So, my earlier post on this was a little premature; anyone who has tried out the code has found out that it pretty much doesn’t work (hey I did warn you!). Now there are a range of fun reasons why this didn’t work, most of which I’ve now solved.
Firstly, it turns out that EABI and ARMv4T are pretty much
incompatible. (I’ll post separately about that!). In short,
thumb interworking doesn’t (can’t) work, so I’ve reverted back to
plain old ARMv4 architecture as my target (the only difference
between ARMv4 and ARMv4T is the thumb stuff, which we can’t use
until the compiler / spec is fixed.). So I’ve updated the
linux-arm.mk to support ARMv4 for now as well.
Of course the next problem that this introduces is that the
bx instruction doesn’t exist on ARMv4, and GCC
(helpfully) complains and stops the compilation. Now a BX
without thumb support is simply a mov pc, instruction,
so I went through and provided a BX macro that expands
to either bx or mov pc,. This is a little
bit nasty/invasive because it touches all the system call bindings,
thankfully these are generated anyway, but it makes the diff quite
large. (When I have time I’ll make it so that generation is part of
the buid system, not a manual process.)
The next problem is that the provided compiler’s
libgcc library is build for ARMv5, and has instructions
that just don’t exist on ARMv4 (shc as clz), so I went and built a new compiler
targeted to ARMv4. There is no reason why this couldn’t be set up as a
multi-lib compiler that supports both, but I don’t have enough GCC
wizardry in me to work that out right now. So a new compiler.
This got things to a booting stage, but not able to mount
/system or /data. Basically, Android by
default uses yet another flash
file-system (YAFFS), but for some reasons, which I couldn’t
fully work out initially, the filesystem just didn’t seem to cleanly
initialise and then mount. So, without diving too deep, I figured I
could just use jffs2 instead, which I know works on the
target. So I upgraded the Android build system to support allowing you
to choose which filesystem type to use, and providing jffs2 as an
option. This was going much better, and I got a lot further, far
enough that I needed to recompile my kernel with support for some of
the Android specific drivers like ashmem, binder and
logger. Unfortunately I was getting a hang on an mmap
call, for reasons that I couldn’t quite work out. After a lot of
tedious debugging (my serial console is broken, so I have to rely on
graphics console, which is really just an insane way to try and debug
anything), anyway, it turns out that part of what the Dalvik virtual
machine does when optimising class files is to mmap the
file as writable memory. This was what was failing, with the totally
useless error invalid argument. Do you know how many unique
paths along the mmap system call can set EINVAL? Well it’s a lot. Anyway,
long story short, it turns out that the jffs2 filesystem doesn’t support writable
mmaps! %&!#.
After I finished cursing, I decided to go back to using yaffs and
working out what the real problem is. After upgrading u-boot (in a
pointless attempt to fix my serial console), I noticed a new
write yaffs[1] command. This wasn’t there in the old
version. Ok, cool, maybe this has something do to with the
problem. But what is this the deal with yaffs versus yaffs1? Well it
turns out that NAND has different pagesize, 512 bytes, and 2k (or
multiples thereof, maybe??). And it turns out that YAFFS takes
advantage of this and has different file systems for different sized
NAND pages, and of course, everything that can go wrong will so, the
filesystem image that the build system creates is YAFFS2 which is for
2k pages not 512b pages. So, I again updated the build system to
firstly build both the mkyaffs2image and the
mkyaffsimage tool, and then set off building a YAFFS file
system.
Now, while u-boot supports yaffs filesystem, device firmware update
doesn’t (appear to). So this means I need to copy the image to memory
first, then on the device copy it from memory to flash. Now, the other
fun thing is that dfu can only copy 2MB or so to RAM at a time, and
the system.img file is around 52MB or so, which means that
it takes around 26 individual copies of 2MB sections.... very, very painful.
But in the end this more or less worked. So now I have a 56MB partition
for the system, and a 4MB partition for the user and things are looking good.
Good that is, right up until the point where dalvik starts up and
writes out cached version of class files to /data. You
see, it needs more than 4MB, a lot more, so I’m kind of back to square
one. I mean, if I’d looked at the requirements I would have read 128MB
of flash, but meh, who reads requirements? The obvious option would be
some type of MMC card, but as it turns out the number of handy Fry’s
stores on Boeing 747 from Sydney to LA number in the zeroes.
So the /system partition is read-only, and since the
only problem with jffs2 was when we were writing to it, it seems that
we could use jffs2 for the read-only system partition, which has the
advantage of jffs2 doing compression, and fitting in about 30MB, not
about 50MB, leaving plenty of room for the user data partition, which
is where the Dalvik cached files belong. This also has the advantage
of being able to use normal DFU commands to install the image
(yay!). So after more updates to the build system to now support
individually setting the system filesystem type and the user
filesystem type things seem a lot happier.
Currently, I have a system that boots init, starts up
most of the system services, including the Dalvik VM, runs a bunch of
code, but bombs out with an out-of-memory error in the pixelflinger
code which I’m yet to have any luck tracing. Currently my serial
console is fubar, so I can’t get any useful logging, which makes
things doubly painful. The next step is to get adb
working over USB so I have at least an output of the errors and
warning, which should give me half a chance of tracking down the
problem.
So if you want to try and get up to this point, what are the steps?
Well, firstly go and download the android
toolchain source code. and compile it for a v4 target. You use the
--target=armv4-android-eabi argument to configure if I
remember correctly.
Once you have that done, grab my latest patch and apply it to the Android source code base. (That is tar file with diffs for each individual project, apply these correctly is left as an exercise for the reader). Then you want to compile it with the new toolchain. I use a script like this:
#!/bin/sh
make TARGET_ARCH_VERSION=armv4 \
MKJFFS2_CMD="ssh nirvana -x \"cd `pwd`; mkfs.jffs2\"" \
SYSTEM_FSTYPE=jffs2 \
USERDATA_FSTYPE=yaffs \
TARGET_TOOLS_PREFIX=/opt/benno/bin/armv4-android-eabi- $@
Things you will need to change it the tools prefix, and the mkjffs2 command. The evil-hackery above is to run it on my linux virtual machine (I’m compiling the rest under OS X, and I can’t get mkfs.jffs2 to compile under it yet.)
