Zhigang Gong [Fri, 5 Sep 2014 08:19:30 +0000 (16:19 +0800)]
GBE: fix a bug with LLVM 3.3.
Signed-off-by: Zhigang Gong <zhigang.gong@intel.com>
Tested-by: Meng, Mengmeng <mengmeng.meng@intel.com>
Junyan He [Fri, 5 Sep 2014 08:27:30 +0000 (16:27 +0800)]
Add the missing function prototypes of any() and atom_add()
Signed-off-by: Junyan He <junyan.he@linux.intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Zhigang Gong [Fri, 5 Sep 2014 06:04:38 +0000 (14:04 +0800)]
GBE: avoid one optimization pass to generate wide integer.
Integer type wider than 64 bit is hard to handle on Gen.
Let's try to prevent ScalarReplAggregates pass to generate
such type of integer.
v2:
fix compilation error with LLVM 3.3.
Signed-off-by: Zhigang Gong <zhigang.gong@intel.com>
Reviewed-by: "Yang, Rong R" <rong.r.yang@intel.com>
Zhigang Gong [Thu, 4 Sep 2014 04:30:44 +0000 (12:30 +0800)]
GBE: remove the user defined macro cl_khr_fp64.
This is not a predefined macro according to the spec. Let's not
define it by default. This patch also disable the fp64 when enter
user kernels.
v2:
Some internal .cl files require cl_khr_fp64 enabled. Fixed that issue
by move the enable macro to ocl_types.h.
Signed-off-by: Zhigang Gong <zhigang.gong@intel.com>
Reviewed-by: Junyan He <junyan.he@linux.intel.com>
Junyan He [Mon, 1 Sep 2014 02:22:54 +0000 (10:22 +0800)]
Delete all the unused files of old huge header.
Signed-off-by: Junyan He <junyan.he@linux.intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Junyan He [Mon, 1 Sep 2014 07:28:18 +0000 (15:28 +0800)]
Use the PCH to accelerate the parsing speed of the ocl.h
We disable the valid check for the PCH to avoid path
and modified time check, which brings us some trouble.
Signed-off-by: Junyan He <junyan.he@linux.intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Junyan He [Mon, 1 Sep 2014 07:28:02 +0000 (15:28 +0800)]
Enable libocl and disable the usage of the old huge header.
The llvm ir print out is modified.
From the OCL_OUTPUT_LLVM_BEFORE_EXTRA_PASS and
OCL_OUTPUT_LLVM, we change to
OCL_OUTPUT_LLVM_BEFORE_LINK
OCL_OUTPUT_LLVM_AFTER_LINK
OCL_OUTPUT_LLVM_AFTER_GEN
The first one print out the IR before link the bitcode lib.
The second one print out the IR result after linking.
Then last one print out the IR after gen translating.
Signed-off-by: Junyan He <junyan.he@linux.intel.com>
Reviewed-by: "Song, Ruiling" <ruiling.song@intel.com>
Junyan He [Mon, 1 Sep 2014 09:16:22 +0000 (17:16 +0800)]
Add the bit code linker into the module pass.
The bit code linker will load the beignet.bc as a
lib module and link the module of the kernel together.
Then we will filter out all the dead bit code by create
an InternalizePass for the module.
After this stage, the ir will include the bitcode just
used by the cl kernel.
Signed-off-by: Junyan He <junyan.he@linux.intel.com>
Reviewed-by: "Song, Ruiling" <ruiling.song@intel.com>
Junyan He [Mon, 1 Sep 2014 02:19:42 +0000 (10:19 +0800)]
Add memcpy, memset and barrier bitcode files into libocl
Signed-off-by: Junyan He <junyan.he@linux.intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@intel.com>
Junyan He [Mon, 1 Sep 2014 02:16:17 +0000 (10:16 +0800)]
Add the ocl_defines header file into libocl
This file will be used to define some common defines
for both CL and the backend source code.
Signed-off-by: Junyan He <junyan.he@linux.intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@intel.com>
Junyan He [Mon, 1 Sep 2014 02:15:35 +0000 (10:15 +0800)]
Add the relational module into libocl as template
Signed-off-by: Junyan He <junyan.he@linux.intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@intel.com>
Junyan He [Mon, 1 Sep 2014 02:14:45 +0000 (10:14 +0800)]
Add the math function into libocl as template
Signed-off-by: Junyan He <junyan.he@linux.intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@intel.com>
Junyan He [Mon, 1 Sep 2014 02:13:04 +0000 (10:13 +0800)]
Add the integer module into libocl as template
Signed-off-by: Junyan He <junyan.he@linux.intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@intel.com>
Junyan He [Mon, 1 Sep 2014 02:12:42 +0000 (10:12 +0800)]
Add the common module into the libocl as template
Signed-off-by: Junyan He <junyan.he@linux.intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@intel.com>
Junyan He [Mon, 1 Sep 2014 02:12:26 +0000 (10:12 +0800)]
Add the gen_vector script into the libocl
This script will genenrate function defines and function
prototypes for all the vector functions.
