add built-in function "frexp"

Signed-off-by: Homer Hsing <homer.xing@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>

Improve the clGetMemObjectInfo API, add more info option

Improve the clGetMemObjectInfo API, add more info option.
CL_MEM_ASSOCIATED_MEMOBJECT and CL_MEM_OFFSET need create
subbuffer implememted firstly.
Attach the test case in get_cl_info.cpp

Signed-off-by: Junyan He <junyan.he@linux.intel.com>
Reviewed-by: "Xing, Homer" <homer.xing@intel.com>

Add the support for clSetMemObjectDestructorCallback API

Reviewed-by: "Xing, Homer" <homer.xing@intel.com>

Improve the clEnqueueMapBuffer and clCreateBuffer API

In clCreateBuffer API, add the CL_MEM_ALLOC_HOST_PTR and
CL_MEM_USE_HOST_PTR flag support.
CL_MEM_ALLOC_HOST_PTR flag seem nothings special to do.
CL_MEM_USE_HOST_PTR flag will request clEnqueueMapBuffer API:
1> The host_ptr specified in clCreateBuffer is guaranteed to
contain the latest bits in the region being mapped when the
clEnqueueMapBuffer command has completed.
2> The pointer value returned by clEnqueueMapBuffer will be
derived from the host_ptr specified when the buffer object is created.

We improve the clEnqueueMapBuffer to setup a map for the mapped
address and do the data sync problem based on the address when
mapped and unmapped.

Signed-off-by: Junyan He <junyan.he@linux.intel.com>
Tested-by: Zhigang Gong <zhigang.gong@linux.intel.com>

fix vectorial built-in functions "min, max, clamp"

vectorial versions of "min,max,clamp" was missing.

Signed-off-by: Homer Hsing <homer.xing@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>

test built-in function "sign"

Signed-off-by: Homer Hsing <homer.xing@intel.com>
Tested-by: Zhigang Gong <zhigang.gong@linux.intel.com>

built-in function "sign"

Signed-off-by: Homer Hsing <homer.xing@intel.com>
Tested-by: Zhigang Gong <zhigang.gong@linux.intel.com>

test function "mad_sat"

Signed-off-by: Homer Hsing <homer.xing@intel.com>
Tested-by: Zhigang Gong <zhigang.gong@linux.intel.com>

add built-in function "mad_sat"

currently only "char, uchar, short, ushort" types of arguments are supported

Signed-off-by: Homer Hsing <homer.xing@intel.com>
Tested-by: Zhigang Gong <zhigang.gong@linux.intel.com>

utest: Query the device driver version and the open cl c version.

Signed-off-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Reviewed-by: Song, Ruiling <ruiling.song@intel.com>

CL: Refine the version string handling.

Now concentrate the version assignment at the root cmake files.
All the other place will refer the specified macros other than
hard coded a number.

Signed-off-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Reviewed-by: Song, Ruiling <ruiling.song@intel.com>

test function "smoothstep"

Signed-off-by: Homer Hsing <homer.xing@intel.com>
Tested-by: Zhigang Gong <zhigang.gong@linux.intel.com>

support built-in function "smoothstep"

Signed-off-by: Homer Hsing <homer.xing@intel.com>
Tested-by: Zhigang Gong <zhigang.gong@linux.intel.com>

test API function "clGetImageInfo"

Signed-off-by: Homer Hsing <homer.xing@intel.com>
Reviewed-by: He Junyan <junyan.he@inbox.com>

support clGetImageInfo

clGetImageInfo() is an OpenCL API. It returns information of an image.

This patch makes Piglit test case "clGetImageInfo" pass.

Signed-off-by: Homer Hsing <homer.xing@intel.com>
Reviewed-by: He Junyan <junyan.he@inbox.com>

improve clCreateContext conformance

OpenCL specification says "clCreateContext" function returns
CL_INVALID_PROPERTY if the same property name appears more than once.

But "clCreateContext" did not follow that.

This patch uses a local temp integer to track whether a property name has
appeared. If so, returns CL_INVALID_PROPERTY.

This patch makes Piglit test case "clCreateContext" pass.

