linaro-toolchain

linaro-toolchain@lists.linaro.org

8 participants
5694 discussions

[ANNOUNCE] Linaro GCC 4.8, 4.7 and 4.6 2013.04 released

by Yvan Roux

The Linaro Toolchain Working Group is pleased to announce the 2013.04 release of Linaro GCC 4.8, Linaro GCC 4.7 and Linaro GCC 4.6. Linaro GCC 4.8 2013.04 is the first release in the 4.8 series. Based off the latest GCC 4.8.0+svn197294 release, it includes performance improvements and bug fixes. Interesting changes include: * Our first 4.8 based release * Updates to GCC 4.8.0+svn197294 * Initial optimized support for Cortex-A53 for arm*-*-* targets * Improved support for new ARMv8-A instructions for arm*-*-* and aarch64*-*-* targets. * Backport of optimizations concerning whether to use Neon for 64-bit bitops for arm*-*-* targets. Linaro GCC 4.7 2013.04 is the thirteenth and last development release in the 4.7 series before entering maintenance. Based off the latest GCC 4.7.2+svn197188 release, it includes ARM-focused performance improvements and bug fixes. Interesting changes include: * Updates to GCC 4.7.2+svn197188 * Includes arm/aarch64-4.7-branch up to svn revision 196381 * Backport vectorizer cost model * Turn off 64-bit Bitops in Neon Linaro GCC 4.6 2013.04 is the 26th release in the 4.6 series. Based off the latest GCC 4.6.3+svn197511 release, this is the thirteenth release after entering maintenance and the last regular one. Interesting changes include: * Updates to 4.6.3+svn197511 The source tarballs are available from: https://launchpad.net/gcc-linaro/+milestone/4.8-2013.04 https://launchpad.net/gcc-linaro/+milestone/4.7-2013.04 https://launchpad.net/gcc-linaro/+milestone/4.6-2013.04 Downloads are available from the Linaro GCC page on Launchpad: https://launchpad.net/gcc-linaro More information on the features and issues are available from the release pages: https://launchpad.net/gcc-linaro/4.8/4.8-2013.04 https://launchpad.net/gcc-linaro/4.7/4.7-2013.04 https://launchpad.net/gcc-linaro/4.6/4.6-2013.04 Mailing list: http://lists.linaro.org/mailman/listinfo/linaro-toolchain Bugs: https://bugs.launchpad.net/gcc-linaro/ Questions? https://ask.linaro.org/ Interested in commercial support? Inquire at support(a)linaro.org

13 years

cbuild binutils job

by Will Newton

Hi all, Currently the binutils job that gets run via cbuild configures with --enable-gold. I guess this could be useful to ensure the gold build is not broken, but has the downside of slowing down the build and causing make check to fail. I propose we do not enable gold until such a time as we wish to formally support gold and fix the broken make check. Does anybody have any objections to doing that? Thanks, -- Will Newton Toolchain Working Group, Linaro

13 years

[PATCH] libhadoop: CRC: ARM NEON support.