After some time passes you should end up with a ramdisk.img, userdata.img and system.img files. The next step is to get a usable kernel.
I’m using the OpenMoko stable kernel, which is 2.6.24 based. I’ve patched this with bits of the Android kernel (enough, I think, to make it run). Make sure you configure support for yaffs, binder, logger and ashmem. Here is the kernel config I’m currently using.
At this stage it is important you have a version of u-boot supporting the yaffs write commands, if you don’t your next step is to install that. After this the next step is to re-partition your flash device. In case it isn’t obvious this will trash your current OS. The useful parts from my uboot environment are:
mtdids=nand0=neo1973-nand
bootdelay=-1
mtdparts=mtdparts=neo1973-nand:256k(uboot)ro,16k(uboot-env),752k(ramdisk),2m(kernel),36m(system),24m(userdata)
rdaddr=0x35000000
kaddr=0x32000000
bootcmd=setenv bootargs ${bootargs_base} ${mtdparts} initrd=${rdaddr},${rdsize}; nand read.e ${kaddr} kernel; nand read.e ${rdaddr} ramdisk; bootm ${kaddr}
bootargs_base=root=/dev/ram rw console=tty0 loglevel=8
Note the mtdparts which defines the partitions, and the bootcmd. (I’m not entirely happy with the boot command, mostly because when I install new RAM image I need to manually update $rdsize, which is a pain).
With this in place you are ready to start. The first image to move across is your userdata image. Now to make this happen we first copy it into memory using dfu-util:
sudo dfu-util -a 0 -R -D source/out/target/product/generic/userdata.img -R
Then you need to use the nand write.yaffs1 command to copy it to the data partition. Note, at this stage I get weird behaviour, I’m not convinced that the yaffs support truly works yet! Afterwards I get some messed up data in other parts of the flash (which is why we are doing it first). After you have copied it in, I suggest reseting the device, and you may find you need to reinitialise u-boot (using dyngen, and resetting up the environment as above.
After this you are good to use dfu-util to copy accross the kernel, system.img and ramdisk.img. After copying the ramdisk.img across update the rdsize variable with the size of the ramdisk.
Once all this is done, you are good to boot, I wish you luck! If you have a working
serial console you can probably try the logcat command to see why
graphics aren’t working. If you get this far please email me the results!
After a lot of stuffing around installing new hard drives so I had enough space to actually play with the source code, getting screwed by Time Machine when trying to convert my filesystem from case-insenstive to case-insensitive (I gave up and am now usuing a case-sensitive disk image on top of my case-insenstive file system.. sigh), I finally have the Android source code compiling, yay!.
Compiling is fairly trivial, just make and away it
goes. The fun thing is trying to work out exactly what the hell the
build system is actually doing. I’ve got to admit though, it is a
pretty clean build system, although it isn’t going to win any speed
records. I’m going to go into more details on the build sstem when i
have more time, and I’ve actually worked out what the hell is
happening.
Anyway, after a few false starts I now have the build system compiling for ARMv4T processors (such as the one inside the Neo1973), and hopefully at the same time I haven’t broken compilation from ARMv5TE.
For those interested I have a patch
available. Simply apply this to the checked out code, and the build
using make TARGET_ARCH_VERSION=armv4t. Now, of course I haven’t
actually tried to run this code yet, so it might not work, but
it seems to compile fine, so that is a good start! Now once I work out how to make
git play nice I'll actually put this into a branch and make it available, but
the diff will have to suffice for now. Of course I’m not the only one looking
at this, check out Christopher’s
page for more information. (Where he actually starts solving some problems instead of just
working around them ;)
The rest of this post documents the patch. For those interested it should give you some idea of the build system and layout, and hopefully it is something that can be applied to mainline.
The first changes made are to the linux-arm.mk file. A new make variable
TARGET_ARCH_VERSION is added. For now this is defaulted to armv5te,
but it can be overridden on the command line as shown above.
project build/ diff --git a/core/combo/linux-arm.mk b/core/combo/linux-arm.mk index adb82d3..a43368f 100644 --- a/core/combo/linux-arm.mk +++ b/core/combo/linux-arm.mk @@ -7,6 +7,8 @@ $(combo_target)TOOLS_PREFIX := \ prebuilt/$(HOST_PREBUILT_TAG)/toolchain/arm-eabi-4.2.1/bin/arm-eabi- endif +TARGET_ARCH_VERSION ?= armv5te + $(combo_target)CC := $($(combo_target)TOOLS_PREFIX)gcc$(HOST_EXECUTABLE_SUFFIX) $(combo_target)CXX := $($(combo_target)TOOLS_PREFIX)g++$(HOST_EXECUTABLE_SUFFIX) $(combo_target)AR := $($(combo_target)TOOLS_PREFIX)ar$(HOST_EXECUTABLE_SUFFIX)
The next thing is to make the GLOBAL_CFLAGS variable dependent on the architecture
version. The armv5te defines stay in place, but an armv4t architecture version is
added. Most of the cflags are pretty similar, except we change the -march
flag, and change the pre-processor defines. These will become important later in the
patch as they provide the mechanism for distinguishing between versions in the code.
@@ -46,6 +48,7 @@ ifneq ($(wildcard $($(combo_target)CC)),) $(combo_target)LIBGCC := $(shell $($(combo_target)CC) -mthumb-interwork -print-libgcc-file-name) endif +ifeq ($(TARGET_ARCH_VERSION), armv5te) $(combo_target)GLOBAL_CFLAGS += \ -march=armv5te -mtune=xscale \ -msoft-float -fpic \ @@ -56,6 +59,21 @@ $(combo_target)GLOBAL_CFLAGS += \ -D__ARM_ARCH_5__ -D__ARM_ARCH_5T__ \ -D__ARM_ARCH_5E__ -D__ARM_ARCH_5TE__ \ -include $(call select-android-config-h,linux-arm) +else +ifeq ($(TARGET_ARCH_VERSION), armv4t) +$(combo_target)GLOBAL_CFLAGS += \ + -march=armv4t \ + -msoft-float -fpic \ + -mthumb-interwork \ + -ffunction-sections \ + -funwind-tables \ + -fstack-protector \ + -D__ARM_ARCH_4__ -D__ARM_ARCH_4T__ \ + -include $(call select-android-config-h,linux-arm) +else +$(error Unknown TARGET_ARCH_VERSION=$(TARGET_ARCH_VERSION)) +endif +endif $(combo_target)GLOBAL_CPPFLAGS += -fvisibility-inlines-hidden
The next bit we update is the prelink-linux-arm.map file.
