clang/test/CodeGenCUDA/bf16.cu

   1 // REQUIRES: nvptx-registered-target
   2 // REQUIRES: x86-registered-target
   3
   4 // RUN: %clang_cc1 "-aux-triple" "x86_64-unknown-linux-gnu" "-triple" "nvptx64-nvidia-cuda" \
   5 // RUN:    -fcuda-is-device "-aux-target-cpu" "x86-64" -O1 -S -o - %s | FileCheck %s
   6
   7 #include "Inputs/cuda.h"
   8
   9 // CHECK-LABEL: .visible .func _Z8test_argPDF16bDF16b(
  10 // CHECK:        .param .b64 _Z8test_argPDF16bDF16b_param_0,
  11 // CHECK:        .param .align 2 .b8 _Z8test_argPDF16bDF16b_param_1[2]
  12 //
  13 __device__ void test_arg(__bf16 *out, __bf16 in) {
  14 // CHECK-DAG:     ld.param.u64  %[[A:rd[0-9]+]], [_Z8test_argPDF16bDF16b_param_0];
  15 // CHECK-DAG:     ld.param.b16  %[[R:rs[0-9]+]], [_Z8test_argPDF16bDF16b_param_1];
  16   __bf16 bf16 = in;
  17   *out = bf16;
  18 // CHECK:         st.b16         [%[[A]]], %[[R]]
  19 // CHECK:         ret;
  20 }
  21
  22
  23 // CHECK-LABEL: .visible .func (.param .align 2 .b8 func_retval0[2]) _Z8test_retDF16b(
  24 // CHECK:        .param .align 2 .b8 _Z8test_retDF16b_param_0[2]
  25 __device__ __bf16 test_ret( __bf16 in) {
  26 // CHECK:        ld.param.b16    %[[R:rs[0-9]+]], [_Z8test_retDF16b_param_0];
  27   return in;
  28 // CHECK:        st.param.b16    [func_retval0], %[[R]]
  29 // CHECK:        ret;
  30 }
  31
  32 __device__ __bf16 external_func( __bf16 in);
  33
  34 // CHECK-LABEL: .visible .func  (.param .align 2 .b8 func_retval0[2]) _Z9test_callDF16b(
  35 // CHECK:        .param .align 2 .b8 _Z9test_callDF16b_param_0[2]
  36 __device__ __bf16 test_call( __bf16 in) {
  37 // CHECK:        ld.param.b16    %[[R:rs[0-9]+]], [_Z9test_callDF16b_param_0];
  38 // CHECK:        st.param.b16    [param0], %[[R]];
  39 // CHECK:        .param .align 2 .b8 retval0[2];
  40 // CHECK:        call.uni (retval0),
  41 // CHECK-NEXT:   _Z13external_funcDF16b,
  42 // CHECK-NEXT:   (
  43 // CHECK-NEXT:   param0
  44 // CHECK-NEXT    );
  45 // CHECK:        ld.param.b16    %[[RET:rs[0-9]+]], [retval0];
  46   return external_func(in);
  47 // CHECK:        st.param.b16    [func_retval0], %[[RET]]
  48 // CHECK:        ret;
  49 }