by Steve Capper

Hello linaro-toolchain, This is a patch to add a NEON optimised buffer folding routine to Hadoop 2.0.2-alpha. It allows for faster CRC32c checksum computation. The code is in the form of NEON intrinsics, and details of the implementation can be found at: https://wiki.linaro.org/LEG/Engineering/CRC This patch is against Hadoop 2.0.2-alpha: https://github.com/apache/hadoop-common/tree/branch-2.0.2-alpha I've sent this to linaro-toolchain because I would like the NEON intrinsic implementation to be scrutinised. I would be particularly interested in any superfluous instructions being identified or any bad practices that could be cleaned up before opening these patches up more. Any comments/critique would really be appreciated. Thanks, -- Steve Signed-off-by: Steve Capper <steve.capper(a)linaro.org> --- .../hadoop-common/src/JNIFlags.cmake | 2 + .../native/src/org/apache/hadoop/util/bulk_crc32.c | 142 +++++++++++++++++++- 2 files changed, 142 insertions(+), 2 deletions(-) diff --git a/hadoop-common-project/hadoop-common/src/JNIFlags.cmake b/hadoop-common-project/hadoop-common/src/JNIFlags.cmake index aba4c18..5c9336a 100644 --- a/hadoop-common-project/hadoop-common/src/JNIFlags.cmake +++ b/hadoop-common-project/hadoop-common/src/JNIFlags.cmake @@ -36,6 +36,8 @@ endif (JVM_ARCH_DATA_MODEL EQUAL 32) # Determine float ABI of JVM on ARM Linux if (CMAKE_SYSTEM_PROCESSOR MATCHES "^arm" AND CMAKE_SYSTEM_NAME STREQUAL "Linux") + message("Enabling NEON support.") + set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mfpu=neon") find_program(READELF readelf) if (READELF MATCHES "NOTFOUND") message(WARNING "readelf not found; JVM float ABI detection disabled") diff --git a/hadoop-common-project/hadoop-common/src/main/native/src/org/apache/hadoop/util/bulk_crc32.c b/hadoop-common-project/hadoop-common/src/main/native/src/org/apache/hadoop/util/bulk_crc32.c index 8822b5c..6282c60 100644 --- a/hadoop-common-project/hadoop-common/src/main/native/src/org/apache/hadoop/util/bulk_crc32.c +++ b/hadoop-common-project/hadoop-common/src/main/native/src/org/apache/hadoop/util/bulk_crc32.c @@ -41,6 +41,16 @@ static inline uint32_t crc_val(uint32_t crc); static uint32_t crc32_zlib_sb8(uint32_t crc, const uint8_t *buf, size_t length); uint32_t crc32c_sb8(uint32_t crc, const uint8_t *buf, size_t length); +#ifdef __ARM_NEON__ +#include <arm_neon.h> +static uint32_t crc32c_neon(uint32_t crc, const uint8_t *buf, size_t length); +#undef USE_PIPELINED +#define CRC32C_FUNC crc32c_neon +#else +#define CRC32C_FUNC crc32c_sb8 +#endif /* __ARM_NEON__ */ + + #ifdef USE_PIPELINED static void pipelined_crc32c(uint32_t *crc1, uint32_t *crc2, uint32_t *crc3, const uint8_t *p_buf, size_t block_size, int num_blocks); #endif @@ -58,7 +68,7 @@ int bulk_calculate_crc(const uint8_t *data, size_t data_len, crc_update_func = crc32_zlib_sb8; break; case CRC32C_POLYNOMIAL: - crc_update_func = crc32c_sb8; + crc_update_func = CRC32C_FUNC; break; default: return -EINVAL; @@ -100,7 +110,7 @@ int bulk_verify_crc(const uint8_t *data, size_t data_len, do_pipelined = 1; #endif } else { - crc_update_func = crc32c_sb8; + crc_update_func = CRC32C_FUNC; } break; default: @@ -256,6 +266,134 @@ static uint32_t crc32_zlib_sb8( return crc; } +#ifdef __ARM_NEON__ + +/* + * Functions to reduce the size of the input buffer (fold) on ARM + * NEON. The smaller buffer has the same CRC32c checksum as the + * original. + * + * Most of the NEON buffer folding work takes place in the function + * below. We do the following: + * 1) 4 sets of vmull.p8's + * 2) Combine these to give a "vmull.p32" (lf3) + * 3) Shift left 1 bit to account for the endianess of multiplication. + * + * The folding and multiplication logic can be found documented at: + * https://wiki.linaro.org/LEG/Engineering/CRC + */ +static inline uint64x1_t crc32c_neon_proc_part(poly8x8_t lhs, poly8x8_t rhs1, + poly8x8_t rhs2, poly8x8_t rhs3, poly8x8_t