This reverts commit ab5265ebe6.
This broke the mips64 build (which, stupidly, I hadn't tested):
system/core/libpixelflinger/codeflinger/MIPS64Assembler.h:24:35: fatal error: tinyutils/KeyedVector.h: No such file or directory
#include "tinyutils/KeyedVector.h"
Change-Id: I4fa1c04e196592fd1ef893a83c75261a55d32290
Added ARMtoMips64Assembler class that translates ARM instructions to MIPS64r6
instructions.
Added MIPS64r6 assembly implementations for col32cb16_blend and t32cb16_blend
functions.
Added MIPS32r2 assembly implementation for col32cb16_blend function.
Added tests for MIPS64r6 (assembler, disassembler and assembly implementation
functions).
Added MIPS32 tests for assembly implementation functions.
Minor bug fixes for MIPS32r2 branch.
Change-Id: I69e49622117be5b8167628e9702db6aafb1849d7
See the comment-block at the top of Aarch64Assembler.cpp
for overview on how AArch64 support has been implemented
In addition, this commit contains
[x] AArch64 inline asm versions of gglmul series of
functions and a new unit test bench to test the
functions
[x] Assembly implementations of scanline_col32cb16blend
and scanline_t32cb16blend for AArch64, with unit
test bench
Change-Id: I915cded9e1d39d9a2a70bf8a0394b8a0064d1eb4
Signed-off-by: Ashok Bhat <ashok.bhat@arm.com>
GGLAssembler assumes addresses to be 32-bit and uses ARM 32-bit
instructions to load/store/manipulate addresses. To support, 64-bit
architectures, following changes has been done
1. ARMAssemblerInterface has been extended to support four new
operations ADDR_LDR, ADDR_STR, ADDR_SUB, ADDR_ADD. Base class
implements these virtual functions to use 32bit equivalent
function. This avoids existing 32-bit Assembler backend
implementations like ARMAssembler and MIPSAssembler from
mapping the new functions to existing equivalent routines.
This also allows 64-bit Architectures like AArch64 to override
the function in their assembler backend implementations.
2. GGLAssembler code (spread over GGLAssembler.cpp, GGLAssembler.h
and texturing.cpp) has been changed to use the new operations
for address operations.
Change-Id: I3d7eace4691e3e47cef737d97ac67ce6ef4fb18d
Signed-off-by: Ashok Bhat <ashok.bhat@arm.com>
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%
* 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.
Elliott Hughes