The dynamic libraries in android are laid out explicitly in virtual memory
according to this map file. If I’m not mistaken those address look suspiciously
1MB aligned, which means they should fit nicely in the pagetable, and provides
some opportunity to use fast-address-space-switching techniques. In the port
to ARMv4 I have so far been lazy and instead of fixing up any assembler code
I’ve just gone with existing C code. One outcome of this is that I need the
libffi.so for my foreign function interface, so I’ve added this to the map
for now. I’m not 100% sure that when compiling for ARMv5 this won’t cause
a problem. Will need to see. Fixing up the code to avoid needing libffi
is probably high on the list of things to do.
diff --git a/core/prelink-linux-arm.map b/core/prelink-linux-arm.map index d4ebf43..6e0bc43 100644 --- a/core/prelink-linux-arm.map +++ b/core/prelink-linux-arm.map @@ -113,3 +113,4 @@ libctest.so 0x9A700000 libUAPI_jni.so 0x9A500000 librpc.so 0x9A400000 libtrace_test.so 0x9A300000 +libffi.so 0x9A200000
The next module is the bionic module which is the light-weight C library that is part of Android. This has some nice optimised routines for memory copy and compare, but unfortunately they rely on ARMv5 instructions. I’ve changed the build system to only use the optimised assembler when compiling with ARMv5TE, and falling back to C routines in the other cases. (The strlen implementation isn’t pure assembly, but the optimised C implementation has inline asm, so again it needs to drop back to plain old dumb strlen.)
project bionic/ diff --git a/libc/Android.mk b/libc/Android.mk index faca333..3fb3455 100644 --- a/libc/Android.mk +++ b/libc/Android.mk @@ -206,13 +206,9 @@ libc_common_src_files := \ arch-arm/bionic/_setjmp.S \ arch-arm/bionic/atomics_arm.S \ arch-arm/bionic/clone.S \ - arch-arm/bionic/memcmp.S \ - arch-arm/bionic/memcmp16.S \ - arch-arm/bionic/memcpy.S \ arch-arm/bionic/memset.S \ arch-arm/bionic/setjmp.S \ arch-arm/bionic/sigsetjmp.S \ - arch-arm/bionic/strlen.c.arm \ arch-arm/bionic/syscall.S \ arch-arm/bionic/kill.S \ arch-arm/bionic/tkill.S \ @@ -274,6 +270,18 @@ libc_common_src_files := \ netbsd/nameser/ns_print.c \ netbsd/nameser/ns_samedomain.c + +ifeq ($(TARGET_ARCH),arm) +ifeq ($(TARGET_ARCH_VERSION),armv5te) +libc_common_src_files += arch-arm/bionic/memcmp.S \ + arch-arm/bionic/memcmp16.S \ + arch-arm/bionic/memcpy.S \ + arch-arm/bionic/strlen.c.arm +else +libc_common_src_files += string/memcmp.c string/memcpy.c string/strlen.c string/ffs.c +endif +endif + # These files need to be arm so that gdbserver # can set breakpoints in them without messing # up any thumb code.
Unfortunately, it is clear that this C only code hasn’t been used in a while as there was a trivial bug as fixed by the patch below. This makes me worry about what other bugs that aren’t caught by the compiler may be lurking.
diff --git a/libc/string/memcpy.c b/libc/string/memcpy.c index 4cd4a80..dea78b2 100644 --- a/libc/string/memcpy.c +++ b/libc/string/memcpy.c @@ -25,5 +25,5 @@ * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ -#define MEM_COPY +#define MEMCOPY #include "bcopy.c"
Finally, frustratingly, the compiler’s ffs() implementation
appears to fallback to calling the C library’s ffs() implementation
if it can’t doing something optimised. This happens when compiling for ARMv4, so I’ve
added an ffs() implementation (stolen from FreeBSD).
#include#include /* * Find First Set bit */ int ffs(int mask) { int bit; if (mask == 0) return (0); for (bit = 1; !(mask & 1); bit++) mask = (unsigned int)mask >> 1; return (bit); }
The next module for attention is the dalvik virtual machine. Again this has some code that relies on ARMv5, but there is a C version that we fall back on. In this case it also means pulling in libffi. This is probably the module that needs to most attention in actually updating the code to be ARMv4 assembler in the near future.
project dalvik/ diff --git a/vm/Android.mk b/vm/Android.mk index dfed78d..c66a861 100644 --- a/vm/Android.mk +++ b/vm/Android.mk @@ -189,6 +189,7 @@ ifeq ($(TARGET_SIMULATOR),true) endif ifeq ($(TARGET_ARCH),arm) +ifeq ($(TARGET_ARCH_VERSION),armv5te) # use custom version rather than FFI #LOCAL_SRC_FILES += arch/arm/CallC.c LOCAL_SRC_FILES += arch/arm/CallOldABI.S arch/arm/CallEABI.S @@ -204,6 +205,16 @@ else mterp/out/InterpC-desktop.c \ mterp/out/InterpAsm-desktop.S LOCAL_SHARED_LIBRARIES += libffi + LOCAL_SHARED_LIBRARIES += libdl +endif +else + # use FFI + LOCAL_C_INCLUDES += external/libffi/$(TARGET_OS)-$(TARGET_ARCH) + LOCAL_SRC_FILES += arch/generic/Call.c + LOCAL_SRC_FILES += \ + mterp/out/InterpC-desktop.c \ + mterp/out/InterpAsm-desktop.S + LOCAL_SHARED_LIBRARIES += libffi endif LOCAL_MODULE := libdvm
Next is libjpeg, which again, has assembler optimisation that we can’t easily use without real porting work, so we fall back to the C
project external/jpeg/ diff --git a/Android.mk b/Android.mk index 9cfe4f6..3c052cd 100644 --- a/Android.mk +++ b/Android.mk @@ -19,6 +19,12 @@ ifneq ($(TARGET_ARCH),arm) ANDROID_JPEG_NO_ASSEMBLER := true endif +# the assembler doesn't work for armv4t +ifeq ($(TARGET_ARCH_VERSION),armv4t) +ANDROID_JPEG_NO_ASSEMBLER := true +endif + + # temp fix until we understand why this broke cnn.com #ANDROID_JPEG_NO_ASSEMBLER := true
For some reason compiling with ARMv4 doesn’t allow the prefetch loop array compiler optimisation, so we turn it off for ARMv4.