rhs4) +{ + poly16x8_t lm1, lm2, lm3, lm4; + poly16x4x2_t lz1, lz2; + uint16x4_t le1, le2; + uint32x2_t le3; + uint32x4_t ls1, ls2, lf1, lf2; + uint64x2_t ls3, le4; + uint64x1_t lf3, lf4; + + lm1 = vmull_p8(lhs, rhs1); + lm2 = vmull_p8(lhs, rhs2); + lz1 = vuzp_p16(vget_low_p16(lm2), vget_high_p16(lm2)); + le1 = veor_u16(vreinterpret_u16_p16(lz1.val[0]), + vreinterpret_u16_p16(lz1.val[1])); + ls1 = vshll_n_u16(le1, 8); + lf1 = veorq_u32(ls1, vreinterpretq_u32_p16(lm1)); + + lm3 = vmull_p8(lhs, rhs3); + lm4 = vmull_p8(lhs, rhs4); + lz2 = vuzp_p16(vget_low_p16(lm4), vget_high_p16(lm4)); + le2 = veor_u16(vreinterpret_u16_p16(lz2.val[0]), + vreinterpret_u16_p16(lz2.val[1])); + ls2 = vshll_n_u16(le2, 8); + lf2 = veorq_u32(ls2, vreinterpretq_u32_p16(lm3)); + + le3 = veor_u32(vget_low_u32(lf2), vget_high_u32(lf2)); + ls3 = vshll_n_u32(le3, 16); + le4 = veorq_u64(ls3, vreinterpretq_u64_u32(lf1)); + lf3 = vreinterpret_u64_u32(veor_u32(vget_low_u32(vreinterpretq_u32_u64(le4)), + vget_high_u32(vreinterpretq_u32_u64(le4)))); + lf4 = vshl_n_u64(lf3, 1); + return lf4; +} + +static uint32_t crc32c_neon(uint32_t crc, const uint8_t *buf, size_t length) { + poly8x8_t xor_constant, lhs1, lhs2, lhs3, lhs4, rhs1, rhs2, rhs3, rhs4; + poly8x16_t lhl1, lhl2; + + uint64_t residues[4]; + uint32_t loop; + + if (length % 32) + return crc32c_sb8(crc, buf, length); + + /* + * because crc32c has an initial crc value of 0xffffffff, we need to + * pre-fold the buffer before folding begins proper. + * The following constant is computed by: + * 1) finding a 8x32 bit value that gives a 0xffffffff crc (with initial value 0) + * (this will be 7x32 bit 0s and 1x32 bit constant) + * 2) run a buffer fold (with 0 xor_constant) on this 8x32 bit value to get the + * xor_constant. + */ + xor_constant = vcreate_p8(0x3E43E474A2870290); + + if (crc != 0xffffffff) + return crc32c_sb8(crc, buf, length); + + /* k1 = x^288 mod P(x) - bit reversed */ + /* k2 = x^256 mod P(x) - bit reversed */ + + rhs1 = vcreate_p8(0x510AC59A9C25531D); /* k2:k1 */ + rhs2 = vcreate_p8(0x0A519AC5259C1D53); /* byte swap */ + rhs3 = vcreate_p8(0xC59A510A531D9C25); /* half word swap */ + rhs4 = vcreate_p8(0x9AC50A511D53259C); /* byte swap of half word swap */ + + lhl1 = vld1q_p8((const poly8_t *) buf); + lhl2 = vld1q_p8((const poly8_t *) buf + 16); + + lhs1 = vget_low_p8(lhl1); + lhs2 = vget_high_p8(lhl1); + lhs3 = vget_low_p8(lhl2); + lhs4 = vget_high_p8(lhl2); + + /* pre-fold lhs4 */ + lhs4 = vreinterpret_p8_u16(veor_u16(vreinterpret_u16_p8(lhs4), + vreinterpret_u16_p8(xor_constant))); + + for(loop = 0; loop < (length - 32)/32; ++loop) { + uint64x1_t l1f4, l2f4, l3f4, l4f4; + + l1f4 = crc32c_neon_proc_part(lhs1, rhs1, rhs2, rhs3, rhs4); + l2f4 = crc32c_neon_proc_part(lhs2, rhs1, rhs2, rhs3, rhs4); + l3f4 = crc32c_neon_proc_part(lhs3, rhs1, rhs2, rhs3, rhs4); + l4f4 = crc32c_neon_proc_part(lhs4, rhs1, rhs2, rhs3, rhs4); + + lhl1 = vld1q_p8((const poly8_t *) (buf + 32 * (loop + 1))); + lhl2 = vld1q_p8((const poly8_t *) (buf + 32 * (loop + 1) + 16)); + + __builtin_prefetch(buf + 32 * (loop + 2)); + + lhs1 = vget_low_p8(lhl1); + lhs2 = vget_high_p8(lhl1); + lhs3 = vget_low_p8(lhl2); + lhs4 = vget_high_p8(lhl2); + + lhs1 = vreinterpret_p8_u64(veor_u64(vreinterpret_u64_p8(lhs1), l1f4)); + lhs2 = vreinterpret_p8_u64(veor_u64(vreinterpret_u64_p8(lhs2), l2f4)); + lhs3 = vreinterpret_p8_u64(veor_u64(vreinterpret_u64_p8(lhs3), l3f4)); + lhs4 = vreinterpret_p8_u64(veor_u64(vreinterpret_u64_p8(lhs4), l4f4)); + } + + vst1q_p8((poly8_t *) &residues[0], vcombine_p8(lhs1, lhs2)); + vst1q_p8((poly8_t *) &residues[2], vcombine_p8(lhs3, lhs4)); + + return crc32c_sb8(0, (const uint8_t *)residues, 32); +} + +#endif /*__ARM_NEON__ */ + /////////////////////////////////////////////////////////////////////////// // Begin code for SSE4.2 specific hardware support of CRC32C /////////////////////////////////////////////////////////////////////////// -- 1.7.9.5