Some modules need very verbose vector functions after their
scalar version. We will write a template for all the scalar
version and use this script the generate the vector version
and append them to the template to generate the header or
source file.
Signed-off-by: Junyan He <junyan.he@linux.intel.com>
Signed-off-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@intel.com>
Junyan He [Mon, 1 Sep 2014 02:11:52 +0000 (10:11 +0800)]
Add the convert and as modules into the libocl
The convert and as function suites have very similar
format for all tye types and vectors, and they are
really verbose. So the two scripts will generate the
code for Convert and AS separatedlly.
Signed-off-by: Junyan He <junyan.he@linux.intel.com>
Signed-off-by: Simon Richter <Simon.Richter@hogyros.de>
Reviewed-by: Zhigang Gong <zhigang.gong@intel.com>
Junyan He [Mon, 1 Sep 2014 02:11:39 +0000 (10:11 +0800)]
Add thw workitem module into the libocl
Signed-off-by: Junyan He <junyan.he@linux.intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@intel.com>
Junyan He [Mon, 1 Sep 2014 02:11:24 +0000 (10:11 +0800)]
Add vload module into the libocl
Signed-off-by: Junyan He <junyan.he@linux.intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@intel.com>
Junyan He [Mon, 1 Sep 2014 02:10:05 +0000 (10:10 +0800)]
Add printf module into libocl
Signed-off-by: Junyan He <junyan.he@linux.intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@intel.com>
Junyan He [Mon, 1 Sep 2014 02:09:28 +0000 (10:09 +0800)]
Add the sync module into the libocl
Signed-off-by: Junyan He <junyan.he@linux.intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@intel.com>
Junyan He [Mon, 1 Sep 2014 02:08:34 +0000 (10:08 +0800)]
Add the misc module into the libocl
Signed-off-by: Junyan He <junyan.he@linux.intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@intel.com>
Junyan He [Tue, 2 Sep 2014 13:28:13 +0000 (21:28 +0800)]
Add the image module into the libocl
Signed-off-by: Junyan He <junyan.he@linux.intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@intel.com>
Junyan He [Mon, 1 Sep 2014 02:07:04 +0000 (10:07 +0800)]
Add the geometric module into the libocl
Signed-off-by: Junyan He <junyan.he@linux.intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@intel.com>
Junyan He [Mon, 1 Sep 2014 02:06:35 +0000 (10:06 +0800)]
Add the atomic module into the libocl
Signed-off-by: Junyan He <junyan.he@linux.intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@intel.com>
Junyan He [Mon, 1 Sep 2014 02:06:09 +0000 (10:06 +0800)]
Add the async module into the libocl
Signed-off-by: Junyan He <junyan.he@linux.intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@intel.com>
Junyan He [Mon, 1 Sep 2014 02:05:54 +0000 (10:05 +0800)]
Add common define header files to initialize the libocl
Signed-off-by: Junyan He <junyan.he@linux.intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@intel.com>
Zhigang Gong [Thu, 4 Sep 2014 00:01:25 +0000 (08:01 +0800)]
GBE: fixup/refine a bug for image1D array's extra binding index handling.
Due to hardware limitation on Gen7/Gen75 when sampling a
surface with clamp address mode and nearest filter mode
on a integer image1Darray type surface, we have to bind
one buffer to to bti. The previous implementation hard
coded it to 128 + original index and when check whether
it is such type bti in driver layer, assume the bti reserved
is 3 which is wrong now.
This patch fixed those hard coded functions and use the
macros defined in the program.h.
Signed-off-by: Zhigang Gong <zhigang.gong@intel.com>
Reviewed-by: "Song, Ruiling" <ruiling.song@intel.com>
Zhigang Gong [Thu, 28 Aug 2014 01:26:06 +0000 (09:26 +0800)]
GBE: refine the unaligned data gathering.
Save some unecessary duplicate instructions.
Signed-off-by: Zhigang Gong <zhigang.gong@intel.com>
Reviewed-by: "Song, Ruiling" <ruiling.song@intel.com>
Zhigang Gong [Thu, 28 Aug 2014 00:44:58 +0000 (08:44 +0800)]
GBE: relax the batch byte/short load vector size restrication.
Previous restrication is that the vector size must be multiple
of DWORD. This restrication prevent the vload2/3 of char or
vload3 of ushort to be optimized. This patch relax this restrication
on the vload path.
Signed-off-by: Zhigang Gong <zhigang.gong@intel.com>
Reviewed-by: "Song, Ruiling" <ruiling.song@intel.com>
Zhigang Gong [Wed, 27 Aug 2014 03:13:15 +0000 (11:13 +0800)]
GBE: optimize unaligned char and short data vector's load.
The gather the contiguous short/char loads into a single load instruction
could give us a good pportunity to use untyped load to optimize them.
