a789b588cd
added cgemm_kernel_4x2_bulldozer.S
2013-06-12 15:55:27 +02:00
8eaa04acbb
added zgemm_kernel_2x2_bulldozer.S
2013-06-11 12:00:49 +02:00
d854b30ae6
Added UNROLL values for 3M to getarch_2nd.c, Makefile.system and Makefile.L3
2013-06-09 17:26:42 +02:00
d65bbec99b
added new sgemm kernel for BULLDOZER
2013-06-09 15:57:42 +02:00
e4c39c7c26
changed stack touching
2013-06-08 10:43:08 +02:00
25491e42f9
New dgemm kernel for BULLDOZER: dgemm_kernel_8x2_bulldozer.S
2013-06-08 09:40:17 +02:00
9f59f384d8
Refs #223 . Fixed s/dgemv bug on windows.
2013-06-04 16:01:05 +08:00
23965f164c
Fixed overflow internal buffer bug of (s/d/c/z)gemv on x86_64.
2013-05-29 19:48:31 +08:00
6a72840945
Fixed overflow internal buffer bug of (s/d/c/z)gemv on x86.
2013-05-29 13:23:12 +08:00
69aa6c8fb1
bad performance with some data
2013-04-28 11:14:23 +02:00
60b263f3d2
removed trsm_kernel_RT_4x4_bulldozer.S. wrong results
2013-04-27 17:23:08 +02:00
7ac306e0da
added trsm_kernel_RT_4x4_bulldozer.S
2013-04-27 16:48:48 +02:00
4cb454cdf2
added trsm_kernel_LT_4x4_bulldozer.S
2013-04-27 14:30:00 +02:00
19ad2fb128
prefetch improved. Defined 2 different kernels for inner loop
2013-04-27 13:40:49 +02:00
6821677489
minor improvements and code cleanup
2013-04-26 20:05:42 +02:00
3326f3152c
Merge pull request #213 from wernsaar/develop
...
Merged some improvements into dgemm_kernel_4x4_bulldozer.S.
2013-04-17 23:56:09 -07:00
7641f6e253
Merged some improvements into dgemm_kernel_4x4_bulldozer.S.
...
Changed the copy functions to generic to solve prefetch conflicts
2013-04-16 19:05:06 +02:00
3ad29452d1
Merge pull request #211 from wernsaar/develop
...
New version of dgemm_kernel_4x4_bulldozer.S
2013-04-15 00:20:55 -07:00
6e3f6f25a5
New version of dgemm_kernel_4x4_bulldozer.S
...
The peak performance with 8 cores is now 90 GFlops
2013-04-12 17:55:51 +02:00
724ae159ce
Fixed the Windows x86_64 ABI bug in s/daxpy kernels.
2013-03-08 22:28:34 +08:00
f300ce3df5
new optimization of dgemm kernel for bulldozer: 10% performance increase
2013-03-06 17:26:03 +01:00
66e64131ed
optimized again bulldozer dgemm kernel
2013-03-05 19:51:37 +01:00
9405f26f4b
new dgemm_kernel for bulldozer
2013-03-04 17:37:38 +01:00
5c8bf6ae0e
Merge branch 'bulldozer' into develop
2013-02-10 01:19:42 +08:00
a1ead62f28
Disable the warning of sgemm bulldozer kernel.
2013-02-09 17:03:13 +01:00
0133580148
Used sgemm bulldozer kernel on 64 bit.
2013-02-09 16:29:14 +01:00
274246651d
Merge branch 'bulldozer' of git://github.com/wernsaar/OpenBLAS into bulldozer
2013-02-09 16:25:07 +01:00
299b5a44dc
Merge branch 'develop' of github.com:xianyi/OpenBLAS into bulldozer
2013-02-09 16:22:04 +01:00
d311236dfd
Refs #189 . Fixed the bug of s/cdot about invalid reading NAN on x86_64.
2013-01-25 20:56:14 +08:00
0b08f7479e
Refs #154 . Fixed gemv_t bug about overflow 16MB buffer on x86.
