clang/test/Driver/hip-toolchain-rdc.hip

   1 // REQUIRES: x86-registered-target
   2 // REQUIRES: amdgpu-registered-target
   3
   4 // RUN: %clang -### --target=x86_64-linux-gnu \
   5 // RUN:   -x hip --cuda-gpu-arch=gfx803 --cuda-gpu-arch=gfx900 \
   6 // RUN:   --hip-device-lib=lib1.bc --hip-device-lib=lib2.bc \
   7 // RUN:   --hip-device-lib-path=%S/Inputs/hip_multiple_inputs/lib1 \
   8 // RUN:   --hip-device-lib-path=%S/Inputs/hip_multiple_inputs/lib2 \
   9 // RUN:   -fuse-ld=lld -fgpu-rdc -nogpuinc \
  10 // RUN:   -fhip-dump-offload-linker-script \
  11 // RUN:   %S/Inputs/hip_multiple_inputs/a.cu \
  12 // RUN:   %S/Inputs/hip_multiple_inputs/b.hip \
  13 // RUN: 2>&1 | FileCheck -check-prefixes=CHECK,LNX %s
  14
  15 // RUN: %clang -### --target=x86_64-pc-windows-msvc \
  16 // RUN:   -x hip --cuda-gpu-arch=gfx803 --cuda-gpu-arch=gfx900 \
  17 // RUN:   --hip-device-lib=lib1.bc --hip-device-lib=lib2.bc \
  18 // RUN:   --hip-device-lib-path=%S/Inputs/hip_multiple_inputs/lib1 \
  19 // RUN:   --hip-device-lib-path=%S/Inputs/hip_multiple_inputs/lib2 \
  20 // RUN:   -fuse-ld=lld -fgpu-rdc -nogpuinc \
  21 // RUN:   -fhip-dump-offload-linker-script \
  22 // RUN:   %S/Inputs/hip_multiple_inputs/a.cu \
  23 // RUN:   %S/Inputs/hip_multiple_inputs/b.hip \
  24 // RUN: 2>&1 | FileCheck -check-prefixes=CHECK,MSVC %s
  25
  26 // check code object alignment in dumped llvm-mc input
  27 // LNX: .protected __hip_fatbin
  28 // LNX: .type __hip_fatbin,@object
  29 // LNX: .section .hip_fatbin,"a",@progbits
  30 // MSVC: .section .hip_fatbin, "dw"
  31 // CHECK: .globl __hip_fatbin
  32 // CHECK: .p2align 12
  33 // CHECK: __hip_fatbin:
  34 // CHECK: .incbin "[[BUNDLE:.*hipfb]]"
  35
  36 // emit objects for host side path
  37 // CHECK: [[CLANG:".*clang.*"]] "-cc1" "-triple" [[HOST:"x86_64-[^"]+"]]
  38 // CHECK-SAME: "-aux-triple" "amdgcn-amd-amdhsa"
  39 // CHECK-SAME: "-emit-obj"
  40 // CHECK-SAME: {{.*}} "-main-file-name" "a.cu"
  41 // CHECK-SAME: "-fgpu-rdc"
  42 // CHECK-SAME: {{.*}} "-o" [[A_OBJ_HOST:".*o"]] "-x" "hip"
  43 // CHECK-SAME: {{.*}} [[A_SRC:".*a.cu"]]
  44
  45 // CHECK: [[CLANG]] "-cc1" "-triple" [[HOST]]
  46 // CHECK-SAME: "-aux-triple" "amdgcn-amd-amdhsa"
  47 // CHECK-SAME: "-emit-obj"
  48 // CHECK-SAME: {{.*}} "-main-file-name" "b.hip"
  49 // CHECK-SAME: "-fgpu-rdc"
  50 // CHECK-SAME: {{.*}} "-o" [[B_OBJ_HOST:".*o"]] "-x" "hip"
  51 // CHECK-SAME: {{.*}} [[B_SRC:".*b.hip"]]
  52
  53 // generate image for device side path on gfx803
  54 // CHECK: [[CLANG]] "-cc1" "-triple" "amdgcn-amd-amdhsa"
  55 // CHECK-SAME: "-aux-triple" [[HOST:"x86_64-[^"]+"]]
  56 // CHECK-SAME: "-emit-llvm-bc"
  57 // CHECK-SAME: {{.*}} "-main-file-name" "a.cu"
  58 // CHECK-SAME: "-fcuda-is-device" "-fcuda-allow-variadic-functions" "-fvisibility=hidden"
  59 // CHECK-SAME: "-fapply-global-visibility-to-externs"
  60 // CHECK-SAME: "{{.*}}lib1.bc" "{{.*}}lib2.bc"
  61 // CHECK-SAME: "-target-cpu" "gfx803"