13 years

[ACTIVITY] 1 - 5 April 2013

by Omair Javaid

== Progress == * Investigated gdb test cases that are failing on arm-remote gdbserver configuration. -- Fixed some failures by updating host cross compiler version -- Fixed some failures by fixing environment issue where not being loaded properly. -- All test cases that need to build a shared library and transfer it to remote target FAIL due to problems which seems like dejaganu limitations. * Ran GDB test suite on x86_64 remote gdbserver configuration and compared performance of same configuration on arm. * Got most documentation ready for Ireland Visa application, still waiting on invitation letter and hotel booking details. *** Still No blue-print available to log work in JIRA. == Plan == * Investigate compiler version specific and general failures on arm remote gdbserver configuration. * Start on investigation/fixing of arm specific failures in gdb test suite results. * Submit Ireland visa application after receiving invitation letter and hotel booking details. * Planned Holiday -- Planned Day off on Friday 12th April 2013 for car checkup in workshop.

13 years

[ACTIVITY] 2 - 5 Apr 2013

by Kugan

== Summary == - benchmarking coremark with VRP based extension elimination * extension elimination in some cases affecting other optimizations * With this improvements are marginal (details below) == Plan == - study crc where extension elimination is resulting in bad code - Find a solution ==Details== If an assignment gimple statement has RHS expression value that can fit in LHS type, truncation is redundant. Zero/sign extensions are redundant in this case and rtl statement can be replaced as from: (insn 12 11 0 (set (reg:SI 110 [ D.4128 ]) (zero_extend:SI (subreg:HI (reg:SI 117) 0))) c5.c:8 -1 (nil)) to: (insn 12 11 0 (set (subreg/s/u:HI (reg:SI 110 [ D.4128 ]) 0) (subreg:HI (reg:SI 117) 0)) c5.c:8 -1 (nil)) With this change, for the following case: short unPack( unsigned char c ) { /* Only want lower four bit nibble */ c = c & (unsigned char)0x0F ; if( c > 7 ) { /* Negative nibble */ return( ( short )( c - 16 ) ) ; } else { /* positive nibble */ return( ( short )c ) ; } } asm without elimination unPack: @ args = 0, pretend = 0, frame = 0 @ frame_needed = 0, uses_anonymous_args = 0 @ link register save eliminated. and r0, r0, #15 cmp r0, #7 subhi r0, r0, #16 uxthhi r0, r0 sxth r0, r0 bx lr .size asm with elimination unPack: @ args = 0, pretend = 0, frame = 0 @ frame_needed = 0, uses_anonymous_args = 0 @ link register save eliminated. and r0, r0, #15 cmp r0, #7 subhi r0, r0, #16 sxth r0, r0 bx lr In some cases, changed rtl statement is not eliminated by later passes and is generated as a mov instruction. Worse, it also seems to affect the other optimization passes and resulting in worse code for crc. Not found the cause for it yet.