2013-01-20 21:22:12 +08:00
99d1978df7
Fixed #180 . the typos in kernel/x86_64/sgemv_t.S
2013-01-12 12:31:14 +08:00
08bf6674d5
Refs #177 . Fixed sgemv_t compiling bug on Win64.
2013-01-05 11:36:39 +08:00
69200884e1
Refs #173 . Fixed overflow internal buffer bug of gemv_n on x86
2012-12-25 09:27:49 +08:00
0d1518add9
Refs #173 . Fixed overflow internal buffer bug of sgemv_t on x86
2012-12-25 09:10:17 +08:00
91ed4e4450
Refs #171 . Prevent loading the dirty number from the buffer in sgemv_t x86 kernel.
2012-12-23 23:14:17 +08:00
fd3046b32a
Refs #173 . Fixed overflow internal buffer bug of gemv_t on x86.
2012-12-23 21:47:22 +08:00
9fb341a9f8
set parameters for CORE_ATHLON
...
else dgemm_p is set to zero leading to a segfault in alloc_mmap due to
allocsize being zero
2012-12-15 16:05:33 +01:00
d48cff8cf1
Added optimized sgemm_kernel
2012-12-08 18:50:53 +01:00
f19af5ecc0
Refs #54 . Added AMD Bulldozer x86_64 dgemm kernel developed by Werner Saar <wernsaar at googlemail.com>
...
Based on the dgemm kernel for AMD Barcelona, he used AVX and FMA4 instructions.
Thank Werner Saar!
2012-12-07 01:05:11 +08:00
bfaaa975e6
Added BULLDOZER target. So far it uses barcelona kernels.
2012-12-07 00:53:31 +08:00
b7c0fa6bd2
Init AMD Bulldozer codebase.
2012-12-06 07:29:54 -05:00
cea1a885b5
Refs #154 . Fixed the build bug of dgemv_t on MinW64.
2012-11-27 07:24:04 +08:00
5f0117385e
Refs #154 . Fixed a SEGFAULT bug of dgemv_t when m is very large.
...
It overflowed the internal buffer. Thus, we split vector x into blocks when m is very large.
Thank @wangqian for this patch.
2012-11-19 22:32:27 +08:00
2573311308
refs #140 . Fixed zdot incompatibility ABI issue with GCC 4.7 on Win 32.
...
GCC 4.7 uses MSVC ABI on Win 32. This means the caller pops the hidden pointer for returning
aggregate structures larger than 8 bytes.
2012-09-24 20:34:33 +08:00
d0e731e8b8
provide support for passing CFLAGS, FFLAGS, PFLAGS, FPFLAGS to make on the command line
2012-08-21 00:31:12 -04:00
25f1a573fd
Fixed the build bug when DYNAMIC_ARCH=0.
2012-07-07 12:12:24 +08:00
857a0fa0df
Fixed the issue of mixing AVX and SSE codes in S/D/C/ZGEMM.
2012-06-25 19:00:37 +08:00
d34fce56e4
Refs #83 Fixed S/DGEMM calling conventions bug on windows.
2012-06-20 19:53:18 +08:00
6cfcb54a28
Fixed align problem in S and C precision GEMM kernels.
2012-06-20 07:38:39 +08:00
3ef96aa567
Fixed bug in MOVQ redefine and ALIGN SIZE problem.
2012-06-19 20:37:22 +08:00
f76f952547
Refs #83 #53 . Adding Intel Sandy Bridge (AVX supported) kernel codes for BLAS level 3 functions.
2012-06-19 16:37:12 +08:00
eefd30881c
Refs #113 . Fixed the build bug on AMD Bobcat 64-bit OS.
2012-06-02 21:34:23 +08:00
d3b67d0bd8
Refs #113 . Fixed the typo BOBCATE -> BOBCAT
2012-05-31 22:40:15 +08:00
d6cab3f37e
Refs #113 . Support AMD Bobcate using Barcelona kernel codes. Replace 3DNow! with MMX.