This patch enable the short/char load gathering at the load store optimize
pass. Then at the backend, it will load corresponding DWORDs then covert to
short/char accordingly by applying shift and bitwise operations.
The benchmark shows, for vload4/8/16 char or vload/2/4/8/16 short, this patch brings
about 80%-100% improvement.
Signed-off-by: Zhigang Gong <zhigang.gong@intel.com>
Reviewed-by: "Song, Ruiling" <ruiling.song@intel.com>
Zhigang Gong [Wed, 27 Aug 2014 02:33:42 +0000 (10:33 +0800)]
Add new vload benchmark/test case.
v2:
refine the benchmark case and don't mix it with normal
unit test cases.
Signed-off-by: Zhigang Gong <zhigang.gong@intel.com>
Reviewed-by: "Song, Ruiling" <ruiling.song@intel.com>
Zhigang Gong [Fri, 29 Aug 2014 02:04:38 +0000 (10:04 +0800)]
GBE: fix error in the rootn fastpath function for some special input.
The fastpath is to lose some accuracy but get fast speed. It is not
to generate error result. The rootn has many special input and need
to be taken care before we call the native pow directly.
This patch fix all the pow related failures at the OpenCV 3.0 test
suite.
Signed-off-by: Zhigang Gong <zhigang.gong@intel.com>
Reviewed-by: "Song, Ruiling" <ruiling.song@intel.com>
Zhigang Gong [Tue, 2 Sep 2014 02:34:33 +0000 (10:34 +0800)]
utests: fix two utest bugs.
Similar as the bug found by junyan, some events are
accessed before assigned.
Signed-off-by: Zhigang Gong <zhigang.gong@intel.com>
Reviewed-by: He Junyan <junyan.he@inbox.com>
Junyan He [Tue, 2 Sep 2014 02:37:02 +0000 (10:37 +0800)]
Fix a bug for runtime_barrier_list.cpp, event array out of bound
Signed-off-by: Junyan He <junyan.he@linux.intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Junyan He [Mon, 1 Sep 2014 08:18:45 +0000 (16:18 +0800)]
Fix the global string bug for printf.
When there are multi printf statements in multi kernel
fucntions within the same translate unit, if they have
the same sting parameter, the Clang will just generate
one global string named .strXXX to represent that string.
So when translating the kernel to gen, we can not unref
that global var. Just ignore it to avoid assert.
Signed-off-by: Junyan He <junyan.he@linux.intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Yang Rong [Mon, 1 Sep 2014 05:05:06 +0000 (13:05 +0800)]
Fix two bugs.
1. A INSERT_REGINSERT_REG typo.
2. Release main_buf in utest sub_buffer_check.
Signed-off-by: Yang Rong <rong.r.yang@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Yang Rong [Mon, 1 Sep 2014 05:05:05 +0000 (13:05 +0800)]
Two minor fix.
1. Some systems don't define ulong type, use unsigned long instead of..
2. Use sA, sB... instead of sa, sb... to access vector 16, because sometimes sa, sb will cause clang error.
Signed-off-by: Yang Rong <rong.r.yang@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Guo Yejun [Thu, 28 Aug 2014 23:51:49 +0000 (07:51 +0800)]
remove dependency for non-X runtime environment
Signed-off-by: Guo Yejun <yejun.guo@intel.com>
Reviewed-by: Yang Rong <rong.r.yang@intel.com>
Yang Rong [Fri, 29 Aug 2014 03:26:35 +0000 (11:26 +0800)]
Update Beignet.mdwn X11 dependency.
And also remove libgbe in the external dependencies section.
Signed-off-by: Yang Rong <rong.r.yang@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Zhigang Gong [Thu, 28 Aug 2014 05:58:31 +0000 (13:58 +0800)]
Revert "GBE: refine post register allocation scheduling for global buffers."
Different BTI buffer may point to the same memory. Let's not
change the load/store sequence. This fix a regression at the
opencv test suite OCL_Channels/MixChannels.Accuracy/8.
This reverts commit
435f63e9fde93c38331bf0231df5ee8625f88a62.
Singed-off-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Reviewed-by: Ruiling Song <ruiling.song@intel.com>
Yang Rong [Thu, 28 Aug 2014 06:37:44 +0000 (14:37 +0800)]
Only compiler X11 files and do X11 operations when found X11.
Add a build flag HAS_X11 for it.
Signed-off-by: Yang Rong <rong.r.yang@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Ruiling Song [Tue, 26 Aug 2014 07:39:24 +0000 (15:39 +0800)]
GBE: refine the llvm multi-thread related code.
LLVM 3.5 remove llvm_start/stop_multithreaded() API, instead multi-thread
support is determined when build llvm(build option LLVM_ENABLE_THREADS).
llvm_is_multithreaded() is used to check whether llvm is built with
muti-thread support.
