This patch adds a dozen more "shortcut" scanline-processing functions
to pixel-flingers. All of them avoid using the JIT for the corresponding
operation (on ARM), or using the generic and _extremely_ slow 'scanline'
function (on x86, where there is no JIT).
The shortcuts were selected by running the system under emulation
(build full_x86-eng, then launch emulator-x86), and correspond to
operations that are in use when using the system's typical UI features.
This makes it much more responsive and amenable to testing most
applications, at least those that don't use OpenGL ES heavily.
Note that HW OpenGLES emulation is under completion and should solve this
problem entirely, though is not there yet.
Change-Id: I9c73ba21ad158d6cc5532fabe7ed2419e00ecb3f
Merge commit '8e0e372a388434a0553810e2b958e59a26a6bd96' into gingerbread-plus-aosp
* commit '8e0e372a388434a0553810e2b958e59a26a6bd96':
Set PROT_EXEC on the whole pixelflinger code cache.
The pointer difference between word pointers is a number
of words, and it needs to be multiplied by the size of a word
to get a proper byte size.
Without this, we tend to see crashes when the code crosses
a page boundary.
Bug: 3026204
Bug: 3097482
Change-Id: I37776d26d5afcdb1da71680de02fbb95e6548371
According to the ARM Architecture Reference Manual, the comment on
STM instruction should be in reverse order.
Change-Id: I4af852a0478798ff7b02ab9c29c68e320ff78696
Signed-off-by: Kan-Ru Chen <kanru@0xlab.org>
Merge commit '4ea1a52b1bbbd6e78c5909c75d773416108f5b84' into gingerbread-plus-aosp
* commit '4ea1a52b1bbbd6e78c5909c75d773416108f5b84':
libpixelflinger: ARMv6 specific objects are not used. Remove.
Merge commit '6090dacd1894429baaf13f7b30b2f6e9e2c1617f' into gingerbread-plus-aosp
* commit '6090dacd1894429baaf13f7b30b2f6e9e2c1617f':
libpixelflinger: Move codeflinger test function to test-opengl-codegen
From the functionality, the function gglBitBlti() implemented bit-block
(image) transfer operation in which several bitmaps are combined into
one using a raster operator. Renaming to gglBitBlit() should be
explicit.
Change-Id: I23886623b5421131cf7c7bfb8ec67c3809566c93
This patch attempts to remove the generation of libpixelflinger_armv6,
which contains the ARMv6 specific object - rotate90CW_4x4_16v6.o. But,
libpixelflinger_armv6 is not actually used by other projects.
Change-Id: I0e26fa2cf8b4dacfd208dd4279539c3e5b07580e
There is no need to include codeflinger test function in
libpixelflinger. Instead, it should exist in test-opengl-codegen.
Change-Id: Ic5b0bd1f754b2ac678d4e16565568c70ceb8d325
This introduces UBFX instruction generation abilities to the Pixelflinger JIT,
and also modifies the component extraction function to generate the
instruction.
The extract function contains defines to prevent generation of UBFX on pre-v7
cores. The JIT itself retains the ability to produce the instruction even on
v5/6.
This patch only generates UBFX when MOV, AND or BIC can't be used. Based on
the TRM, this appears to be faster on A9 than using UBFX in all cases.
On startup, Pixelflinger JITs three chunks of code. UBFX improves these as
follows:
00000077:03515104_00000000_00000000
(Blends a single colour into an RGB565 buffer.)
Before: 27 inst/pixel, After: 24 inst/pixel, Improvement: 12.5%
00000077:03545404_00000A01_00000000
(Blends RGBA8888 texture into an RGB565 buffer using alpha.)
Before: 30 inst/pixel, After: 27 inst/pixel, Improvement: 11.1%
00000077:03545404_00000A04_00000000
(Blends RGB565 texture into an RGB565 buffer using alpha.)
Before: 29 inst/pixel, After: 27 inst/pixel, Improvement: 7.4%
Instead of allocating memory from the (non executable) heap,
allocate memory using mspace and ensure that we use mprotect
to mark it as PROT_EXEC. This allows pixelflinger to
continue to work even when NX protections are enabled.
Testing: Using the ApiDemos market app, verify that
Apidemos -> Graphics -> OpenGL ES -> GLSurfaceView
works when "adb shell setprop debug.egl.hw 0" is set.
Change-Id: Ib569cd2543c6fa25688ee76325a712bc2347450b
The Pixelflinger disassembler does not handle LDM addressing modes correctly,
assuming that the P and U bits in the instruction mean the same in both LDM and
STM. This results in the disassembler producing sequences like:
stmfd r13!, {r4-r11, r14}
...
...
...
ldmea r13!, {r4-r11, r14}
This small patch fixes it by EORing the P and U bits with the Load/Store bit.
Change-Id: Ic7a1556642c4e29415fc3697019f1239b6c26fc2
This is a shortcut for the needs descriptor
00000077:03515104_00000000_00000000. It requires blending a single 32-bit
colour value into a 16-bit framebuffer.
It's used when fading out the screen, eg. when a modal requester pops-up.
The PF JIT produces code for this using 24 instructions/pixel. The NEON
implementation requires 2.1 instructions/pixel. Performance hasn't been
benchmarked, but the improvement is quite visible.
This code has only been tested by inspection of the fading effect described
above, when press+holding a finger on the home screen to pop up the
Shortcuts/Widgets/Folders/Wallpaper requester.
Along with the NEON version, a fallback v5TE implementation is also provided.
This ARM version of col32cb16blend is not fully optimised, but is a reasonable
implementation, and better than the version produced by the JIT. It is here as
a fallback, if NEON is not available.
* Add support for UXTB16 to the disassembler
* Add encoding of the UXTB16 instruction to the Pixelflinger JIT.
Introducing the UXTB16 instruction allows removal of some masking code, and is
beneficial from a pipeline point of view - lots of UXTB16 followed by MUL
sequences.
Also, further rescheduling and use of SMULWB brings extra performance
improvements.
* Use UXTB16 in bilinear filtered texturing
Uses UXTB16 to extract channels for SIMD operations, rather than creating and
ANDing with masks. Saves a register and is faster on A8, as UXTB16 result can
feed into first stage of multiply, unlike AND.
Also, used SMULWB rather than SMULBB, which allows removal of MOVs used to
rescale results.
Code has been scheduled for A8 pipeline, specifically aiming to allow
multiplies to issue in pipeline 0, for efficient dual issue operation.
Testing on SpriteMethodTest (http://code.google.com/p/apps-for-android/) gives
8% improvement (12.7 vs. 13.7 fps.)
SMULBB to SMULWB trick could be used in <v6 code path, but this hasn't been
implemented.
when ONE / ONE_MINUS_SRC_ALPHA blending mode was used, the code wasn't saturating the color component.
the reason was that this mode is used for premltiplied alpha blending, however, if used with a non
premultiplied source, the color component would wrap.
unfortunately, this costs 6 extra cycles per pixels, however... "correctness" prevails.
this should not impact the UI since it's using h/w acceleration most of the time it also doesn't
impact games which should be using h/w GL. This change will slow the emulator down a bit.
A better fix would be to banish this qemu tracing stuff from libhardware
and also banish it from non-emulator builds, but this at least gets the
minimal build building again.
