文件列表:

g729\uClinux-dist-2008R1.5-RC3\lib\libbfgdots\COPYING (1770, 2007-08-17)
g729\uClinux-dist-2008R1.5-RC3\lib\libbfgdots\g729\doc\DEVMKT-056A_G.729AB_Developers_Guide_for_Blackfin.doc (942080, 2007-08-17)
g729\uClinux-dist-2008R1.5-RC3\lib\libbfgdots\g729\include\g729ab_codec.h (8995, 2007-08-17)
g729\uClinux-dist-2008R1.5-RC3\lib\libbfgdots\g729\Makefile (1815, 2007-08-20)
g729\uClinux-dist-2008R1.5-RC3\lib\libbfgdots\g729\script\findglobals.pl (563, 2007-08-17)
g729\uClinux-dist-2008R1.5-RC3\lib\libbfgdots\g729\script\flat2fdpic.pl (6010, 2007-08-17)
g729\uClinux-dist-2008R1.5-RC3\lib\libbfgdots\g729\script\globals.txt (91, 2007-08-17)
g729\uClinux-dist-2008R1.5-RC3\lib\libbfgdots\g729\script\Makefile (1080, 2007-08-17)
g729\uClinux-dist-2008R1.5-RC3\lib\libbfgdots\g729\script\simgot.c (238, 2007-08-17)
g729\uClinux-dist-2008R1.5-RC3\lib\libbfgdots\g729\script\test1.c (124, 2007-08-17)
g729\uClinux-dist-2008R1.5-RC3\lib\libbfgdots\g729\script\test3.c (162, 2007-08-17)
g729\uClinux-dist-2008R1.5-RC3\lib\libbfgdots\g729\script\tgot.c (49, 2007-08-17)
g729\uClinux-dist-2008R1.5-RC3\lib\libbfgdots\g729\src.fdpic\acelp_code_a.asm (17919, 2007-08-17)
g729\uClinux-dist-2008R1.5-RC3\lib\libbfgdots\g729\src.fdpic\allinitialdnew.asm (13495, 2007-08-17)
g729\uClinux-dist-2008R1.5-RC3\lib\libbfgdots\g729\src.fdpic\allinitialenew.asm (18946, 2007-08-17)
g729\uClinux-dist-2008R1.5-RC3\lib\libbfgdots\g729\src.fdpic\coder.asm (7960, 2007-08-17)
g729\uClinux-dist-2008R1.5-RC3\lib\libbfgdots\g729\src.fdpic\cod_ld8a.asm (29046, 2007-08-17)
g729\uClinux-dist-2008R1.5-RC3\lib\libbfgdots\g729\src.fdpic\decoder.asm (9735, 2007-08-17)
g729\uClinux-dist-2008R1.5-RC3\lib\libbfgdots\g729\src.fdpic\dec_ld8a.asm (22607, 2007-08-17)
g729\uClinux-dist-2008R1.5-RC3\lib\libbfgdots\g729\src.fdpic\dec_sid.asm (11185, 2007-08-17)
g729\uClinux-dist-2008R1.5-RC3\lib\libbfgdots\g729\src.fdpic\dspfunc.asm (6050, 2007-08-17)
g729\uClinux-dist-2008R1.5-RC3\lib\libbfgdots\g729\src.fdpic\dtx.asm (38745, 2007-08-17)
g729\uClinux-dist-2008R1.5-RC3\lib\libbfgdots\g729\src.fdpic\g729comc.asm (18278, 2007-08-17)
g729\uClinux-dist-2008R1.5-RC3\lib\libbfgdots\g729\src.fdpic\g729comd.asm (4539, 2007-08-17)
g729\uClinux-dist-2008R1.5-RC3\lib\libbfgdots\g729\src.fdpic\G729_const.h (1994, 2007-08-17)
g729\uClinux-dist-2008R1.5-RC3\lib\libbfgdots\g729\src.fdpic\gainped.asm (5649, 2007-08-17)
g729\uClinux-dist-2008R1.5-RC3\lib\libbfgdots\g729\src.fdpic\globals.txt (2100, 2007-08-17)
g729\uClinux-dist-2008R1.5-RC3\lib\libbfgdots\g729\src.fdpic\l1_function.asm (38515, 2007-08-17)
g729\uClinux-dist-2008R1.5-RC3\lib\libbfgdots\g729\src.fdpic\lpc.asm (12751, 2007-08-17)
g729\uClinux-dist-2008R1.5-RC3\lib\libbfgdots\g729\src.fdpic\lpcfunc.asm (8160, 2007-08-17)
g729\uClinux-dist-2008R1.5-RC3\lib\libbfgdots\g729\src.fdpic\lspdec.asm (5682, 2007-08-17)
g729\uClinux-dist-2008R1.5-RC3\lib\libbfgdots\g729\src.fdpic\lspgetq.asm (9608, 2007-08-17)
g729\uClinux-dist-2008R1.5-RC3\lib\libbfgdots\g729\src.fdpic\Makefile (2119, 2008-04-29)
g729\uClinux-dist-2008R1.5-RC3\lib\libbfgdots\g729\src.fdpic\pitch_a.asm (16599, 2007-08-17)
g729\uClinux-dist-2008R1.5-RC3\lib\libbfgdots\g729\src.fdpic\postfilt.asm (24116, 2007-08-17)
g729\uClinux-dist-2008R1.5-RC3\lib\libbfgdots\g729\src.fdpic\qua_gain.asm (16187, 2007-08-17)
g729\uClinux-dist-2008R1.5-RC3\lib\libbfgdots\g729\src.fdpic\qua_lsp.asm (15241, 2007-08-17)
g729\uClinux-dist-2008R1.5-RC3\lib\libbfgdots\g729\src.fdpic\tab_dtx.asm (4872, 2007-08-17)
g729\uClinux-dist-2008R1.5-RC3\lib\libbfgdots\g729\src.fdpic\tab_dtxe.asm (4981, 2007-08-17)
... ...