Signed-off-by: Homer Hsing <homer.xing@intel.com>
Reviewed-by: Song, Ruiling <ruiling.song@intel.com>

test built-in function "bitselect"

Signed-off-by: Homer Hsing <homer.xing@intel.com>
Reviewed-by: Yang, Rong <rong.r.yang@intel.com>

support built-in function "bitselect"

Signed-off-by: Homer Hsing <homer.xing@intel.com>
Reviewed-by: Yang, Rong R <rong.r.yang@intel.com>

test const-indexed global constant array

A test case is updated, to test three feature:
  variable-indexed global constant array
  const-indexed global constant array
  global constant scalar

Signed-off-by: Homer Hsing <homer.xing@intel.com>
Reviewed-by: Yang, Rong R <rong.r.yang@intel.com>

support const indexed global constant array

support reading global constant arrays by CONST index.

example:
  constant int o[3] = {71, 72, 73};

  kernel void f(global int *dst) {
    dst[get_global_id(0)] = o[2]; // const index: 2
  }

in llvm converting phase, calculate offset from const index,
then add the offset to array head

Signed-off-by: Homer Hsing <homer.xing@intel.com>
Reviewed-by: Yang, Rong R <rong.r.yang@intel.com>

test built-in functions "degrees" and "radians"

Signed-off-by: Homer Hsing <homer.xing@intel.com>
Tested-by: Yang, Rong R <rong.r.yang@intel.com>

support built-in functions "degrees" and "radians"

Signed-off-by: Homer Hsing <homer.xing@intel.com>
Tested-by: Yang, Rong R <rong.r.yang@intel.com>

test cases for "mul24", "mad24"

Signed-off-by: Homer Hsing <homer.xing@intel.com>
Reviewed-by: He Junyan <junyan.he@inbox.com>

support built-in functions "mul24", "mad24"

Signed-off-by: Homer Hsing <homer.xing@intel.com>
Reviewed-by: He Junyan <junyan.he@inbox.com>

Fix OpenCL C version format

Signed-off-by: Simon Richter <Simon.Richter@hogyros.de>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>

Bump the version number.

The library version is now 0.2, and the Opencl spec number is 1.1.

Signed-off-by: Zhigang Gong <zhigang.gong@linux.intel.com>

test built-in function "upsample"

Signed-off-by: Homer Hsing <homer.xing@intel.com>
Reviewed-by: Song, Ruiling <ruiling.song@intel.com>

support built-in function "upsample"

Signed-off-by: Homer Hsing <homer.xing@intel.com>
Reviewed-by: Song, Ruiling <ruiling.song@intel.com>

Update beignet docs for release v0.2.

Signed-off-by: Zhigang Gong <zhigang.gong@linux.intel.com>

Add the test case for builtin step() function

The step function has two kind of prototype:
gentypen step(gentypen edge, gentypen x)
and
gentypen step(float edge, gentypen x)

The first's test name is compiler_step_floatX
The second's test name is compiler_stepf_floatX

Signed-off-by: Junyan He <junyan.he@linux.intel.com>
Reviewed-by: Song, Ruiling <ruiling.song@intel.com>

Add the step builtin function support

Signed-off-by: Junyan He <junyan.he@linux.intel.com>
Reviewed-by: Song, Ruiling <ruiling.song@intel.com>

CLGL: Refine the hack of gbm extension initialization.

Previous implementation need to refer a EGL internal symbol.
This refinement is also a hack. It just avoid use the EGL internal
symbol, but it does use the internal EGL data structure.

Anyway, before we made some changes to gbm to support our use
model, this ugly method seems the only way to achive sharing
gl 2d/3d textures.

This patch also fix the bug when it failed to get a valid egl
context it may crash.

Signed-off-by: Zhigang Gong <zhigang.gong@linux.intel.com>

GBE: Clear the value map when start a new scalarize pass.

The scalarize pass is a function pass, and the valueMap should
be a per-function data rather than a per-unit data. The reason
we put it in the unit data structure is that the scalarize pass
is before the GenWriter pass thus there is no ir::Function exists.

As there may be multiple kernel functions in one unit, if we don't
clear the valueMap each time running a new scalarize pass, the previous
data may cause some unexpected behaviour. For example, the previous
instructions have been already erased, then latter a new instruction
in this function may be created in the same position of the erased
instruction, then it breaks this valueMap. That's the root cause why
we run the unit test several times and may encounter an assertion
sometime.

This commit also modify the ir::unit layer implementation to remove
the dependency of llvm from that layer. In general, we should not add
llvm related code to the ir layer.

Signed-off-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Tested-by: He Junyan <junyan.he@linux.intel.com>

test scalar global constants

a test case was updated, to test not only constant arrays,
but also constant scalar.