If multi-thread is not support(LLVM3.3/3.4 or 3.5 built with LLVM_ENABLE_THREADS off),
we simply add a mutex when calling clang/llvm related API.
Signed-off-by: Ruiling Song <ruiling.song@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Ruiling Song [Tue, 26 Aug 2014 07:39:11 +0000 (15:39 +0800)]
GBE: clear deadprintfs when current function is done.
It should be cleared, to prevent invalid pointers staying there
when processing next Function.
Signed-off-by: Ruiling Song <ruiling.song@intel.com>
Reviewed-by: He Junyan <junyan.he@inbox.com>
Luo Xionghu [Tue, 26 Aug 2014 02:12:28 +0000 (10:12 +0800)]
fix opencv_test_imgproc subcase OCL_ImgProc/Accumulate.Mask regression.
This regression is caused by structural analysis when check the if-then
node, acturally there are four types of if-then node according to the
topology and fallthrough information. fallthrough check is added in this
patch.
v2: add inversePredicate member and function for BranchInstruction;
print the exact meanning of IF instruction in GEN_IR.
Signed-off-by: Luo Xionghu <xionghu.luo@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Lv Meng [Fri, 15 Aug 2014 01:16:33 +0000 (09:16 +0800)]
Fix compile warnings for CLANG compiler
1.fix data structure redefine warnings.
2.fix 'data' with variable sized type 'union<*>' not at the end of a class warning(in immediate.hpp).
3.fix implicitly conversion warning.
4.fix explicitly assigning a variable type warning.
5.fix comparison of unsigned expression < 0 is always false warning(in cl_api.c).
Signed-off-by: Lv Meng <meng.lv@intel.com>
Reviewed-by: "Song, Ruiling" <ruiling.song@intel.com>
Lv Meng [Thu, 14 Aug 2014 03:33:06 +0000 (11:33 +0800)]
Fix compile warnings for ICC compiler
1.the "const" associated functions' modification is to fix "type qualifier on return type is meaningless" for ICC compile warning.
2.the "operator new" shoud have the corresponding "operator delete" function.
3.In C++0x std::auto_ptr will be deprecated in favor of std::unique_ptr.
Signed-off-by: Lv Meng <meng.lv@intel.com>
Reviewed-by: "Song, Ruiling" <ruiling.song@intel.com>
Ruiling Song [Wed, 13 Aug 2014 01:53:33 +0000 (09:53 +0800)]
cmake: Fix a license issue.
Signed-off-by: Ruiling Song <ruiling.song@intel.com>
Reviewed-by: Yang Rong <rong.r.yang@intel.com>
Lv Meng [Fri, 8 Aug 2014 08:10:03 +0000 (16:10 +0800)]
Fix compile errors for CLANG compiler
Use vector to fix "variable length array of non-POD element type" compiler error.
The /beignet/backend/src/./ir/context.hpp "fn->immediates[index] = imm" would call a private func
'operator=' which would trigger error, and it is not being used.
The undefined reference to `check_copy_overlap' would occur in the following calling.
Signed-off-by: Lv Meng <meng.lv@intel.com>
Reviewed-by: "Song, Ruiling" <ruiling.song@intel.com>
Lv Meng [Fri, 8 Aug 2014 08:08:16 +0000 (16:08 +0800)]
Fix compile error for ICC compiler
fix the pthread_mutex_t undefine compile error and some undefined error would occur when
using math.h in C++ file.for C++ file,it is better using cmath instead off math.h.
Signed-off-by: Lv Meng <meng.lv@intel.com>
Reviewed-by: "Song, Ruiling" <ruiling.song@intel.com>
Ruiling Song [Mon, 11 Aug 2014 05:58:26 +0000 (13:58 +0800)]
GBE: Fix a bug in gatherBTI.
The needNewBTI is a state that only valid for the current candidate.
So need to reset to default value for each candidate.
This fix the regression in opencv 3.0:
./opencv_perf_objdetect OCL_Cascade_Image_MinSize_CascadeClassifier.CascadeClassifier
Signed-off-by: Ruiling Song <ruiling.song@intel.com>
Reviewed-by: Yang Rong <rong.r.yang@intel.com>
Ruiling Song [Mon, 11 Aug 2014 05:49:01 +0000 (13:49 +0800)]
GBE: initialize BTI structure to zero.
Clear to zero to avoid garbage data, as we do not
assign it later for local/constant memory access.
v2:
move initialization code into constructor.
Signed-off-by: Ruiling Song <ruiling.song@intel.com>
Reviewed-by: Yang Rong <rong.r.yang@intel.com>
Ruiling Song [Mon, 11 Aug 2014 05:48:49 +0000 (13:48 +0800)]
GBE: Fix type size for vector3
According to OCL spec, size of vector3 are aligned to vector4.
And for memory load/store, clang already aligned it to vector4.
If we do not calculate private/local memory size as vector4,
out of range memory access will appear.