@@ -29,7 +35,10 @@ LOCAL_SRC_FILES += jidctint.c jidctfst.S endif LOCAL_CFLAGS += -DAVOID_TABLES -LOCAL_CFLAGS += -O3 -fstrict-aliasing -fprefetch-loop-arrays +LOCAL_CFLAGS += -O3 -fstrict-aliasing +ifeq ($(TARGET_ARCH_VERSION),armv5te) +LOCAL_FLAGS += -fprefetch-loop-arrays +endif #LOCAL_CFLAGS += -march=armv6j LOCAL_MODULE:= libjpeg
Next up is libffi, which is just a case of turning it on since we now need it for ARMv4.
project external/libffi/ diff --git a/Android.mk b/Android.mk index f4452c9..07b5c2f 100644 --- a/Android.mk +++ b/Android.mk @@ -6,7 +6,7 @@ # We need to generate the appropriate defines and select the right set of # source files for the OS and architecture. -ifneq ($(TARGET_ARCH),arm) +ifneq ($(TARGET_ARCH_VERSION),armv5te) LOCAL_PATH:= $(call my-dir) include $(CLEAR_VARS)
The external module opencore contains a lot of software implemented codecs. (I wonder about the licensing restrictions on these things...). Not surprisingly these too are tuned for ARMv4, but again we fall back to plain old C.
project external/opencore/ diff --git a/codecs_v2/audio/aac/dec/Android.mk b/codecs_v2/audio/aac/dec/Android.mk index ffe0089..6abdc2d 100644 --- a/codecs_v2/audio/aac/dec/Android.mk +++ b/codecs_v2/audio/aac/dec/Android.mk @@ -150,7 +150,7 @@ LOCAL_SRC_FILES := \ LOCAL_MODULE := libpv_aac_dec LOCAL_CFLAGS := -DAAC_PLUS -DHQ_SBR -DPARAMETRICSTEREO $(PV_CFLAGS) -ifeq ($(TARGET_ARCH),arm) +ifeq ($(TARGET_ARCH_VERSION),armv5te) LOCAL_CFLAGS += -D_ARM_GCC else LOCAL_CFLAGS += -DC_EQUIVALENT diff --git a/codecs_v2/audio/gsm_amr/amr_wb/dec/Android.mk b/codecs_v2/audio/gsm_amr/amr_wb/dec/Android.mk index e184178..3223841 100644 --- a/codecs_v2/audio/gsm_amr/amr_wb/dec/Android.mk +++ b/codecs_v2/audio/gsm_amr/amr_wb/dec/Android.mk @@ -48,7 +48,7 @@ LOCAL_SRC_FILES := \ LOCAL_MODULE := libpvamrwbdecoder LOCAL_CFLAGS := $(PV_CFLAGS) -ifeq ($(TARGET_ARCH),arm) +ifeq ($(TARGET_ARCH_VERSION),armv5te) LOCAL_CFLAGS += -D_ARM_GCC else LOCAL_CFLAGS += -DC_EQUIVALENT diff --git a/codecs_v2/audio/mp3/dec/Android.mk b/codecs_v2/audio/mp3/dec/Android.mk index 254cb6b..c2430fe 100644 --- a/codecs_v2/audio/mp3/dec/Android.mk +++ b/codecs_v2/audio/mp3/dec/Android.mk @@ -28,8 +28,8 @@ LOCAL_SRC_FILES := \ src/pvmp3_seek_synch.cpp \ src/pvmp3_stereo_proc.cpp \ src/pvmp3_reorder.cpp - -ifeq ($(TARGET_ARCH),arm) + +ifeq ($(TARGET_ARCH_VERSION),armv5te) LOCAL_SRC_FILES += \ src/asm/pvmp3_polyphase_filter_window_gcc.s \ src/asm/pvmp3_mdct_18_gcc.s \ @@ -46,7 +46,7 @@ endif LOCAL_MODULE := libpvmp3 LOCAL_CFLAGS := $(PV_CFLAGS) -ifeq ($(TARGET_ARCH),arm) +ifeq ($(TARGET_ARCH_VERSION),armv5te) LOCAL_CFLAGS += -DPV_ARM_GCC else LOCAL_CFLAGS += -DC_EQUIVALENT
Unfortunately it is not just the build file that needs updating in this module. I need to manually go and update the headers so that some optimised inline assembler is only used in the ARMv5 case. To be honest this messes these files up a little bit, so a nicer solution would be preferred.