2012-05-31 18:17:45 +08:00
a53c6e2440
Merge branch 'develop' into sandybridge
2012-05-25 23:16:44 +08:00
5d657c6e67
Fixed #96 a SEGFAULT bug in samax on x86.
2012-04-26 16:50:57 +08:00
03b0eb19f7
Refs #86 . Test alpha=Nan in x86/x86_64 dscale.
2012-04-05 18:16:18 +08:00
19a48b82cf
Init Sandybridge codes based on Nehalem.
2012-03-30 20:01:03 +08:00
3871b6a86d
Merge branch 'loongson3b' into release-0.1.0
2012-03-23 01:26:44 +08:00
83ecfbb9b3
Merge branch 'loongson3a' into release-0.1.0
2012-03-23 01:26:27 +08:00
dff146e306
refs #80 . Used GEMV SSE2 kernels on x86.
2012-03-19 17:56:22 +08:00
8e53b57bb2
Appending gemmkernel and trmmkernel C code in kernel/generic, this code can be used to execute on a new platform which dose not have optimized assemble kernel.
2012-01-10 17:16:13 +00:00
66904fc4e8
BLAS3 used standard MIPS instructions without extensions on Loongson 3B.
2011-11-25 11:20:25 +00:00
0884f6b78d
Merge branch 'loongson3a' of github.com:xianyi/OpenBLAS into loongson3b
2011-11-11 14:26:49 +00:00
2d78fb05c8
Add conjugate condition to gemv.
2011-11-10 15:38:48 +00:00
b95ad4cfaf
Support detecting ICT Loongson-3B CPU.
2011-11-09 19:29:50 +00:00
3bbe3ddb31
Merge branch 'develop' of github.com:xianyi/OpenBLAS into loongson3b
2011-11-09 19:08:29 +00:00
a32e56500a
Fix the compute error of gemv when incx and incy are negative numbers.
2011-11-04 19:32:21 +00:00
c1e618ea2d
Add complete gemv function on Loongson3a platform.
2011-11-03 13:53:48 +00:00
19f5b5c132
Fixed #66 the bug in zgemv kernel with transpose matrix on 64-bit MingW (Windows).
2011-10-18 18:44:23 +08:00
c852ce3981
Ref #65 . Fixed 64-bit Windows calling convention bug in cdot and zdot.
...
According to 64-bit Windows calling convention, the return value is in %rax instead of %xmm0 in cdot kernel.
In zdot, the caller allocates a memory space for return value and sets this memory address to the first hidden parameter. Thus, the callee (zdot) should assign the result to this memory space and return the memory address in %rax.
2011-10-18 10:23:17 +08:00
e08cfaf9ca
Complete all the complex single-precision functions of level3, but the performance needs further improve.
2011-09-16 17:50:40 +00:00
ee4bb8bd25
Add ctrmm part in cgemm_kernel_loongson3a_4x2_ps.S.
2011-09-16 16:08:39 +00:00
7fa3d23dd9
Complete cgemm function, but no optimization.
2011-09-15 16:08:23 +00:00
9679dd077e
Fix some compute error.
2011-09-14 20:00:35 +00:00
7b410b7f0e
Fixed #58 zdot SEGFAULT bug with GCC-4.6. Thank Mr. John for this patch.
...
In i386 calling convention, the caller put the address of return value of zdot into the first hidden parameter.
Thus, the callee should delete this address before return.
Actually, I have fixed the same bug on x86/zdot_sse2.S (issue #32 ). However, that is not a good implementation which uses 3 instructions. Mr. John told me used "ret $0x4" to skip the first hidden address (4 bytes).
2011-09-14 23:52:51 +08:00
d238a768ab
Use ps instructions in cgemm.
2011-09-14 15:32:25 +00:00
b1fe26c45a
refs #55 . Changed DTB_ENTRIES to DTB_DEFAULT_ENTRIES in x86 gemv_n kernel codes.
2011-09-06 14:14:07 +08:00
9fc6764fa7
refs #55 . Added DTB_ENTRIES into dynamic arch setting parameters. Now, it can read DTB_ENTRIES on runtime.
2011-09-05 17:37:07 +08:00
74d4cdb81a
Fix an illegal instruction for strmm_RTLU.