Signed-off-by: Homer Hsing <homer.xing@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>

support global scalar constants

support kind of "constant int n = 1;"

collect scalar value in constant collector
also do virtual register translation for scalar constants

Signed-off-by: Homer Hsing <homer.xing@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>

Add the test case for builtin abs_diff() function

All the integer value types check are supported.
Please use the case named compiler_abs_diff_xxxx,
where xxxx means the data type such as int2, char4

Signed-off-by: Junyan He <junyan.he@linux.intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>

Add the abs_diff builtin function support

Signed-off-by: Junyan He <junyan.he@linux.intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>

Add the vector3 support for builtin abs function

Add the forgetten abs vector3 for all the types.
Because the kernel input alignment, improve the test
case to match the alignment request.

Signed-off-by: Junyan He <junyan.he@linux.intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>

CL: destroy the EGL image which is created for gl sharing when delete the mem object.

Signed-off-by: Zhigang Gong <zhigang.gong@linux.intel.com>

Modify all the builtin function vect return to (vect_name)(e1, e2, e3)

Some builtin functions has the prototype like:
int3 function_name (int3 x) { return (x.s0, x.s1, x.s2);}
which not comply with CL spec and will cause the clang IR
be translated error.
The vector declare should be (vect)(e1, e2, e3)

Signed-off-by: Junyan He <junyan.he@linux.intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>

Disable error message output in release version.

As piglit will got the error message we output to stderr and mark the case 'WARN'.
so, we disable the message to stderr, and use release version to run piglit.

also fix a minor compile fail under release version.

Signed-off-by: Ruiling Song <ruiling.song@intel.com>
Reviewed-by: Yang, Rong R <rong.r.yang@intel.com>

Fix atomic test failed in GT1.

Barrier only ensure one work group finish, can't guarantee all work item's atomic ops
have finished before the last atomic_add.
So use atomic_xchg to update first work group's local buffer to other global buffer position.

Signed-off-by: Yang Rong <rong.r.yang@intel.com>
Reviewed-by: Xing, Homer <homer.xing@intel.com>

utests: increase local size in the two barrier test cases.

Increasing the local size to 256 to bring more pressure
to barrier testing.

Signed-off-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Reviewed-by: Zou, Nanhai <nanhai.zou@intel.com>
Reviewed-by: Yang, Rong R <rong.r.yang@intel.com>
Tested-by: Sun, Yi <yi.sun@intel.com>

GBE: fixed a barrier related bug.

Actually, this commit fixed two bugs related to barrier.
1. We should set useSLM to true if we use barrier.
2. We need to set barrier id to the barrierMsg payload according to
r0.2. And we don't need to reprogram the barrierCount.

And after this fix, we don't need the work around for the local
memory barrier, thus we don't need the memory fence for local memory
barrier.

Signed-off-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Reviewed-by: Zou, Nanhai <nanhai.zou@intel.com>
Reviewed-by: Yang, Rong R <rong.r.yang@intel.com>
Tested-by: Sun, Yi <yi.sun@intel.com>

test cases for built-in functions "mul_hi", "mad_hi"

Signed-off-by: Homer Hsing <homer.xing@intel.com>
Reviewed-by: Song, Ruiling <ruiling.song@intel.com>

support built-in functions "mul_hi", "mad_hi"

Signed-off-by: Homer Hsing <homer.xing@intel.com>
Reviewed-by: Song, Ruiling <ruiling.song@intel.com>

test cases for "hadd", "rhadd"

Signed-off-by: Homer Hsing <homer.xing@intel.com>
Reviewed-by: Song, Ruiling <ruiling.song@intel.com>

support built-in functions "hadd", "rhadd"

backend now support GPU opcode "addc".
add built-in functions "hadd", "rhadd".

Signed-off-by: Homer Hsing <homer.xing@intel.com>
Reviewed-by: Song, Ruiling <ruiling.song@intel.com>

add test case for clGetContextInfo/clGetKernelInfo

Signed-off-by: Ruiling Song <ruiling.song@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>

Implement API clGetKernelInfo

Signed-off-by: Ruiling Song <ruiling.song@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>

Implement clGetContextInfo

Signed-off-by: Ruiling Song <ruiling.song@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>

put 64-bit float test cases at tail

64-bit float test cases set SIMD width to 8,
causing failure in GT1 type of Intel GPU card.