diff --git a/codecs_v2/video/m4v_h263/enc/src/dct_inline.h b/codecs_v2/video/m4v_h263/enc/src/dct_inline.h
index 86474b2..41a3297 100644
--- a/codecs_v2/video/m4v_h263/enc/src/dct_inline.h
+++ b/codecs_v2/video/m4v_h263/enc/src/dct_inline.h
@@ -22,7 +22,7 @@
#ifndef _DCT_INLINE_H_
#define _DCT_INLINE_H_
-#if !defined(PV_ARM_GCC)&& defined(__arm__)
+#if !(defined(PV_ARM_GCC) && defined(__arm__) && defined(__ARCH_ARM_5TE__))
#include "oscl_base_macros.h"
@@ -109,7 +109,7 @@ __inline int32 sum_abs(int32 k0, int32 k1, int32 k2, int32 k3,
#elif defined(__CC_ARM) /* only work with arm v5 */
#if defined(__TARGET_ARCH_5TE)
-
+#error
__inline int32 mla724(int32 op1, int32 op2, int32 op3)
{
int32 out;
@@ -266,7 +266,7 @@ __inline int32 sum_abs(int32 k0, int32 k1, int32 k2, int32 k3,
return abs_sum;
}
-#elif defined(PV_ARM_GCC) && defined(__arm__) /* ARM GNU COMPILER */
+#elif defined(PV_ARM_GCC) && defined(__arm__) && defined(__ARCH_ARM_5TE__) /* ARM GNU COMPILER */
__inline int32 mla724(int32 op1, int32 op2, int32 op3)
{
diff --git a/codecs_v2/video/m4v_h263/enc/src/fastquant_inline.h b/codecs_v2/video/m4v_h263/enc/src/fastquant_inline.h
index 6a35d43..fbfeddf 100644
--- a/codecs_v2/video/m4v_h263/enc/src/fastquant_inline.h
+++ b/codecs_v2/video/m4v_h263/enc/src/fastquant_inline.h
@@ -25,7 +25,7 @@
#include "mp4def.h"
#include "oscl_base_macros.h"
-#if !defined(PV_ARM_GCC) && defined(__arm__) /* ARM GNU COMPILER */
+#if !(defined(PV_ARM_GCC) && defined(__arm__) && defined(__ARCH_ARM_V5TE__)) /* ARM GNU COMPILER */
__inline int32 aan_scale(int32 q_value, int32 coeff, int32 round, int32 QPdiv2)
{
@@ -423,7 +423,7 @@ __inline int32 coeff_dequant_mpeg_intra(int32 q_value, int32 tmp)
return q_value;
}
-#elif defined(PV_ARM_GCC) && defined(__arm__) /* ARM GNU COMPILER */
+#elif defined(PV_ARM_GCC) && defined(__arm__) && defined(__ARCH_ARM_V5TE__) /* ARM GNU COMPILER */
__inline int32 aan_scale(int32 q_value, int32 coeff,
int32 round, int32 QPdiv2)
diff --git a/codecs_v2/video/m4v_h263/enc/src/vlc_encode_inline.h b/codecs_v2/video/m4v_h263/enc/src/vlc_encode_inline.h
index 69857f3..b0bf46d 100644
--- a/codecs_v2/video/m4v_h263/enc/src/vlc_encode_inline.h
+++ b/codecs_v2/video/m4v_h263/enc/src/vlc_encode_inline.h
@@ -18,7 +18,7 @@
#ifndef _VLC_ENCODE_INLINE_H_
#define _VLC_ENCODE_INLINE_H_
-#if !defined(PV_ARM_GCC)&& defined(__arm__)
+#if !(defined(PV_ARM_GCC) && defined(__arm__) && defined(__ARCH_ARM_V5TE__))
__inline Int zero_run_search(UInt *bitmapzz, Short *dataBlock, RunLevelBlock *RLB, Int nc)
{
@@ -208,7 +208,7 @@ __inline Int zero_run_search(UInt *bitmapzz, Short *dataBlock, RunLevelBlock *R
return idx;
}
-#elif defined(PV_ARM_GCC) && defined(__arm__) /* ARM GNU COMPILER */
+#elif defined(PV_ARM_GCC) && defined(__arm__) && defined(__ARCH_ARM_V5TE__) /* ARM GNU COMPILER */
__inline Int m4v_enc_clz(UInt temp)
{
A similar approach is needed in the skia graphics library.
project external/skia/
diff --git a/include/corecg/SkMath.h b/include/corecg/SkMath.h
index 76cf279..5f0264f 100644
--- a/include/corecg/SkMath.h
+++ b/include/corecg/SkMath.h
@@ -162,7 +162,7 @@ static inline int SkNextLog2(uint32_t value) {
With this requirement, we can generate faster instructions on some
architectures.
*/
-#if defined(__arm__) && !defined(__thumb__)
+#if defined(__arm__) && defined(__ARM_ARCH_5TE__) && !defined(__thumb__)
static inline int32_t SkMulS16(S16CPU x, S16CPU y) {
SkASSERT((int16_t)x == x);
SkASSERT((int16_t)y == y);
The sonivox module (no idea what that is!), has the same requirement of updating the build to avoid building ARMv5 specific code.
project external/sonivox/ diff --git a/arm-wt-22k/Android.mk b/arm-wt-22k/Android.mk index 565c233..a59f917 100644 --- a/arm-wt-22k/Android.mk +++ b/arm-wt-22k/Android.mk @@ -73,6 +73,7 @@ LOCAL_COPY_HEADERS := \ host_src/eas_reverb.h ifeq ($(TARGET_ARCH),arm) +ifeq (($TARGET_ARCH),armv5) LOCAL_SRC_FILES+= \ lib_src/ARM-E_filter_gnu.s \ lib_src/ARM-E_interpolate_loop_gnu.s \
The low-level audio code in audioflinger suffers from the same optimisations, and we need to dive into the code on this occasion to fix things up.
