From: David Neto [mailto:dneto.llvm@gmail.com]
Sent: Tuesday, December 07, 2010 1:03 PM
To: Villmow, Micah
Cc: cfe-dev@cs.uiuc.edu; llvmdev@cs.uiuc.edu
Subject: Re: [cfe-dev] [LLVMdev] OpenCL support
>> From: llvmdev-bounces@cs.uiuc.edu [mailto:llvmdev-
bounces@cs.uiuc.edu]
>> On Behalf Of Peter Collingbourne
>> Sent: Monday, December 06, 2010 2:56 PM
>> To: David Neto
>> Cc: cfe-dev@cs.uiuc.edu; llvmdev@cs.uiuc.edu
>> Subject: Re: [LLVMdev] [cfe-dev] OpenCL support
>>
>> Hi David,
>>
>> > What do I think your patch should look like? It's true that the
>> > diag::err_as_qualified_auto_decl is inappropriate for OpenCL when
>> it's
>> > the __local addres space.
>> >
>> > But we need to implement the semantics somehow. Conceptually I
think
>> > of it as a CL source-to-source transformation that lowers
>> > function-scope-local-address-space variables into a more primitive
>> > form.
>> >
>> > I think I disagree that the Clang is an inappropriate spot for
>> > implementing this type of transform: Clang "knows" the source
>> language
>> > semantics, and has a lot of machinery required for the transform.
>> > Also, Clang also knows a lot about the target machine (e.g. type
>> > sizes, builtins, more?).
>> >
>> > So I believe the "auto var in different address space" case should
be
>> > allowed in the AST in the OpenCL case, and the local-lowering
>> > transform should be applied in CodeGen. Perhaps the lowering is
>> > target-specific, e.g. GPU-style, or more generic style as I
proposed.
>> >
>> > Thoughts?
>>
>> I've been rethinking this and perhaps coming around to this way
>> of thinking. Allocating variables in the __local address space
>> is really something that can't be represented at the LLVM level,
>> at least in a standard form.
> [Villmow, Micah] We ran across this problem in our OpenCL
implementation. However, you can create a global variable with an
'__local' address space and it works fine. There is an issue with
collision between auto-arrays in different kernels, but that can be
solved with a little name mangling. There are other ways to do this,
for example, by converting local auto-arrays into kernel local pointer
arguments with a known size.
Here's a little example to show the direction I was heading, with an
illustration as a CL-to-C translation. I believe there are no
namespace issues, but otherwise is essentially the same as the global
variable solution.
The idea is that the func scope local addr variables are like a stack
frame that is shared between the different work items in a group. So
collect all those variables in an anonymous struct, and then create a
function scope private variable to point to the one copy of that
struct. The pointer is returned by a system-defined intrinsic
function dependent on the current work item. (The system knows what
work groups are in flight, which is why you need a system-defined
intrinsic.)
So a kernel function like this:
void foo(__global int*A) {
__local int vint;
__local int *vpint;
__local int const *vcpint;
__local int volatile vvint;
int a = A[0];
vint = a;
vvint = a;
int a2 = vint;
int va2 = vvint;
barrier(CLK_LOCAL_MEM_FENCE);
A[0] = a2 + va2;
}
[Villmow, Micah] This example is incorrect. There is a race condition between the writes to vint and vvint and the reads from vint/vvint. The reason being is that all threads in a work-group share the memory that vint is allocated in. So if you have two work-items in a work-group, both work-items are writing to the same memory location and you don't know which thread is writing to that location. Also, the barrier is in the wrong location as you need to verify that all threads writes to your local before you can safely start reading to it. A quick example using two work items running in parallel.
Cycle WI 1 WI 2
1 read A[0]
2 read A[0]
3 write vint
4 write vint
5 write vvint
6 read vint
7 write vvint
8 read vint
10 read vvint
11 read vvint
12 barrier
13 barrier
It is pretty easy to see why it is a race condition. 1, there is no synchronization between store and load and two both work items are writing to the same memory. This is even more troublesome on AMD GPU's when we run up to 256 work items in parallel. I would keep this in mind when attempting to define some standard way to do this.
is translated to this, which does pass through Clang, with __local
meaning attrib addrspace(2):
extern __local void * __get_work_group_local_base_addr(void); //
intrinsic
void foo(__global int*A) {
__local struct __local_vars_s {
int vint;
int *vpint;
int const *vcpint;
int volatile vvint;
} * const __local_vars
// this is a *private* variable, pointing to *local*
addresses.
// it's a const pointer because it shouldn't change; and
being const may expose optimizations
= __get_work_group_local_base_addr(); // the new intrinsic
int a = A[0];
__local_vars->vint = a; // l-values are translated as memory
stores.
__local_vars->vvint = a;
int a2 = __local_vars->vint; // r-values are translated as memory
loads
int va2 = __local_vars->vvint;
barrier(CLK_LOCAL_MEM_FENCE);
A[0] = a2 + va2;
}
[Villmow, Micah] I'd prefer not to embed them in a structure for the simple reason that it isn't the way we do it. What we do for OpenCL at AMD on the GPU is we turn it into something like the following:
@cllocal_foo_vint int* addressspace(2) [1xi32]; <-- forgive my pseudo LLVM-IR.
@cllocal_foo_vvint int* addressspace(2) volatile [1xi32];
void foo(__global int*A) {
__local int *vint = &cllocal_foo_vint;
__local int *vpint; <-- this is just a pointer and requires no modification
__local int const *vcpint; <-- same here, only arrays or variables require modification
__local int volatile *vvint = &cllocal_foo_vvint;
int a = A[0];
*vint = a;
*vvint = a;
int a2 = *vint;
int va2 = *vvint;
A[0] = a2 + va2;
}
Also, don't forget there is a difference between local int a and local int* a, one is just a pointer to some pre-allocated memory and the other actually
requires memory to be allocated. Any implementation should ignore local pointers variables and only deal with local scalar, vector and array variables.
As an extension, the backend ought to be able to use some smarts to
simplify this down in simple cases. For example if the system only
ever allows one work group at a time, then the intrinsic could boil
down to returning a constant, and then link time optimization can
scrub away unneeded work. Similarly if you have a GPU style
environment where (as Peter described) the "local" addresses are the
same integer value but in different groups point to different storage,
then again the intrinsic returns a constant and again LTO optimizes
the result.
I haven't thought through the implications of a kernel having such
vars calling another kernel having such variables. At least the
OpenCL spec says that the behaviour is implementation-defined for such
a case. It would be nice to be able to represent any of the sane
possibilities.
[Villmow, Micah] A kernel with locally defined variables cannot call another kernel with local variables. This is illegal in OpenCL because when a kernel calls another kernel, that kernel is treated as a normal function. A normal function is not allowed to have local variables declared inside the body. Basically this restricts locally defined variables in a kernel body to the top level kernel function only.
Micah
It was really