_MemoryImage();
> +    gfxASurface::gfxImageFormat format
> +      = static_cast<gfxASurface::gfxImageFormat>(image.format());
> +
> +    nsRefPtr<gfxASurface> surf =

Hrm, I wonder if we always want a Quartz surface here or if we want the ability to do something else. For now this should be fine though.

Comment 3

[Matt Woodrow (:mattwoodrow)]

https://hg.mozilla.org/integration/mozilla-inbound/rev/85d108f0b422

Comment 4

[Matt Woodrow (:mattwoodrow)]

Backed out for causing b2g reftest failures: https://hg.mozilla.org/integration/mozilla-inbound/rev/09ff5b55b27a

Comment 5

[Matt Woodrow (:mattwoodrow)]

https://hg.mozilla.org/integration/mozilla-inbound/rev/fb160edaec7c

Re-landed with the content -> format conversion implemented correctly.

Comment 6

[Ryan VanderMeulen [:RyanVM]]

https://hg.mozilla.org/mozilla-central/rev/fb160edaec7c

Comment 7

[Oleg Romashin (:romaxa)]

sounds like this fix breaking OMTC SW rendering... 

Assertion failure: aDescriptorType == SurfaceDescriptor::TShmem (We can only support Shmem currently), at mozilla-central/gfx/layers/basic/BasicCompositor.cpp:74