• John Colvin (6/6) Oct 15 2014 How unrealistic would it be to compile a small subset of D, e.g.
• Kai Nacke (11/17) Oct 15 2014 As far as I know SPIR is a LLVM IR subset based on LLVM 3.0.
• John Colvin (3/22) Oct 16 2014 SPIR 2.0 is based on llvm 3.2 IIRC. Is ldc likely to require >3.2

"John Colvin" <john.loughran.colvin gmail.com> writes:

How unrealistic would it be to compile a small subset of D, e.g. 
betterc, to OpenCL SPIR https://www.khronos.org/spir ?

It would be fantastic to use all of D's code-generation and 
compile-time abstractions for writing opencl kernels. In 
addition, compile-time introspection of the kernel code would 
allow for significant improvements in the host API.

"Kai Nacke" <kai redstar.de> writes:

Hi John!

On Wednesday, 15 October 2014 at 18:14:22 UTC, John Colvin wrote:
 How unrealistic would it be to compile a small subset of D, 
 e.g. betterc, to OpenCL SPIR https://www.khronos.org/spir ?

 It would be fantastic to use all of D's code-generation and 
 compile-time abstractions for writing opencl kernels. In 
 addition, compile-time introspection of the kernel code would 
 allow for significant improvements in the host API.

As far as I know SPIR is a LLVM IR subset based on LLVM 3.0. 
Currently, I don't know what needs to be done here.

But in general it should be possible. Here is a tutorial which 
shows how to compile Ada to a CUDA kernel: 
http://markmail.org/download.xqy?id=4xfv2vajltcydmhi&number=1
This can be done with D, too. It requires some fiddeling with 
address spaces which is currently not implemented.

Regards,
Kai

"John Colvin" <john.loughran.colvin gmail.com> writes:

On Thursday, 16 October 2014 at 04:32:58 UTC, Kai Nacke wrote:
 Hi John!

 On Wednesday, 15 October 2014 at 18:14:22 UTC, John Colvin 
 wrote:
 How unrealistic would it be to compile a small subset of D, 
 e.g. betterc, to OpenCL SPIR https://www.khronos.org/spir ?

 It would be fantastic to use all of D's code-generation and 
 compile-time abstractions for writing opencl kernels. In 
 addition, compile-time introspection of the kernel code would 
 allow for significant improvements in the host API.

 As far as I know SPIR is a LLVM IR subset based on LLVM 3.0. 
 Currently, I don't know what needs to be done here.

SPIR 2.0 is based on llvm 3.2 IIRC. Is ldc likely to require >3.2 
any time soon?

 But in general it should be possible. Here is a tutorial which 
 shows how to compile Ada to a CUDA kernel: 
 http://markmail.org/download.xqy?id=4xfv2vajltcydmhi&number=1
 This can be done with D, too. It requires some fiddeling with 
 address spaces which is currently not implemented.

 Regards,
 Kai