How to use clang to compile OpenCL to ptx code?

Question

Clang 3.0 is able to compile OpenCL to ptx and use Nvidia\'s tool to launch the ptx code on GPU. How can I do this? Please be specific.

Answer 1

Here is brief guide how to do it with Clang trunk (3.4 at this point) and libclc. I assume you have basic knowledge how to configure and compile LLVM and Clang, so I just listed the configure flags I have used.

square.cl:

__kernel void vector_square(__global float4* input,  __global float4* output) {
  int i = get_global_id(0);
  output[i] = input[i]*input[i];
}

1. Compile llvm and clang with nvptx support:

   ../llvm-trunk/configure --prefix=$PWD/../install-trunk --enable-debug-runtime --enable-jit --enable-targets=x86,x86_64,nvptx
   make install
   

2. Get libclc (git clone http://llvm.org/git/libclc.git) and compile it.

   ./configure.py --with-llvm-config=$PWD/../install-trunk/bin/llvm-config
   make
   

If you have problem compiling this you might need to fix couple of headers in ./utils/prepare-builtins.cpp

-#include "llvm/Function.h"
-#include "llvm/GlobalVariable.h"
-#include "llvm/LLVMContext.h"
-#include "llvm/Module.h"
+#include "llvm/IR/Function.h"
+#include "llvm/IR/GlobalVariable.h"
+#include "llvm/IR/LLVMContext.h"
+#include "llvm/IR/Module.h"

1. Compile kernel to LLVM IR assember:

   clang -Dcl_clang_storage_class_specifiers -isystem libclc/generic/include -include clc/clc.h -target nvptx -xcl square.cl -emit-llvm -S -o square.ll
   

2. Link kernel with builtin implementations from libclc

   llvm-link libclc/nvptx--nvidiacl/lib/builtins.bc square.ll -o square.linked.bc
   

3. Compile fully linked LLVM IR to PTX

   clang -target nvptx square.linked.bc -S -o square.nvptx.s
   

square.nvptx.s:

    //
    // Generated by LLVM NVPTX Back-End
    //
    .version 3.1
    .target sm_20, texmode_independent
    .address_size 32

            // .globl       vector_square

    .entry vector_square(
            .param .u32 .ptr .global .align 16 vector_square_param_0,
            .param .u32 .ptr .global .align 16 vector_square_param_1
    )
    {
            .reg .pred %p<396>;
            .reg .s16 %rc<396>;
            .reg .s16 %rs<396>;
            .reg .s32 %r<396>;
            .reg .s64 %rl<396>;
            .reg .f32 %f<396>;
            .reg .f64 %fl<396>;

            ld.param.u32    %r0, [vector_square_param_0];
            mov.u32 %r1, %ctaid.x;
            ld.param.u32    %r2, [vector_square_param_1];
            mov.u32 %r3, %ntid.x;
            mov.u32 %r4, %tid.x;
            mad.lo.s32      %r1, %r3, %r1, %r4;
            shl.b32         %r1, %r1, 4;
            add.s32         %r0, %r0, %r1;
            ld.global.v4.f32        {%f0, %f1, %f2, %f3}, [%r0];
            mul.f32         %f0, %f0, %f0;
            mul.f32         %f1, %f1, %f1;
            mul.f32         %f2, %f2, %f2;
            mul.f32         %f3, %f3, %f3;
            add.s32         %r0, %r2, %r1;
            st.global.f32   [%r0+12], %f3;
            st.global.f32   [%r0+8], %f2;
            st.global.f32   [%r0+4], %f1;
            st.global.f32   [%r0], %f0;
            ret;
    }