Signed-off-by: Homer Hsing <homer.xing@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>

Add the test case for builtin abs() function

All the integer value types check are supported.
Please use the case named compiler_abs_xxxx,
where xxxx means the data type such as int2, char4

Signed-off-by: Junyan He <junyan.he@linux.intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Reviewed-by: Xing, Homer <homer.xing@intel.com>
Reviewed-by: Song, Ruiling <ruiling.song@intel.com>

Add the builtin function abs() support

Signed-off-by: Junyan He <junyan.he@linux.intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Reviewed-by: Xing, Homer <homer.xing@intel.com>
Reviewed-by: Song, Ruiling <ruiling.song@intel.com>

Clear atomic dst buffer to fix atomic random fail.

Because atomic's address used as src and dst, so need to clear this address.

Signed-off-by: Yang Rong <rong.r.yang@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>

Fix options parse infinite loop bug.

Signed-off-by: Yang Rong <rong.r.yang@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>

Test case for vector type comparison results.

Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>

GBE: fixed the bug when sext a i1 to i8/i16/i32.

We need to extent it to -1 rather than 1.
Reported by Edward Ching <edward.k.ching@gmail.com>.

Signed-off-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Tested-by: Edward Ching <edward.k.ching@gmail.com>

Enable int32 atomic and fp64 extensions.

Signed-off-by: Yang Rong <rong.r.yang@intel.com>
Signed-off-by: Zhigang Gong <zhigang.gong@linux.intel.com>

CL: remove the deprecated function clSetCommandQueueProperty.

According to the Open CL spec:
The clSetCommandQueueProperty API is no longer supported in OpenCL 1.1.
so we remove the corresponding function here.

Signed-off-by: Zhigang Gong <zhigang.gong@linux.intel.com>

Fix some math function error in simd16.

INT DIV splite to simd8 but forget to set quarter_control.
Will fail when predication enable.
Change the atomic test case to trigger this bug.

Signed-off-by: Yang Rong <rong.r.yang@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>

Add atomic test case.

The test case include local memory and global memory, atomic operations from
different threads and different work groups.

Signed-off-by: Yang Rong <rong.r.yang@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>

Add all atomic built-in functions.

Treat all atomic function's operands as unsigned int, except imin/imax.
So use the different function __gen_ocl_atomic_umin and __gen_ocl_atomic_imin.
Overload different address space, local and global.

Signed-off-by: Yang Rong <rong.r.yang@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>

Add atomic help functions.

Use the data port message "Untyped Atomic Operation" to implement openCL atomic functions.
This message can handle both global memory and SLM.
For all atomic functions, the operands will be treated as unsigned int, except imax/imin.
Only add one opcode Atomic for all atomic functions in ir, and identify the atomic function
by ir::AtomicOps.

Signed-off-by: Yang Rong <rong.r.yang@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>

test case for function "rotate"

Signed-off-by: Homer Hsing <homer.xing@intel.com>
Reviewed-by: He Junyan <junyan.he@inbox.com>
Reviewed-by: Song, Ruiling <ruiling.song@intel.com>

support build-in function "rotate"

Signed-off-by: Homer Hsing <homer.xing@intel.com>
Reviewed-by: He Junyan <junyan.he@inbox.com>
Reviewed-by: Song, Ruiling <ruiling.song@intel.com>

Refine the get_local_id/... builtins.

As we could prepare correct value on runtime library side and give
a correct value in the payload for dim 0, 1 and 2. So for these 3
dim argument, we don't need to check it whether in the valid range,
we just read the payload's value.

This way, we can avoid any unecessary branching for normal usage of
these builtin functions. And could avoid a known bool related bug.

Signed-off-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Tested-by: Sun, Yi <yi.sun@intel.com>

utests: Add basic arithmetic test case

test case for + - * / % of data type (u)int8/16/32
remove duplicated cases.

Signed-off-by: Ruiling Song <ruiling.song@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>

GBE: Add more support of char and short arithmetic

add * / % support of char and short type.

Signed-off-by: Ruiling Song <ruiling.song@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>

test case for function "clz"

Signed-off-by: Homer Hsing <homer.xing@intel.com>
Reviewed-by: Junyan He <junyan.he@linux.intel.com>

support zero bit counting

support OpenCL built-in function "clz", returning number of leading zero bits

add GEN GPU "fbh", "fbl" instructions, for counting zero bits

Signed-off-by: Homer Hsing <homer.xing@intel.com>
Reviewed-by: Junyan He <junyan.he@linux.intel.com>

Add the test case for clGetProgramInfo API

Signed-off-by: Junyan He <junyan.he@linux.intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>

Add the test case for clGetCommandQueueInfo API

Because all the get clGetXXXInfo API have similar
structure in function type, we will integrate them
together, and rename the get_program_info.cpp to get_cl_info.cpp

Signed-off-by: Junyan He <junyan.he@linux.intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>

Add the support of the API: clGetCommandQueueInfo

Though we support get the CL_QUEUE_PROPERTIES,
but because the CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE
and CL_QUEUE_PROFILING_ENABLE will never be set when
create the queue, it just return a all 0 bitfield now.