  62 // CHECK-SAME: "-fgpu-rdc"
  63 // CHECK-SAME: {{.*}} "-o" [[A_BC1:".*bc"]] "-x" "hip"
  64 // CHECK-SAME: {{.*}} [[A_SRC]]
  65
  66 // CHECK: [[CLANG]] "-cc1" "-triple" "amdgcn-amd-amdhsa"
  67 // CHECK-SAME: "-aux-triple" [[HOST]]
  68 // CHECK-SAME: "-emit-llvm-bc"
  69 // CHECK-SAME: {{.*}} "-main-file-name" "b.hip"
  70 // CHECK-SAME: "-fcuda-is-device" "-fcuda-allow-variadic-functions" "-fvisibility=hidden"
  71 // CHECK-SAME: "-fapply-global-visibility-to-externs"
  72 // CHECK-SAME: "{{.*}}lib1.bc" "{{.*}}lib2.bc"
  73 // CHECK-SAME: "-target-cpu" "gfx803"
  74 // CHECK-SAME: "-fgpu-rdc"
  75 // CHECK-SAME: {{.*}} "-o" [[B_BC1:".*bc"]] "-x" "hip"
  76 // CHECK-SAME: {{.*}} [[B_SRC]]
  77
  78 // CHECK-NOT: "*.llvm-link"
  79 // CHECK-NOT: ".*opt"
  80 // CHECK-NOT: ".*llc"
  81 // CHECK: {{".*lld.*"}} {{.*}} "-plugin-opt=-amdgpu-internalize-symbols"
  82 // CHECK-SAME: "-plugin-opt=mcpu=gfx803"
  83 // CHECK-SAME: "-o" "[[IMG_DEV1:.*.out]]" [[A_BC1]] [[B_BC1]]
  84
  85 // generate image for device side path on gfx900
  86 // CHECK: [[CLANG]] "-cc1" "-triple" "amdgcn-amd-amdhsa"
  87 // CHECK-SAME: "-aux-triple" [[HOST]]
  88 // CHECK-SAME: "-emit-llvm-bc"
  89 // CHECK-SAME: {{.*}} "-main-file-name" "a.cu"
  90 // CHECK-SAME: "-fcuda-is-device"
  91 // CHECK-SAME: "{{.*}}lib1.bc" "{{.*}}lib2.bc"
  92 // CHECK-SAME: "-target-cpu" "gfx900"
  93 // CHECK-SAME: {{.*}} "-o" [[A_BC2:".*bc"]] "-x" "hip"
  94 // CHECK-SAME: {{.*}} [[A_SRC]]
  95
  96 // CHECK: [[CLANG]] "-cc1" "-triple" "amdgcn-amd-amdhsa"
  97 // CHECK-SAME: "-aux-triple" [[HOST]]
  98 // CHECK-SAME: "-emit-llvm-bc"
  99 // CHECK-SAME: {{.*}} "-main-file-name" "b.hip"
 100 // CHECK-SAME: "-fcuda-is-device"
 101 // CHECK-SAME: "{{.*}}lib1.bc" "{{.*}}lib2.bc"
 102 // CHECK-SAME: "-target-cpu" "gfx900"
 103 // CHECK-SAME: "-fgpu-rdc"
 104 // CHECK-SAME: {{.*}} "-o" [[B_BC2:".*bc"]] "-x" "hip"
 105 // CHECK-SAME: {{.*}} [[B_SRC]]
 106
 107 // CHECK-NOT: "*.llvm-link"
 108 // CHECK-NOT: ".*opt"
 109 // CHECK-NOT: ".*llc"
 110 // CHECK: {{".*lld.*"}} {{.*}} "-plugin-opt=-amdgpu-internalize-symbols"
 111 // CHECK-SAME: "-plugin-opt=mcpu=gfx900"
 112 // CHECK-SAME: "-o" "[[IMG_DEV2:.*.out]]" [[A_BC2]] [[B_BC2]]
 113
 114 // combine images generated into hip fat binary object
 115 // CHECK: [[BUNDLER:".*clang-offload-bundler"]] "-type=o"
 116 // CHECK-SAME: "-bundle-align=4096"
 117 // CHECK-SAME: "-targets={{.*}},hipv4-amdgcn-amd-amdhsa--gfx803,hipv4-amdgcn-amd-amdhsa--gfx900"
 118 // CHECK-SAME: "-input={{.*}}" "-input=[[IMG_DEV1]]" "-input=[[IMG_DEV2]]" "-output=[[BUNDLE]]"
 119
 120 // CHECK: [[MC:".*llvm-mc.*"]] "-triple" [[HOST]] "-o" [[OBJBUNDLE:".*o"]] "{{.*}}.mcin" "--filetype=obj"
 121
 122 // output the executable
 123 // LNX: [[LD:".*ld.*"]] {{.*}}"-o" "a.out" {{.*}} [[A_OBJ_HOST]] [[B_OBJ_HOST]] [[OBJBUNDLE]]
 124 // MSVC: [[LD:".*lld-link.*"]] {{.*}}"-out:a.exe" {{.*}} [[A_OBJ_HOST]] [[B_OBJ_HOST]] [[OBJBUNDLE]]