This can fix the failure of opencv 3.0 case:
OCL_Arithm/MeanStdDev.Mat_Mask
v2:
vec3 constant data should be aligned to vec4.
Signed-off-by: Ruiling Song <ruiling.song@intel.com>
Reviewed-by: Yang Rong <rong.r.yang@intel.com>
Rebecca Palmer [Fri, 8 Aug 2014 11:07:24 +0000 (12:07 +0100)]
Fail gracefully on unsupported hardware
If no compatible hardware is present, clGetDeviceIDs is supposed to
report CL_DEVICE_NOT_FOUND to the caller, but in Beignet this currently
ends the whole program with exit(-1) or assert(0). This fixes this.
This is required to have a "just works" OpenCL in Debian/Ubuntu, as
their package manager doesn't know the hardware and hence commonly will
install Beignet on hardware that doesn't support it; returning an error
allows the caller to try other ICDs until it finds the right one, or to
run without using OpenCL. Previous discussion:
http://lists.alioth.debian.org/pipermail/pkg-opencl-devel/Week-of-Mon-
20140217/000096.html
http://lists.alioth.debian.org/pipermail/pkg-opencl-devel/Week-of-Mon-
20140217/000100.html
Testing if you only have supported hardware: use a chroot, the GPU
isn't visible from inside.
Identical patch in case line wrap mangles this:
https://bugs.debian.org/cgi-bin/bugreport.cgi?msg=27;filename=fail_gracefully_without_hardware;att=1;bug=745363
Signed-off-by: Rebecca Palmer <rebecca_palmer@zoho.com>
Reviewed-by: Yang Rong <rong.r.yang@intel.com>
Ruiling Song [Mon, 11 Aug 2014 02:15:14 +0000 (10:15 +0800)]
GBE: Fix a warning in getConstantPointerRegister.
compiler complains "warning: control reaches end of non-void function"
Signed-off-by: Ruiling Song <ruiling.song@intel.com>
Reviewed-by: "Song, Ruiling" <ruiling.song@intel.com>
Luo Xionghu [Wed, 6 Aug 2014 01:36:31 +0000 (09:36 +0800)]
fix the relational built-in vector function regression.
the relational vector function need return -1 instead of 1 according to
the spec.
Signed-off-by: Luo Xionghu <xionghu.luo@intel.com>
Reviewed-by: "Song, Ruiling" <ruiling.song@intel.com>
Yang Rong [Mon, 4 Aug 2014 07:04:16 +0000 (15:04 +0800)]
Fix a utest compiler_async_stride_copy typo.
And need to convert to char when compare.
Signed-off-by: Yang Rong <rong.r.yang@intel.com>
Reviewed-by: "Song, Ruiling" <ruiling.song@intel.com>
LuoXionghu [Fri, 1 Aug 2014 01:40:09 +0000 (09:40 +0800)]
improve the build performance of vector type built-in function.
expand the gentypen with loop to reduce the redundant inline for more
than 4 components type.
v2: add the greater than 4 componets conditon to avoid performace
degration.
Signed-off-by: Luo Xionghu <xionghu.luo@intel.com>
Ruiling Song [Thu, 31 Jul 2014 08:48:01 +0000 (16:48 +0800)]
GBE: remove some useless code for getting printf buffer address.
This is not used anymore.
Also fix an annoying warning.
Signed-off-by: Ruiling Song <ruiling.song@intel.com>
Zhigang Gong [Tue, 29 Jul 2014 04:55:44 +0000 (12:55 +0800)]
GBE: adjust preferred vector length.
Signed-off-by: Zhigang Gong <zhigang.gong@intel.com>
Reviewed-by: "Song, Ruiling" <ruiling.song@intel.com>
Zhigang Gong [Tue, 29 Jul 2014 04:36:42 +0000 (12:36 +0800)]
GBE: Reduce random behaviour of the code generation
There are two major types of random behviour source. One is the
register spill tick. Now we fix it to increase from 0 for
each new code generation.
The second random source is the register sorting. When two
register has the same startID or endID, their sorting order
is not determined and maybe random in different machine. This
patch mitigate this random source by introduce another comparison
if the main key is identical.
Signed-off-by: Zhigang Gong <zhigang.gong@intel.com>
Reviewed-by: "Song, Ruiling" <ruiling.song@intel.com>
Zhigang Gong [Thu, 24 Jul 2014 05:13:32 +0000 (13:13 +0800)]
utest: add new test for constant expression processing.
If we use 3-component vector in a union, it may introduce
some complex constant expression as below:
float bitcast (i32 trunc (i128 bitcast (<4 x i32> <i32
1065353216, i32
1073741824, i32
1077936128, i32 undef> to i128) to i32) to float).
To test the constant expression processing function.
Signed-off-by: Zhigang Gong <zhigang.gong@intel.com>
Reviewed-by: "Song, Ruiling" <ruiling.song@intel.com>
Zhigang Gong [Wed, 23 Jul 2014 09:02:52 +0000 (17:02 +0800)]
GBE: enable constant expression processing.