project frameworks/base/
diff --git a/libs/audioflinger/AudioMixer.cpp b/libs/audioflinger/AudioMixer.cpp
index 9f1b17f..4c0890c 100644
--- a/libs/audioflinger/AudioMixer.cpp
+++ b/libs/audioflinger/AudioMixer.cpp
@@ -400,7 +400,7 @@ void AudioMixer::process__validate(state_t* state, void* output)
static inline
int32_t mulAdd(int16_t in, int16_t v, int32_t a)
{
-#if defined(__arm__) && !defined(__thumb__)
+#if defined(__arm__) && defined(__ARCH_ARM_5TE__) && !defined(__thumb__)
int32_t out;
asm( "smlabb %[out], %[in], %[v], %[a] \n"
: [out]"=r"(out)
@@ -415,7 +415,7 @@ int32_t mulAdd(int16_t in, int16_t v, int32_t a)
static inline
int32_t mul(int16_t in, int16_t v)
{
-#if defined(__arm__) && !defined(__thumb__)
+#if defined(__arm__) && defined(__ARCH_ARM_5TE__) && !defined(__thumb__)
int32_t out;
asm( "smulbb %[out], %[in], %[v] \n"
: [out]"=r"(out)
@@ -430,7 +430,7 @@ int32_t mul(int16_t in, int16_t v)
static inline
int32_t mulAddRL(int left, uint32_t inRL, uint32_t vRL, int32_t a)
{
-#if defined(__arm__) && !defined(__thumb__)
+#if defined(__arm__) && defined(__ARCH_ARM_5TE__) && !defined(__thumb__)
int32_t out;
if (left) {
asm( "smlabb %[out], %[inRL], %[vRL], %[a] \n"
@@ -456,7 +456,7 @@ int32_t mulAddRL(int left, uint32_t inRL, uint32_t vRL, int32_t a)
static inline
int32_t mulRL(int left, uint32_t inRL, uint32_t vRL)
{
-#if defined(__arm__) && !defined(__thumb__)
+#if defined(__arm__) && defined(__ARCH_ARM_5TE__) && !defined(__thumb__)
int32_t out;
if (left) {
asm( "smulbb %[out], %[inRL], %[vRL] \n"
diff --git a/libs/audioflinger/AudioResamplerSinc.cpp b/libs/audioflinger/AudioResamplerSinc.cpp
index e710d16..88b8c22 100644
--- a/libs/audioflinger/AudioResamplerSinc.cpp
+++ b/libs/audioflinger/AudioResamplerSinc.cpp
@@ -62,7 +62,7 @@ const int32_t AudioResamplerSinc::mFirCoefsDown[] = {
static inline
int32_t mulRL(int left, int32_t in, uint32_t vRL)
{
-#if defined(__arm__) && !defined(__thumb__)
+#if defined(__arm__) && defined(__ARCH_ARM_5TE__) && !defined(__thumb__)
int32_t out;
if (left) {
asm( "smultb %[out], %[in], %[vRL] \n"
@@ -88,7 +88,7 @@ int32_t mulRL(int left, int32_t in, uint32_t vRL)
static inline
int32_t mulAdd(int16_t in, int32_t v, int32_t a)
{
-#if defined(__arm__) && !defined(__thumb__)
+#if defined(__arm__) && defined(__ARCH_ARM_5TE__) && !defined(__thumb__)
int32_t out;
asm( "smlawb %[out], %[v], %[in], %[a] \n"
: [out]"=r"(out)
@@ -103,7 +103,7 @@ int32_t mulAdd(int16_t in, int32_t v, int32_t a)
static inline
int32_t mulAddRL(int left, uint32_t inRL, int32_t v, int32_t a)
{
-#if defined(__arm__) && !defined(__thumb__)
+#if defined(__arm__) && defined(__ARCH_ARM_5TE__) && !defined(__thumb__)
int32_t out;
if (left) {
asm( "smlawb %[out], %[v], %[inRL], %[a] \n"
The AndroidConfig.h header file is included on every compile. We mess with it to convince it that we don’t have an optimised memcmp16 function.
project system/core/ diff --git a/include/arch/linux-arm/AndroidConfig.h b/include/arch/linux-arm/AndroidConfig.h index d7e182a..76f424e 100644 --- a/include/arch/linux-arm/AndroidConfig.h +++ b/include/arch/linux-arm/AndroidConfig.h @@ -249,8 +249,9 @@ /* * Do we have __memcmp16()? */ +#if defined(__ARCH_ARM_5TE__) #define HAVE__MEMCMP16 1 - +#endif /* * type for the third argument to mincore(). */
Next up is the pixelflinger, where things get interesting, because all of a sudden we have armv6 code. I’ve taken the rash decision of wrapping this in conditionals that are only enabled if you actually have an ARMv6 version, not a pesky ARMv5E, but I really need to better understand the intent here. It seems a little strange.
diff --git a/libpixelflinger/Android.mk b/libpixelflinger/Android.mk index a8e5ee4..077cf47 100644 --- a/libpixelflinger/Android.mk +++ b/libpixelflinger/Android.mk @@ -5,7 +5,7 @@ include $(CLEAR_VARS) # ARMv6 specific objects # -ifeq ($(TARGET_ARCH),arm) +ifeq ($(TARGET_ARCH_VERSION),armv6) LOCAL_ASFLAGS := -march=armv6 LOCAL_SRC_FILES := rotate90CW_4x4_16v6.S LOCAL_MODULE := libpixelflinger_armv6 @@ -39,7 +39,7 @@ PIXELFLINGER_SRC_FILES:= \ raster.cpp \ buffer.cpp -ifeq ($(TARGET_ARCH),arm) +ifeq ($(TARGET_ARCH_VERSION),armv5te) PIXELFLINGER_SRC_FILES += t32cb16blend.S endif @@ -67,7 +67,7 @@ ifneq ($(BUILD_TINY_ANDROID),true) LOCAL_MODULE:= libpixelflinger LOCAL_SRC_FILES := $(PIXELFLINGER_SRC_FILES) LOCAL_CFLAGS := $(PIXELFLINGER_CFLAGS) -DWITH_LIB_HARDWARE -ifeq ($(TARGET_ARCH),arm) +ifeq ($(TARGET_ARCH_VERSION),armv6) LOCAL_WHOLE_STATIC_LIBRARIES := libpixelflinger_armv6 endif include $(BUILD_SHARED_LIBRARY)
Finally scanline has an optimised asm version it calls in preference to doing the same thing inline with C code. Again, I take the easy way out, and use the C code.
diff --git a/libpixelflinger/scanline.cpp b/libpixelflinger/scanline.cpp
index d24c988..685a3b7 100644
--- a/libpixelflinger/scanline.cpp
+++ b/libpixelflinger/scanline.cpp
@@ -1312,7 +1312,7 @@ void scanline_t32cb16blend(context_t* c)
const int32_t v = (c->state.texture[0].shade.it0>>16) + y;
uint32_t *src = reinterpret_cast(tex->data)+(u+(tex->stride*v));
-#if ((ANDROID_CODEGEN >= ANDROID_CODEGEN_ASM) && defined(__arm__))
+#if ((ANDROID_CODEGEN >= ANDROID_CODEGEN_ASM) && defined(__arm__) && defined(__ARCH_ARM_5TE__))
scanline_t32cb16blend_arm(dst, src, ct);
#else
while (ct--) {
And that my friends, is that! Now to see if I can actually run this code!