2011-09-02 19:41:06 +00:00
7906146836
Fix an error for strmm_LLTN.
2011-09-02 16:57:33 +00:00
3274ff47b8
Fix an error for strmm_LLTN.
2011-09-02 16:50:50 +00:00
a059c553a1
Fix a compute error for strmm.
2011-09-02 16:00:04 +00:00
23e182ca7c
Fix stack-pointer bug for strmm.
2011-09-02 15:28:01 +00:00
a15bc95824
Add strmm part.
2011-09-02 09:15:09 +00:00
09f49fa891
Using PS instructions to improve the performance of sgemm and it is 4.2Gflops now.
2011-08-31 21:24:03 +00:00
cb0214787b
Modify compile options.
2011-08-30 20:57:00 +00:00
2e8cdd1542
Using ps instruction.
2011-08-30 20:54:19 +00:00
c8360e3ae5
Complete all the plura single precision functions of level3 on Loongson3a, the performance is 2.3GFlops.
2011-07-18 17:03:38 +00:00
68532fa9ec
Merge branch 'loongson3a' of github.com:xianyi/OpenBLAS into loongson3a
2011-06-24 09:28:12 +00:00
708d2b6255
Fix compute error in ztrmm.
2011-06-24 09:27:41 +00:00
e72113f06a
Add ztrmm and ztrsm part on loongson3a. The average performance is 2.2G.
2011-06-23 21:11:00 +00:00
14f81da375
Change prefetch length of A and B, the performance is 2.1G now.
2011-06-23 10:46:58 +00:00
fc21f7ad28
Merge branch 'release-v0.1alpha2' into loongson3a
2011-06-23 16:08:23 +08:00
1c96d345e2
Improve zgemm performance from 1G to 1.8G, change block size in param.h.
2011-06-21 22:16:23 +00:00
c4efde7713
Merge branch 'loongson3a' into release-v0.1alpha2
2011-06-21 17:50:00 +08:00
32353a9d30
Refs #20 . Fixed the installation bug with DYNAMIC_ARCH=1.
2011-06-21 17:39:08 +08:00
b3d1887745
Fixed #35 a build bug with NO_LAPACK=1 DYNAMIC_ARCH=1 FC=gfortran. I forgot to test it with gfortran in last bug fixed commit.
2011-06-09 22:59:49 +08:00
8d50a9fd1a
Fixed #35 a build bug with NO_LAPACK=1 & DYNAMIC_ARCH=1.
2011-06-09 11:38:59 +08:00
4335bca2f7
Fixed #33 ztrmm bug on Nehalem.
2011-06-07 12:53:25 +08:00
31040e4d80
Fixed #32 a SEGFAULT bug with gcc-4.6. According to i386 calling convention, The called funtion should remove the hidden return value address from the stack.
2011-06-03 13:19:54 +08:00
88d94d0ec8
Fixed #30 strmm computational error on Loongson3A.
2011-05-28 09:48:34 +00:00
fc84909115
Modify single precision compiler conditions, increasing single precision kernel code on Loongson3a.
2011-05-27 09:47:17 +00:00
5ca4e51df0
Remove the useless code, modify code comments and format.
2011-05-18 10:54:51 +00:00
fcb5ce011b
Fixed #28 . Convert the result to double precision in MIPS64 dsdot_k kernel.
2011-05-17 21:24:00 +00:00
a9320f896e
Fixed #25 dtrmm and dtrsm computational error on Loongson3A.
2011-05-14 22:00:57 +00:00
b206fc7075
Fixed #28 . Convert the result to double precision in the end of dsdot kernel.
2011-05-13 02:34:30 +08:00
29dce62b8f
Finish dtrsm_kernel_Rx.S on Loongson3A.
2011-05-11 10:44:23 +00:00
432c309f63
Finish dtrsm_kernel_Lx.S on Loongson3A.
2011-05-10 12:48:43 +00:00
d2f351d819
Modify dtrsm compiler options
2011-05-09 17:31:58 +00:00
5a991b7149
Fixed #24 drmm error on Loongson3A
2011-05-09 17:28:20 +00:00
9320933520
Completely dtrmm function.