Signed-off-by: Zhigang Gong <zhigang.gong@intel.com>
Reviewed-by: "Song, Ruiling" <ruiling.song@intel.com>
Zhigang Gong [Tue, 22 Jul 2014 10:36:28 +0000 (18:36 +0800)]
GBE: complete constant expression processing.
The target is to process all possible complex nested constant expression as below:
const = type0 OP0 (const0)
const0 = type1 OP1 (const1, const2)
const1 = ...
The supported OPs are as below:
BITCAST,
ADD,
SUB,
MUL,
DIV,
REM,
SHL,
ASHR,
LSHR,
AND,
OR,
XOR
We also add support for array/vector type of immediate. Some possible examples are as below:
float bitcast (i32 trunc (i128 bitcast (<4 x i32> <i32
1064178811, i32
1064346583, i32
1062836634, i32 undef> to i128) to i32) to float)
float bitcast (i32 trunc (i128 lshr (i128 bitcast (<4 x i32> <i32
1064178811, i32
1064346583, i32
1062836634, i32 undef> to i128), i128 32) to i32) to float)
v2:
separate all private method implementations to immediate.cpp.
Signed-off-by: Zhigang Gong <zhigang.gong@intel.com>
Reviewed-by: "Song, Ruiling" <ruiling.song@intel.com>
Zhigang Gong [Tue, 22 Jul 2014 09:01:06 +0000 (17:01 +0800)]
GBE: simplify processConstant.
Preparation to support generic constant expression.
Signed-off-by: Zhigang Gong <zhigang.gong@intel.com>
Reviewed-by: "Song, Ruiling" <ruiling.song@intel.com>
Zhigang Gong [Tue, 22 Jul 2014 07:56:08 +0000 (15:56 +0800)]
GBE: refactor the immediate class to support vector data type.
Signed-off-by: Zhigang Gong <zhigang.gong@intel.com>
Reviewed-by: "Song, Ruiling" <ruiling.song@intel.com>
Zhigang Gong [Wed, 30 Jul 2014 07:49:29 +0000 (15:49 +0800)]
GBE: refine post register allocation scheduling for global buffers.
Signed-off-by: Zhigang Gong <zhigang.gong@intel.com>
Reviewed-by: Ruiling Song <ruiling.song@intel.com>
Zhigang Gong [Wed, 30 Jul 2014 07:36:01 +0000 (15:36 +0800)]
GBE: cleanup image base index related code.
Signed-off-by: Zhigang Gong <zhigang.gong@intel.com>
Reviewed-by: Ruiling Song <ruiling.song@intel.com>
Ruiling Song [Wed, 30 Jul 2014 05:59:30 +0000 (13:59 +0800)]
GBE: Handle bti allocation for internal buffer used by printf.
1. Move the bti/Register map from gbe::Context to ir::Function.
2. use GlobalVariable instead of 'call' to get internal buffer (used for printf) base address.
Signed-off-by: Ruiling Song <ruiling.song@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Ruiling Song [Wed, 30 Jul 2014 05:59:29 +0000 (13:59 +0800)]
GBE: Refine bti usage in backend & runtime.
Previously, we simply map 2G surface for memory access,
which has obvious security issue, user can easily read/write graphics
memory that does not belong to him. To prevent such kind of behaviour,
We bind each surface to a dedicated bti. HW provides automatic
bounds check. For out-of-bound write, it will be ignored. And for read
out-of-bound, hardware will simply return zero value.
The idea behind the patch is for a load/store instruction, it will search
through the LLVM use-def chain until finding out where the address
comes from. Then the bti is saved in ir::Instruction and used for
the later code generation. And for mixed pointer case, a load/store
will access more than one bti.
To simplify some code, '0' is reserved for constant address space,
'1' is reserved for private address space. Other btis are assigned
automatically by backend.
Signed-off-by: Ruiling Song <ruiling.song@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Ruiling Song [Tue, 29 Jul 2014 07:41:38 +0000 (15:41 +0800)]
runtime: set correct state for constant buffer on hsw.
According to spec, should set I965_SURCHAN_SELECT_XXX on hsw.
Then we can use sampler message to read constant surface.
This fix the regression in unit test brought by:
'GBE: Optimize constant load with sampler.'
Signed-off-by: Ruiling Song <ruiling.song@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Ruiling Song [Mon, 28 Jul 2014 01:19:30 +0000 (09:19 +0800)]
utests: Fix a bug in image_1D_buffer.
Should use buffer_sz to clCreateBuffer().
Signed-off-by: Ruiling Song <ruiling.song@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Ruiling Song [Mon, 28 Jul 2014 01:19:29 +0000 (09:19 +0800)]
GBE: align the fields in union ImageInfoKey.
To avoid possible garbage data.