G729ab Blackfin uClinux codec README 12 October 2006 David Rowe (subcontractor to Analog Devices) Directories ----------- doc - original Analog doc file (pre uClinux port) include - include file u need to use G729A(B) libraries script - Perl script for flat-fdpic conversion plus some test files src.orig - original flat mode asm, slightly modified to support fdpic conversion by script. Perform ALL asm changes here, not in src.fdpic. src.simgot - partially converted asm code, used to incrementally test fdpic conversion scheme during development. May be useful for testing if Perl script is developed further. src.fdpic - fdpic asm files, generated auto magically by Perl script from src.orig. Do not modify any asm code in this directory. It's OK to modify globals.txt and the Makefile. Building the Libraries and Test Programs ---------------------------------------- Configure uclinux-dist to built libbfgdots. Libraries created: src.orig/libg729.a : flat mode lib src.simgot/libg729.a : simulated GOT flat mode lib (used for development) src.fdpic/libg729.a : fdpic static lib src.fdpic/libg729.so : fdpic shared lib Test programs are created in test/: g729ab : flat mode test program g729asimgot : simulated GOT flat mode test program g729ab_fdpic : fdpic test program, statically linked G729 g729ab_fdpic_so : fdpic test program, dynamically linked G729 Using the library ----------------- Please refer to the test program test/g729ab_test.c for library APIs. Bitstream formats: using 0x6b21 as SYNC word for each frame [SYNCWORD][BIT_NUM][bits] Packed mode: - g729a: [0x6b21][0x0050][80-bit] - g729ab: three kinds of frames -- BIT_NUM = 0x0050 -- BIT_NUM = 0x0010 -- BIT_NUM = 0x0000 Unpacked mode: each bit is encoded using a 16-bit word: - 0x007F for 0 - 0x0081 for 1 Running the Test Programs ------------------------- 1. Download the entire "test_data" directory and the scripts "quick.sh" and "alltests.sh" to the target. 2. Download one or more of the programs "g729ab", "g729ab_fdpic", "g729ab_fdpic_so". If you are testing the .so also download "libg729ab.so". 3. To run a quick test: root:/var/tmp> ./quick.sh g729ab_testfdpic Average MIPs: 6.60 Average MIPs: 1.33 Average MIPs: 4.69 Average MIPs: 1.54 root:/var/tmp> The four cases tested are G729A enc, G729A dec, G729AB enc, G729AB dec. If there are no other messages the test passed. It will stop with a "diff" error message if the vectors don't match, indicating a fail. The test above was run on a BF533 STAMP with write-back cache , 400MHz CCLK, 80MHz SCLK. MIPs are measured based on cycles used, so are independent of the actual clock. The same test on a BF537 STAMP with write-through cache, 500MHz CCLK, 100MHz SCLK: root:/var/tmp> ./quick.sh g729ab_testfdpic Average MIPs: 8.02 Average MIPs: 2.20 Average MIPs: 5.91 Average MIPs: 2.19 root:/var/tmp> The difference is speed may be due to the write-through cache. 4. To run all the test vectors: root:/var/tmp> ./alltests.sh g729ab_testfdpic root:/var/tmp> As with the quick test - no output means the tests passed. 5. Test multi-channel operation: root:/var/tmp> ./g729ab_testfdpic --multi Multi-threaded test, 8 encoders and 8 decoders Average MIPs: 8.03 Average MIPs: 8.71 Average MIPs: 8.*** Average MIPs: 8.65 Average MIPs: 8.67 Average MIPs: 8.67 Average MIPs: 8.58 Average MIPs: 8.58 Average MIPs: 3.14 Average MIPs: 3.10 Average MIPs: 3.06 Average MIPs: 3.02 Average MIPs: 3.02 Average MIPs: 2.94 Average MIPs: 2.92 Average MIPs: 2.88 The test above was run on a BF533 STAMP with write-back cache, 400MHz CCLK, 80MHz SCLK. The larger values are the average MIPs for each encoder, the smaller values are decoder threads. If any of the threads failed to process correctly (as compared to the test vectors) an error message will be printed and the test terminated. The MIPs values are slightly higher than (3) above as they include some overhead for testing and file I/O. Due to the design of the test program and threading model is was difficult to separate the pure G729 execution time from other work being done by the CPU. The purpose of the --multi mode is to prove simultaneous channels can run without interfering with each other (for example state variable clashing), rather than accurately measure MIPs. Test (3) above is a better test for measuring MIPs as the test design guarantees that the entire