2011-04-17 20:26:49 +00:00
921caefa56
Increased handling trmm part, no edge handling. Test size(M and N) must be a multiple of 4 .
2011-04-15 21:56:25 +00:00
ecd4c1f3d9
Modify prefetching C.
2011-04-11 22:46:36 +00:00
ab9e4ce351
Adjust kc size from 112 to 116 .
2011-04-11 22:17:57 +00:00
782205a693
Add dgemm compiler Options in KERNEL.LOONGSON3A.
2011-04-06 10:38:34 +00:00
ac494c0d04
New kernel in LOONGSON3A.
2011-04-06 10:36:44 +00:00
f405b5bcc5
Fixed the bug about Loongson3A gsLQC1 & gsSQC1 instructions in daxpy kernel. Now daxpy is correct.
2011-03-18 23:05:56 +00:00
d5cffd506a
Modified the default kernel makefile in MIPS64 arch.
2011-03-07 11:23:12 +00:00
5838f12995
Support unalign address in daxpy on loongson3a simd..
2011-03-05 10:17:10 +08:00
5444a3f8f7
Unroll to 16 in daxpy on loongson3a.
2011-03-04 17:50:17 +08:00
88cbfcc5b5
Merge commit 'origin/x86' into loongson3a
2011-03-04 14:11:52 +00:00
ce78abe37e
Merge branch 'x86' of github.com:xianyi/OpenBLAS into x86
2011-03-04 11:53:04 +08:00
8f1090d32a
Support NO_LAPACK=1 to build the lib without LAPACK functions.
2011-03-04 11:51:32 +08:00
272f62a2b6
Changed movlps macro name in capital in x86/zdot_sse2.S file.
2011-03-03 00:46:39 +08:00
36016fe349
On x86 32bits, gcc 4.4.3 generated wrong codes (movsd) from movlps in zdot_sse2.S line 191.
...
This would casue zdotu & zdotc failures. Instead, use movlpd to walk around it. Fixed #8 . Fixed #9 .
2011-03-02 18:45:43 +08:00
6eb02bbb9c
Merge remote branch 'origin/x86' into loongson3a
2011-03-02 13:52:05 +08:00
12214e1d0f
Fixed #7 . Modified axpy kernel codes to avoid unloop with incx==0 or incy==0 in x86 32bits arch.
2011-02-23 20:08:34 +08:00
0cfd29a819
Fixed #7 . 1)Disable the multi-thread and 2) Modified kernel codes to avoid unloop in axpy function when incx==0 or incy==0.
2011-02-21 00:24:21 +08:00
bfaa80c316
fixed #4 csrot & drot returned the wrong result when incx==incy==0 on i686 arch.
2011-02-18 03:00:58 +08:00
c5852d4e30
fixed #4 csrot returned the wrong result when incx==incy==0.
2011-02-16 23:39:43 +08:00
84ba64e65b
fixed a bug in drot whe incx or incy equals to zero.
2011-02-16 23:35:41 +08:00
1e671b49f3
Did the experiment with Loongson 3A 128bit load & store instruction.
2011-01-29 03:05:27 +08:00
77b7020d69
changed prefetch order.
2011-01-29 03:03:34 +08:00
e003b811ab
load x & y contiguously in axpy.
2011-01-28 11:18:50 +08:00
ebe2da8474
Modified aligned size. Added additional prefetch instruction because of cache line is 32 bytes in Loongson 3A.
2011-01-27 23:07:06 +08:00
c0b5992fab
added axpy kernel with prefetch for Loongson3A. To-Do: tuning prefetch distance & instruction order.
2011-01-26 22:34:33 +08:00
342bbc3871
Import GotoBLAS2 1.13 BSD version codes.
2011-01-24 14:54:24 +00:00
wernsaar
Zhang Xianyi
wangqian
Julian Taylor
Jameson Nash
Xianyi Zhang
unknown
Xianyi Zhang
Zhang Xiianyi