Usually when companies say release 4th quarter 2008
you
usually see something around January 2009, and to be honest that was
when I was expecting the Android source code to finally drop. So
I was a little surprised to see
that the code was released
early this morning.
Stay tuned, more to come as I start playing.
It’s been almost two years now since I help organise and run linux.conf.au 2007, and I thought it was time to jump back into the fray again. This time I’ll not be doing something as silly as trying to organise the whole conference, but I will be running a small miniconf on the first two days of the conference. So I’d like to invite you all to the Open Mobile Miniconf in January next year. And if you think you’ve got something cool you’d like to share with other developers, please take a look at the call for presentations, and drop me a line.
So last week I posted on the difference between trusted and trustworthy code. Picking up that thread again, if there is some code in my system that is trusted how can I tell if it is actually trustworthy?
Now ruling out truly malicious code, our assessment of the trustworthiness of code, really comes down to an assessment of the quality of the code, and a pretty reasonable proxy for code quality is the number of errors in the code, or more specifically, the lack thereof. So the question is how can we determine whether a piece of code has a small number of bugs?
The number of errors in a body of code is going to be the product
of two important factors: the defect density and the size of the
code. So, in general the defect density is measured in bugs per
thousand lines of code (KLOC), and the size of the code is measured in
lines of code. Now there are plenty of arguments on what the
”right“ method is for measuring lines of code, and in general
you can only know the exact defect density after all the bugs are found,
and of course, program testing can be used to show the presence of
bugs, but never to show their absence!
*. So, to lower the number of
bugs that exist in a code base there are really two options: reduce
the number of lines of code, or improve the defect density.
So, how do we improve, that is reduce, the defect density. Well, there are a number of pretty well known ways. Effective testing, despite its caveats, goes a long way to reducing the number of bugs in a code base (assuming that you actually fix the bugs you find!). Static analysis (in its various forms), is also an important tool to help increase the quality of code, and is often a great complement to testing as it can expose bugs in code that is impractical to test. And of course there are other formal methods like model checking which can help eliminate bugs from the design phase. A great example of this is the SPIN model checker. Code reviews, while time intensive, are also a great way of finding bugs that would otherwise fly under the radar. Another way to improve code quality is to write simple, rather than complex code. McCabe’s cyclomatic complexity measure can be one good indicator of this. There is, of course, just a sampling of some of the aspects of software quality. Wikipedia and McConnell’s Code Complete for more information.
Now, how do you know if some code base has actually undergone testing, how do you know if static analysis has been performed on the source code? How do you know if the code has under gone a thorough code review? Well, generally there are two ways, you trust the developer of that code, or you get someone you do trust to do the quality analysis (which may be yourself). Now, is the point where things quickly shift from technological solutions into the fuzzy world of economics, social science, psychology and legal theory as we try to determine the trustworthiness of another entity, be it person or corporation. The one technologically relevant part is that it is much more difficult to do a 3rd party analysis of code quality without access to the source code. Note: this I am not saying that open source software is more trustworthy, simply that making the source available enables 3rd party assessments of code quality, which may make it easier for some people to trust the code.
So, improving code quality, and thereby reducing defect density, is one side of the equation, but even if you have a very low defect density, for example, less than 1/KLOC, you can still have a large number of bugs if your code base is large. So it is very important to reduce the size of the code base as well. A small code base doesn’t just have the direct benefit of reducing the code size part of the equation, it also helps improve the defect density part of the equation. Why? Well, almost all the techniques mentioned above are more effective or tractable on a small code base. You can usually get much better code coverage, and even a reasonable amount of path coverage, with a smaller code base. Code reviews can be more comprehensive. Now to some extent those techniques can work for large code bases, just through more programmers at it, but using static analysis is another matter. Many of the algorithms and techniques involved with static analysis have polynomial, or even exponential computational complexity with n based on number of lines of code. So an analysis that may take an hour on a 10,000 line code base, could end up taking all week to run on a code base of 100,000 lines of code.
Of course, this doesn’t address the problem of how you assure that the code you think is in your TCB is really what you think it is. That topic really gets us into trusted boot, trusted protection module, code signing and so on, which I’m not going to try and address in this post.
Now, it should be very clear that if you want to be able to trust your trusted computing base, then it is going to need to be both small and high quality.
If you’ve seen me give a presentation recently, or just been talking about some of the stuff I’ve been doing recently, you’ve probably heard me mention the term trusted computing base or TCB. (Not to be confused with thread control blocks, the other TCB in operating systems). So what is the trusted computing base?
The TCB for a given system is all the components, both hardware and software, that we must be relied upon to operate correctly if the security of the system is to be maintained. In other words, an error that occurs in the TCB can affect the overall system security, while an error outside the TCB can not affect the overall system security.
Now, the TCB depends on the scope of the system and the defined security policy. For example, if we are talking about a UNIX operating system, and its applications, then the trusted computing base contains at least the operating system kernel, and probably any system daemons and setuid programs. As the kernel enforces the security mechanism of process boundaries, it should be obvious that an error in the kernel can affect the overall system security. Traditionally on UNIX, there is a user, root, who is all powerful, and can change the system security policies, so an error in any piece of software that runs with root privileges also forms part of the trusted computing base. Of course, any applications are outside the trusted computing base. An error in a database server should not affect the overall system security.
Of course, if we are using a UNIX operating system as the foundation of a database server, then the definition of the TCB changes. In this case not only is the operating system part of the TCB, but the database server is as well. This is because the database server is enforcing the security of which users can access which rows, tables and columns in the database, so an error in the database server can clearly impact the security of the system.
OK, so we now know we have to trust all the code that falls inside the TCB if we want to put any trust into our security system. The problem is, just because we have to trust this code does not give us any rational reason to believe that we can trust this code. Just because code is trusted doesn’t give us any indication at all as to whether the code is, in fact, trustworthy.
To put any faith in the security of the system we should ensure that any trusted code is trustworthy code.
There are a number of things that we can do to increase our confidence in the trustworthiness of our code, which I will explore in coming posts. For more information on the trusted computing base, the Wikipedia page gives a good overview, and links to some useful papers.
