[LLVMdev] How to generate constant memory for ptx code by NVPTX?

[Antony Yu]

Hello,

I work on compiling OpenCL kernel to PTX code by clang and NVPTX with
libclc.
I have a kernel that contains constant variable declared in file scope like
this:

   constant one_f = 1.0f;
__kernel void test( ...){    ...  }

Then it is compiled to llvm-ir:

@one_f = addrspace(4) const float 1.000000e+00, align 4
define void test(...){   ...   }

Finally ptx:

.visible .global .align 4 .f32 one_f = 0f3F800000;
.entry test( ...) { ... }

one_f is placed in global memory and load it by ld.global.f32
But I want one_f to be placed in constant memory and load it by
ld.const.f32. Is it any way to achieve this ??

Thanks in advance.
Antony Yu



--
View this message in context: http://llvm.1065342.n5.nabble.com/How-to-generate-constant-memory-for-ptx-code-by-NVPTX-tp58384.html
Sent from the LLVM - Dev mailing list archive at Nabble.com.