Signed-off-by: Ruiling Song <ruiling.song@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Guo Yejun [Thu, 24 Jul 2014 22:00:27 +0000 (06:00 +0800)]
delete GEPInst when it is no longer used
Signed-off-by: Guo Yejun <yejun.guo@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Guo Yejun [Thu, 17 Jul 2014 23:16:34 +0000 (07:16 +0800)]
clean llvm resource in compiler (libgbe.so)
since we have separated the compiler (libgbe.so) and the interpreter
(libgbeinterp.so), the LLVM resource cleanup task should be done in
the compiler instead of the GenProgram::~GenProgram which has no way
to clean llvm resources in libgbeinterp.so
Signed-off-by: Guo Yejun <yejun.guo@intel.com>
Reviewed-by: "Luo, Xionghu" <xionghu.luo@intel.com>
Guo Yejun [Wed, 23 Jul 2014 23:51:18 +0000 (07:51 +0800)]
fix three memory leaks
Signed-off-by: Guo Yejun <yejun.guo@intel.com>
Reviewed-by: "Song, Ruiling" <ruiling.song@intel.com>
Guo Yejun [Thu, 17 Jul 2014 18:38:33 +0000 (02:38 +0800)]
free build_log when the cl program is released
Signed-off-by: Guo Yejun <yejun.guo@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Zhigang Gong [Thu, 17 Jul 2014 02:37:09 +0000 (10:37 +0800)]
NEWS: update for 0.9.2.
Signed-off-by: Zhigang Gong <zhigang.gong@intel.com>
Zhigang Gong [Thu, 17 Jul 2014 02:14:12 +0000 (10:14 +0800)]
docs: add a NEWS document to point to the release notes pages.
Signed-off-by: Zhigang Gong <zhigang.gong@intel.com>
Guo Yejun [Wed, 16 Jul 2014 17:26:23 +0000 (01:26 +0800)]
remove requirment as drm master in non-x environment
Signed-off-by: Guo Yejun <yejun.guo@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Lv Meng [Wed, 16 Jul 2014 07:38:48 +0000 (15:38 +0800)]
improve the clEnqueueCopyBufferRect performance in some cases
Signed-off-by: Lv Meng <meng.lv@intel.com>
Reviewed-by: He Junyan <junyan.he@inbox.com>
Reviewed-by: Yang Rong <rong.r.yang@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>
LuoXionghu [Wed, 16 Jul 2014 01:31:06 +0000 (09:31 +0800)]
add utest load_program_from_gen_bin.
this test case would check whether genProgramSerializeToBinary in
backend can generator gen binary correctly.
rename load_program_from_bin to load_program_from_bin_file.
the difference is load_program_from_bin_file could either load program
from llvm binary or gen binary file generated by gbe_bin_generator.
Signed-off-by: LuoXionghu <xionghu.luo@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>
LuoXionghu [Wed, 16 Jul 2014 01:31:05 +0000 (09:31 +0800)]
add platform info in the gen binary code.
the size of the platform info is 3 bytes, right after the '/0GENC'.
check the header magic number and platform info before deserializeFromBin.
v2: supports IVB/BYT/HSW binary on its' platform, and BYT binary runs on IVB.
v3: fix 'BYT' overwritten by 'IVB';
Signed-off-by: LuoXionghu <xionghu.luo@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Ruiling Song [Mon, 14 Jul 2014 09:24:38 +0000 (17:24 +0800)]
GBE: Optimize constant load with sampler.
Signed-off-by: Ruiling Song <ruiling.song@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Ruiling Song [Mon, 14 Jul 2014 09:24:37 +0000 (17:24 +0800)]
GBE: Use varying register to save one instruction
Signed-off-by: Ruiling Song <ruiling.song@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Zhigang Gong [Fri, 11 Jul 2014 15:36:39 +0000 (23:36 +0800)]
gbe: add the new else instruction to the assert checking.
Else is the new branch instruction. We already added it to the
gen_encoder's patchJMPI but forget it for gen75_encoder. Now
fix it.
Signed-off-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Yang Rong [Thu, 10 Jul 2014 03:38:41 +0000 (11:38 +0800)]
Add some hsw missed pci ids (reserved PCI IDs).
v2:
modified according to Matthias's suggestion.
Signed-off-by: Yang Rong <rong.r.yang@intel.com>
Signed-off-by: Matthias Sattler <Matthias.Sattler@t-online.de>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Zhigang Gong [Thu, 10 Jul 2014 10:57:53 +0000 (18:57 +0800)]
runtime: fix some subtle event bugs.
This patch fix the following two bugs in event handling.
1. When it's time to call a event's user call back function, we need to
set the executed to true before the call. As that call back function
may call into clReleaseEvent(), and if we don't set the executed status
to true, it will enter infinite recursive loop.
2. After the user call clEnqueueNDRangeKernel to get a valid event, the
user set a call back function to that event, and in that call back
function, it will release that event. This scenario is totally correct.