Signed-off-by: Junyan He <junyan.he@linux.intel.com>
Reviewed-by: Song, Ruiling <ruiling.song@intel.com>

Docs: Rearrange documents according to wiki page on fd.o.

We use fd.o wiki to host our documents. To make the maintainence easier,
I change the direcotry and some links in the markdown files. And make
them the same as fd.o's.

Signed-off-by: Zhigang Gong <zhigang.gong@linux.intel.com>

utests: Enable the local memory barrier test case.

And fix one typo. Without the local memory barrier work around,
this test case will fail.

Signed-off-by: Zhigang Gong <zhigang.gong@linux.intel.com>

GBE: work around the local memory barrier fence issue.

According to the spec, we do not need to add a memory fence
for SLM's access. But, in pratice, we have to add it to make
sure get correct result. Don't know the root cause yet, just
use this work around now.

Signed-off-by: Zhigang Gong <zhigang.gong@linux.intel.com>

GBE: Fix a bug in encoding MATH instruction

For std::vector, a push_back may cause memory relocation if no enough
memory in the vector pool. And iterator or pointer got previously will
become invalid after relocation.

Here in GenEncoder::next(), which will call push_back(), memory
relocation may occur. Then relocation will make 'insn' point to
invalid memory that does not belong to GenEncoder::store anymore.
Then, the setting of execution_width will fail.

Signed-off-by: Ruiling Song <ruiling.song@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>

utests: Fix a typo.

Signed-off-by: Zhigang Gong <zhigang.gong@linux.intel.com>

Add one test case for clGetProgramInfo

Signed-off-by: Junyan He <junyan.he@linux.intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>

Fix several typos in unit test.

compiler_sub_bytes and compiler_sub_shorts

Signed-off-by: Ruiling Song <ruiling.song@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>

Enable generation of convert_ and as_ functions for double

Signed-off-by: Simon Richter <Simon.Richter@hogyros.de>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>

Define double vector types

Add the definition for the "doubleN" vector types

Signed-off-by: Simon Richter <Simon.Richter@hogyros.de>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>

Enable cl_khr_fp64 extension for OpenCL stdlib header

This allows the stdlib header to define overloads for doubles.

Signed-off-by: Simon Richter <Simon.Richter@hogyros.de>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>

Add long and ulong types to generated functions.

This enables all generated functions for 64 bit integers.

Signed-off-by: Simon Richter <Simon.Richter@hogyros.de>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Tested-by: He Junyan <junyan.he@linux.intel.com>

Define all convert_* functions.

These functions convert between vectors of the same length by casting each
member in turn.

Signed-off-by: Simon Richter <Simon.Richter@hogyros.de>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Tested-by: He Junyan <junyan.he@linux.intel.com>

Generate all supported as_* functions

This adds support for all reinterpreting type conversions currently
possible.

The conversion functions can be updated by invoking the
update_as.sh script.

Signed-off-by: Simon Richter <Simon.Richter@hogyros.de>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Tested-by: He Junyan <junyan.he@linux.intel.com>

add a lost special register name

Signed-off-by: Homer Hsing <homer.xing@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>

test cases for 64-bit float

Signed-off-by: Homer Hsing <homer.xing@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Tested-by: Yang Rong <rong.r.yang@intel.com>

Support 64-bit float

support:
  arithmetic(+ - *)
  store load
  immediate_value
  if else
  select

other change:
  add "nib control" field in machine instruction format
  support "nib control"
  fix "directly store float-64 after load float-64".
  change hard coded store size (4) to flexible size (4 or 8)
  when using float-64 load(store), change to SIMD8

example:

/* support arithmetic store load immediate_value */
kernel void f(global double *src, global double *dst) {
  int i = get_global_id(0);
  double d = 1.234567890123456789;
  dst[i] = d * (src[i] + d);
}

/* support if else */
kernel void f(global float *src, global double *dst) {
  int i = get_global_id(0);
  float d = 1.234567890123456789f;
  if (i < 14)
    dst[i] = d * (d + src[i]);
  else
    dst[i] = 14;
}

/* support select */
kernel void f(global float *src, global double *dst) {
  int i = get_global_id(0);
  float d = 1.234567890123456789f;
  dst[i] = i < 14 ? d : 14;
}

Signed-off-by: Homer Hsing <homer.xing@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Tested-by: Yang Rong <rong.r.yang@intel.com>

GBE: Fixed one bug in scalarize pass

I met segfault at void Scalarize::dce() randomly when I integrate
a openCL kernel to the chromium's GPU process. After discuss with
Yang Rong, I found one bug in this function. As it use two loops
to erase the dead instructions, but it doesn't set the pointer to
NULL at the first loop when it already erased the instruction. Thus
at the second loop, when it call (*i)->getParent, the (*i) may
already be deleted then it may refer a freed region and may cause
segfault.