CPU is available for the G729 code under test. This program starts 8 encoders and 8 decoders, each running a set of test data files. The results of each thread are checked against the output test vectors files. In this mode (--multi flag) each thread models typical execution on a real time system, e.g.: ... encode sleep 10ms encode sleep 10ms ... In practice threads sleep most of the time, and only wake up when they are presented with input data from an I/O process. Most of the time each thread is sleeping, in that time the CPU can process other threads. Requirements ------------ The requirements for this work were: REQ1/ Convert the flat mode G729AB assembler to fdpic. Object code compiled as fdpic can (i) be relocated to L1 for efficient operation and (ii) be placed in a .so to support mixing of GPL with non-GPL licensed code. REQ2/ To reduce future maintenance develop a single code base than can be used for flat and fdpic modes. REQ3/ Use the minimum amount of L1 memory, for example 8 channels shouldn't use 8 times the amount of L1 as 1 channel. REQ4/ Develop an efficient threading model to support simultaneous operation by multiple threads. There should be no undue delays or excessive CPU usage due to spin locks. Flat to FDPIC Conversion ------------------------ The G729AB code consists of 12,000 lines of intricate hand-crafted assembler. This is is a large amount of very complex code in a very difficult development environment (optimised assembler on an embedded platform). Any modifications to the source are likely to introduce bugs that will be difficult and time consuming to find. Therefore a strong theme of this work was to minimise the need for modification to the asm code. The code is non-re-entrant, as all of the states are stored as globals rather than offsets into a structure, for example: I0.H = sharp; I0.L = sharp; To meet (REQ1) all of the global references need to be changed to the form: R0 = [P3+sharp@GOT17M4]; IO = [R0]; In fdpic mode P3 is a dedicated GOT table pointer that must be preserved. However the flat mode asm uses every pointer register in many routines, therefore to free P3 throughout the code would mean a major re-write of the code, and large amounts of complex debugging. After analysing the code it was discovered that M2 was rarely used - this was therefore chosen as the GOT offset. However this results in more complex code for each variable lookup: [--SP] = R0; [--SP] = P3; P3 = M2; R0 = [P3+sharp@GOT17M4]; I0 = R0 P3 = [SP++]; R0 = [SP++]; Note we save R0/P3 to the stack to make sure our look ups don't affect operation of the surrounding code. This resulted in more MIPs to execute the code (especially with all code and variables in external memory), however once placed in L1 the efficiency was acceptable (see tests above). There are a large number of global look ups that need to be converted to fdpic. However as the conversion of each look up was very similar a Perl script (flat2fdpic.pl) was written to automate the process and reduce the chance of error. A Makefile is used to generate the fdpic asm from the flat mode code automatically. Thus if the flat mode source is modified, the fdpic code is automatically updated. This meets the single code base requirement REQ2. The asm code in src.fdpic should not be modified directly. Modify the code in src.orig instead. Function Calls in FDPIC ----------------------- During testing of the .so a problem was discovered with calls to internal functions. The .so linking process forces all calls to jump to a PLT table to look up the address of the actual function being called. The PLT table is generated when the .so is linked. For example: $ bfin-uclinux-objump -d libg729.so 6972: ff e3 6c e2 CALL 2e4a <__init+0x37e>; which calls this code: 2e4a: 19 e5 a2 ff P1=[P3+-376]; 2e4e: 1b e5 a3 ff P3=[P3+-372]; 2e52: 51 00 JUMP (P1); However the PLT code assumes that several registers are reserved (P1 and P3). The G729 asm does