But our current event handling doesn't have a deadicated timer thread to
update those on-the-fly events' status. Thus those events will not have
a chance to get updated, and those call back function will not executed
forever. To introduce a complete timer style thread to maintain this type
of events is too heavy for this fix release. This patch choose an easy
way to work around it. It will make sure the last gpgpu event to be finished
before current task to be enqueued.
After this patch, most of the OpenCV 3.0 cases could run smoothly without
any serious issue.
Signed-off-by: Zhigang Gong <zhigang.gong@intel.com>
Reviewed-by: Yang Rong <rong.r.yang@intel.com>
Zhigang Gong [Thu, 10 Jul 2014 08:52:48 +0000 (16:52 +0800)]
runtime/driver: refine error handlings.
We should always check whether a dri_bo_map success or fail.
Signed-off-by: Zhigang Gong <zhigang.gong@intel.com>
Reviewed-by: Yang Rong <rong.r.yang@intel.com>
Yongjia Zhang [Thu, 17 Jul 2014 18:16:52 +0000 (02:16 +0800)]
GBE: fix empty block disassemble bug.
If a block is empty, there will be more than one label pointing the
same instruction we need to step over all those labels.
Signed-off-by: Zhigang Gong <zhigang.gong@intel.com>
Reviewed-by: Yongjia Zhang <zhang_yong_jia@126.com>
Yongjia Zhang [Thu, 17 Jul 2014 18:14:41 +0000 (02:14 +0800)]
Enable structural analysis
enable structural analysis and use if, else and endif in the
generated asm.
Signed-off-by: Yongjia Zhang <yongjia.zhang@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Yongjia Zhang [Thu, 17 Jul 2014 18:14:40 +0000 (02:14 +0800)]
Use instruction if else and endif manipulate structures
Use instruction if, else and endif manipulate the control flow of
identified if-then and if-else structures at backend. but this
is not enabled, just add the necessary code to backend.
Signed-off-by: Yongjia Zhang <yongjia.zhang@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Yongjia Zhang [Thu, 17 Jul 2014 18:14:39 +0000 (02:14 +0800)]
Add structure identification on ir level
Add tool structures and functions for identifying if-then and
if-else structures on Gen IR level.
Signed-off-by: Yongjia Zhang <yongjia.zhang@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Yongjia Zhang [Thu, 17 Jul 2014 18:14:38 +0000 (02:14 +0800)]
Add Gen instruction 'else'
Add Gen instruction 'else' for future use.
Signed-off-by: Yongjia Zhang <yongjia.zhang@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Yongjia Zhang [Thu, 17 Jul 2014 18:14:37 +0000 (02:14 +0800)]
Add Gen IR IF, ELSE and ENDIF
Add Gen IR IF, ELSE and ENDIF to mark the strucutred region.
Signed-off-by: Yongjia Zhang <yongjia.zhang@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Luo [Tue, 8 Jul 2014 01:07:54 +0000 (09:07 +0800)]
remove lspci, gbe_bin_genenrater would generator llvm binary by default.
driver can get chipset id by ioctl instead of calling lspci in cmake;
user could generator gen binary by configuring cmake option
-DGEN_PCI_ID=xxxx or calling the gbe_bin_generater with option -t
GEN_PCI_ID.
v2: add "\0GENC" magic code for gen binary, fix typo.
Signed-off-by: Luo <xionghu.luo@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Ruiling Song [Mon, 7 Jul 2014 05:59:24 +0000 (13:59 +0800)]
GBE: Fix builtin tanpi.
To meet precision requirement of OCL Spec .
Signed-off-by: Ruiling Song <ruiling.song@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Zhigang Gong [Tue, 8 Jul 2014 06:34:57 +0000 (14:34 +0800)]
Build: Change versioning policy.
Don't set patch version number on master branch.
Signed-off-by: Zhigang Gong <zhigang.gong@intel.com>
Yi Sun [Fri, 4 Jul 2014 03:35:57 +0000 (11:35 +0800)]
Remove the generated test cases list.
Signed-off-by: Yi Sun <yi.sun@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Zhigang Gong [Fri, 4 Jul 2014 02:04:33 +0000 (10:04 +0800)]
Build: check whether lspci exists.
Signed-off-by: Zhigang Gong <zhigang.gong@intel.com>
Zhigang Gong [Thu, 3 Jul 2014 03:33:10 +0000 (11:33 +0800)]
runtime: fix a gpgpu event and thread local gpgpu handling bug.
When pending a command queue, we need to record the whole gpgpu
structure not just the batch buffer. For the following reason:
1. We need to keep those private buffer, for example those printf buffers.
2. We need to make sure this gpgpu will not be reused by other enqueuement.
v2:
Don't try to flush all user event attached to the queue.
Just need to flush the current event when doing command queue flush.
Signed-off-by: Zhigang Gong <zhigang.gong@intel.com>
Reviewed-by: "Yang, Rong R" <rong.r.yang@intel.com>