13 years

[ACTIVITY] 1 - 5 Apr 2013

by Venkataramanan Kumar

== Progress == * gc sections tests Completed adding gc-section test cases. 1. TLS and GOT tests. http://sourceware.org/ml/binutils/2013-03/msg00273.html 2. PLT tests. http://sourceware.org/ml/binutils/2013-03/msg00273.html * Evaluate gprof work. There is one function hook "find_call" which is machine dependent and is written for i386.c/sparc Then there are changes in configure.in to add architecture details. * 1-1 with Matt. == Plan == Continue gprof support work for Aarch64

13 years

[ACTIVITY] Week 14

by Renato Golin

== Progress == * Buildbots - All ARM build-bots GREEN! Hurray!!! :D - Improving sort comparison functions makes output reproducible on all machines - Fixed fpcmp, which fixes sqlite3 report - A bad commit broke lots of tests, reverted - Adding a self-hosting check-all Pandaboard - In theory, it works, but config is not perfect yet * Arndale - Suse image is not stable enough, installed Linaro (325) - Fathi got GCC bootstrapping with it, should work - Installed a heat-sink on the board, should reduce problems - Still got network issues, may be the mac address - Someone could liaise with Dave Pigott to get a new MAC on the DHCP * EuroLLVM 2013 - Badges, finishing touches == Plan == * Holidays next week, then * Check with Galina about Beagle bots * Finish Panda self-host + test-suite A9 bot * Gather info and hardware for 3.3 release tests * Plan for the future!

13 years

[ACTIVITY] Week 14

by Will Newton

== Progress == * Short week (2 days) * Tested memcpy on big endian. * Updated newlib memcpy patch. * More digging into glibc IFUNC. == Issues == * None. == Plan == * Try and get newlib mempcy patch accepted. * IFUNC... -- Will Newton Toolchain Working Group, Linaro

13 years

[ACTIVITY] Week 14

by Yvan Roux

== Issues == * None == Progress == * Boehm-gc AArch64 support backport in GCC: - Ping maintainers, no response yet. * Libunwind AArch64 support: - No significant improvements, task preempted by review roster. * Merge request review roster: - Lots of reviews done and ongoing to prepare 4.6, 4.7 en 4.8 releases. == Plan == * Libunwind AArch64 support: - Fix and submit upstream.

13 years

[ACTIVITY] 2-5 April 2013

by Christophe Lyon

== Progress == * Disable-peeling: - Still waiting for the results from cbuild. Because of many merges (pre-release week), there are a lot of jobs in the queue :-( * Libsanitizer: - tried to understand why isatty(2) returns true when executing the testsuite via qemu. * Neon intrinsics codegen: - benchmarking in bare-machine shows that GCC is actually ~5% faster than RVCT on this sample codec. - to be further discussed internally * Neon intrinsics testsuite: - looking at FP16 support. * Turnoff 64bits ops in Neon: - backport in 4.8 done by Matt during the release merges. * Arndale: a few attempts to use it, but it's still unstable. * Internal support == Next == * Disable-peeling: Analyze bench results when available. * Revert-coalesce-vars: Analyze bench results when available. * Libsanitizer: contine isatty(2) study wrt qemu/expect. * Neon intrinsics: check internally for other codecs

13 years

Jump to page:

2026

2025

2024

2023

2022

2021

2020

2019

2018

2017

2016

2015

2014

2013

2012

2011

2010

linaro-toolchain