Signed-off-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Reviewed-by: Yang Rong <rong.r.yang@intel.com>

Fix some piglit constant buffer tests fail.

If indirect move's source is scalrar reg, such as using cb[0] in kernel,
should not unpack.

Change test case compiler_function_constant0 to trigger this bug.

Signed-off-by: Yang Rong <rong.r.yang@intel.com>
Reviewed-by: Ruiling Song <ruiling.song@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>

update to OpenCL 1.1 header

Based on our current progress, we plan to implement OpenCL 1.1 rather
than OpenCL 1.2 for the next release. Thus we downgrade the header file
to 1.1 in this commit.

put OpenCL 1.1 header in include/CL/

add OpenCL 1.2 defs, used by some code, by
 "#ifndef CL_VERSION_1.2"
 " some OpenCL 1.2 defs"
 "#endif"

add OCL_CREATE_IMAGE2D, OCL_CREATE_IMAGE3D,
    OCL_CREATE_GL_IMAGE2D, OCL_CREATE_GL_IMAGE3D

update test cases

Signed-off-by: Homer Hsing <homer.xing@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>

add the support of clGetProgramBuildInfo and clGetProgramInfo

For clGetProgramBuildInfo,
CL_BUILD_IN_PROGRESS not support now
and CL_PROGRAM_BUILD_LOG need do add the info collection
logic in backend and not support too, just return null
string now.
clGetProgramInfo all are fully supported.

Signed-off-by: Junyan He <junyan.he@linux.intel.com>
Reviewed-by: Ruiling Song <ruiling.song@intel.com>

Support multi-source CL program

Also fix several errors on clBuildProgram.

v2: minor refine of cl_program_release_sources()

Signed-off-by: Ruiling Song <ruiling.song@intel.com>
Reviewed-by: He Junyan <junyan.he@linux.intel.com>

utests: Add a new local memory barrier case

We fail this case right now, disalbe it. And need more work
to check the root cause. Change the local memory barrier can
pass it, but it doesn't comply with BSPEC which says SLM doesn't
need a memory fence.

Signed-off-by: Yang Rong <rong.r.yang@intel.com>
Signed-off-by: Zhigang Gong <zhigang.gong@linux.intel.com>

Utest: Add a test case for validating built-in function get_global_size()

v1:
According to the OpenCL v1.1 & v1.2 chapter 6.11, the behavior of function get_global_size should be as following:

get_global_size(-1) = 1 (dimension:1)
get_global_size(0) = 3 (dimension:1)
get_global_size(1) = 1 (dimension:1)
get_global_size(2) = 1 (dimension:1)

get_global_size(-1) = 1 (dimension:2)
get_global_size(0) = 3 (dimension:2)
get_global_size(1) = 4 (dimension:2)
get_global_size(2) = 1 (dimension:2)
get_global_size(3) = 1 (dimension:2)

get_global_size(-1) = 1 (dimension:3)
get_global_size(0) = 3 (dimension:3)
get_global_size(1) = 4 (dimension:3)
get_global_size(2) = 5 (dimension:3)
get_global_size(3) = 1 (dimension:3)
get_global_size(4) = 1 (dimension:3)

if defined:
  globals[0] = 3;
  globals[1] = 4;
  globals[2] = 5;

v2:
    Handle different version.

    Add #if and #elif to make the test case be suitable to different version.

Signed-off-by: Yi Sun <yi.sun@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>

GBE: Fixed GBE: Fix some builtin functions' return value.

Signed-off-by: Zhigang Gong <zhigang.gong@linux.intel.com>
Tested-by: Sun, Yi <yi.sun@intel.com>

utests: enable test case for global memory barrier.

Signed-off-by: Zhigang Gong <zhigang.gong@linux.intel.com>

Add memory fence before barrier to support global memory barrier.

Signed-off-by: Yang Rong <rong.r.yang@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>