not reserve these registers, in fact they are sometimes even used to pass function arguments. To preserve these registers a large amount of asm code would need to be modified and debugged. What we need is a way to use regular function calls (non-PLT) for local functions. That way we can use any register and argument passing conventions we like internally. Fortunately, in .so's, functions with local context use regular (non PLT) linking. For example this code in C: static void foo() { } void bar() { foo(); } would result in a .so that used a PLT to look up the address of bar(), but regular linking (ie a direct function call) to foo(). However our asm functions are in different modules, the technique above assumes that the local functions are in the same module. The trick is to use ld and objcopy to partially link all of the object files, then convert the global function symbols to locals. When the .so is built, only a select few functions remain globals. These are the functions in the .so that are called by external programs. Only these functions are called via the PLT method. The Makefile in src.fdpic handles this process. Support for Multiple Codec Instances ------------------------------------ To support REQ3 the codec swaps state variables in and out of its local storage at the start and end of each operation. The local storage is placed in L1 using linker options. This introduces some overhead for the swapping but minimises L1 usage. It also means that only one instance of the codec can run at any one time, requiring a mutex to block access to the encode/decode functions by simultaneous threads. In practice this mutex has little effect on performance, as the codec executes very quickly (300-400us worst case). For example in a typical 10ms processing frame, in most cases the codec will execute immediately. In the case where thread B is blocked while thread A calls the codec, B will sleep until A has finished. This is a reasonable strategy as it minimises CPU cycles, thus meeting REQ4. Note that the multi-tasking approach is designed for "one application, multiple threads". Multiple apps using the .so may work, but hasn't been tested. It will use an amount of L1 proportional to the number of applications using the .so. Notes and Further Work ---------------------- + The "simulated GOT" code in src.simgot was used to incrementally develop the fdpic conversion process. It generates code that can execute in flat mode, however the assembler is very close to the fdpic code. By converting one asm file at a time the fdpic conversion process was debugged without having to convert and test the entire code base in one "big bang". + The Perl script could be modified to generate more optimised code. This will provide small performance gains in L1, but significant gains if operating in external memory. + To best optimise the code further it is recommended that the asm be rewritten to free P3/R0 (or any Rx register). This is a significant job. + After some experimentation it was decided to use one mutex for both the encode and decode functions. When two mutexes were tried, the encoder was getting corrupted. This suggests that there are state variables shared between the encoder and decoder. + If during assembly you get something like: vad.asm:520: Error: pcrel too far BFD_RELOC_BFIN_10 This means that the Perl script has introduced enough asm source code to make a relative jump too far away to assemble. You need to modify the src.orig file to use an absolute jump. + The Perl script relies on global look ups to be in this form: I0.H = sharp; I0.L = sharp; This: I0.L = sharp; I0.H = sharp; and this: I0.H = sharp; (some other asm) I0.L = sharp; will cause the Perl script to stop with an error. Most of the src.orig mods were to get the src into a from the Perl script could understand. After each mod the flat code (g729ab_test) was run to ensure no errors had been introduced.

近期下载者：

相关文件：

评论：[我要评论] [举报此文件]

收藏者：