I recently spoke at the Open Mobile Exchange (OMX), at the O’Reilly Open Source Conference (OSCON). Now, a totally fantastic thing about OSCON, is that there is an easy way for audience members to provide feedback to speakers via the conference website. (Unfortunately, I was spending too much time prepping my slides to give useful feedback to my other speakers, which I must apologise for.) I really hope that the zookeepr guys add similar functionality so linux.conf.au can have similar feedback mechanism.
So, the great thing is that I got some great feedback from my last talk, which confirmed something I was worried about in my talks. When giving a talk, there is always a lot of background material that I feel I need to cover to adequately explain my position, but the problem is that I then have a lot less time to present the crux of my position.
So I’ve decided that I’m going to try and cover background information on my blog, so that I can refer to that in my talk, rather than going into lots of detail during the talk. This also gives all those people with laptops something useful to do during my talk.
Stay tuned!
On the 10th anniversary of the creation of Symbian Ltd, Nokia has announced that they will be acquiring Symbian Ltd with the aim of opening up the Symbian OS under an Eclipse based license. The mobile operating system market is really getting a shake up at the moment!
The Symbian Foundation has been created to build a new platform for mobile phones based on Symbian OS, S60, UIQ and MOAP(S). The foundation is expected to launch in H1 2009.
The new Symbian Foundation Platform will be made up of a common setup application suites, runtimes, UI framework, middleware, OS, tools and SDK, with Foundation members able to provide differentiated experiences on top. The platform is expected to be released to foundation members in 2009, and eventually open sourced over the following two years.
This obviously makes a huge change in the market place. It wll be interesting to see how Symbian Platform, vs. Android, vs. LiMo, vs. Windows Mobile vs. iPhone.
Of course it is all about developers, developers, developers, and it will be extremely interesting to see where developers will want to go.
I’ve, recently started using memoize.py,
as the core of my build system for a new project I’m working on. This
simplicity involved is pretty neat. Rather than manually needing to
work out the dependencies, (or having specialised tools for
determining the dependencies), with memoize.py, you
simply write the commands you need to build your project, and
memoize.py works out all the dependencies for you.
So, what’s the catch? Well, the way memoize.py works
is by using strace to
record all the system calls that a program makes during its
execution. By analyzing this list memoize.py can work out
all the files that are touched when a command is run, and then stores
this as a list of dependencies for that command. Then, the next time
you run the same command memoize.py first checks to see
if any of the dependencies have change (using either md5sum, or
timestamp), and only runs the command if any of the dependencies have
changed. So the catch of course is that this only runs on Linux (as
far as I know, you can’t get strace anywhere else, although that
doesn’t mean the same techniques couldn’t be used with a different
underlying system call tracing tool).
This technique is quite a radical difference to other tools which
determine a large dependency graph of the entire build, and then,
recursively work through this graph to fulfil unmet dependencies. As
a result this form is a lot more imperative, rather than declarative
style. Traditional tools (SCons, make, etc), provide a language which
allows you to essentially describe a dependency graph, and then the
order in which things are executed is really hidden inside the tool.
Using memoize.py is a lot different. You go through
defining the commands you want to run (in order!), and that is
basically it.
Some of the advantages of this approach are:
There are however some disadvantages:
ptrace to perform the system call tracing.memoize.py
a little so that you could simply choose not to run strace. Obviously
you can’t determine dependencies in this case, but you can at least build the
thing.As with may good tools in your programming kit, memoize.py is
available under a very liberal BSD style license, which is nice, because I’ve
been able to fix up some problems and add some extra functionality. In particular
I’ve added options to:
The patch and full file are available. These have of course been provided upstream, so with any luck, some or most of them will be merged upstream.
So, if you have a primarily Linux project, and want to try something
different to SCons, or make, I’d recommend considering memoize.py.
I recently posted my videos from linux.conf.au earlier this year. I ended up spending a lot of time in post-production with these, probably more than I spent in preparing for the talk (and coming up with all the demos for the talk was a lot of work too!).
I ending up shelling out for Final Cut Express (FCE) as I really couldn’t find anything in the free/open source arena that could really do all the effects that I wanted. My biggest shock was how bloody difficult it was to actually use! Don’t let the express part fool you, the learning curve is far from quick. I was also a bit surprised how film oriented FCE is. It is much more geared towards production of video captured on tape that will be viewed on a real screen, than towards digitally captured video destined for the web. (Or at least that was my impression).
The other surprising bit of the process was that I really couldn’t find a suitable place to host my video on the web. Most of the free video places didn’t want hour long movies, and I found the quality of the video once it was transcoded to be pretty terrible in most cases. This is probably due to the fine detail that I’m attempting to show, which probably doesn’t get treated too nicely by most encoders. In any case, I ended up hosting the video using Amazon web services, since the storage a transfer fees were a lot more attractive than slicehost (where the rest of my website is hosted).
Any way, as with most of my posts, the main point of this one was to remind future Benno how to export decent quality movies with FCE. (There are about a million different options to play with, and it took a lot of tweaking to get right). So, in summary, you want something along the lines of:
Format: QuickTime Movie
Options
-Video
-Settings
Compression Type: H.264
Motion:
Frame Rate: Current
Key Frames: Automatic
Frame Reording: x
Data Rate:
Data Rate: Automatic
Compressor:
Quality: Best
Encoding: Best
-Filter: None
-Size:
640x480
Preserve: using letterbox
Deinterlace
-Sound
Linear PCM
Stereo L R
Rate: 48khz
Render Settings: Quality: Normal Linear PCM Seetings: Sample Size: 16 Litte Endian: x
-Prepare for stream --- nope
To get Ogg Theora output, using the XiphQT tools.
One of the best/worst things about doing your own post production is that you become very familiar with your own annoying habits and tics. If you watch the video, um, I’m sure you will, um, realise, um, what I, um, mean. (Note to self: rehearse my talks more!)
By the end of the editing process I was both sick of my own voice,
and sick of anyone who says computers are fast enough
,
when you spending a good 14 hours encoding and compressing a video,
you realise that for some things, computers are still damn slow. I would
expect most encoding and compression is reasonably easily paralellisable
(if that is a real word?), so this massively multi-core revolution will
hopefully help my future video editing projects.