clang/test/Driver/hip-device-libs.hip

   1 // Test if oclc_daz_opt_on or if oclc_daz_opt_off is linked depending on
   2 // expected denormal mode.
   3
   4 // Test subtarget with flushing on by default.
   5 // RUN: %clang -### --target=x86_64-linux-gnu \
   6 // RUN:  --cuda-gpu-arch=gfx803 \
   7 // RUN:  --rocm-path=%S/Inputs/rocm   \
   8 // RUN:   %S/Inputs/hip_multiple_inputs/b.hip \
   9 // RUN: 2>&1 | FileCheck %s --check-prefixes=ALL,FLUSHD,ROCMDIR
  10
  11
  12 // Test subtarget with flushing off by ddefault.
  13 // RUN: %clang -### --target=x86_64-linux-gnu \
  14 // RUN:  --cuda-gpu-arch=gfx900 \
  15 // RUN:  --rocm-path=%S/Inputs/rocm \
  16 // RUN:   %S/Inputs/hip_multiple_inputs/b.hip \
  17 // RUN: 2>&1 | FileCheck %s --check-prefixes=ALL,NOFLUSHD,ROCMDIR
  18
  19
  20 // Test explicit flag, opposite of target default.
  21 // RUN: %clang -### --target=x86_64-linux-gnu \
  22 // RUN:   --cuda-gpu-arch=gfx900 \
  23 // RUN:   -fgpu-flush-denormals-to-zero \
  24 // RUN:   --rocm-path=%S/Inputs/rocm \
  25 // RUN:   %S/Inputs/hip_multiple_inputs/b.hip \
  26 // RUN: 2>&1 | FileCheck %s --check-prefixes=ALL,FLUSHD,ROCMDIR
  27
  28
  29 // Test explicit flag, opposite of target default.
  30 // RUN: %clang -### --target=x86_64-linux-gnu \
  31 // RUN:   --cuda-gpu-arch=gfx803 \
  32 // RUN:   -fno-gpu-flush-denormals-to-zero \
  33 // RUN:   --rocm-path=%S/Inputs/rocm \
  34 // RUN:   %S/Inputs/hip_multiple_inputs/b.hip \
  35 // RUN: 2>&1 | FileCheck %s --check-prefixes=ALL,NOFLUSHD,ROCMDIR
  36
  37
  38 // Test explicit flag, same as target default.
  39 // RUN: %clang -### --target=x86_64-linux-gnu \
  40 // RUN:   --cuda-gpu-arch=gfx900 \
  41 // RUN:   -fno-gpu-flush-denormals-to-zero \
  42 // RUN:   --rocm-path=%S/Inputs/rocm \
  43 // RUN:   %S/Inputs/hip_multiple_inputs/b.hip \
  44 // RUN: 2>&1 | FileCheck %s --check-prefixes=ALL,NOFLUSHD,ROCMDIR
  45
  46
  47 // Test explicit flag, same as target default.
  48 // RUN: %clang -### --target=x86_64-linux-gnu \
  49 // RUN:   --cuda-gpu-arch=gfx803 \
  50 // RUN:   -fgpu-flush-denormals-to-zero \
  51 // RUN:   --rocm-path=%S/Inputs/rocm \
  52 // RUN:   %S/Inputs/hip_multiple_inputs/b.hip \
  53 // RUN: 2>&1 | FileCheck %s --check-prefixes=ALL,FLUSHD,ROCMDIR
  54
  55
  56 // Test last flag wins, not flushing
  57 // RUN: %clang -### --target=x86_64-linux-gnu \
  58 // RUN:   --cuda-gpu-arch=gfx803 \
  59 // RUN:   -fgpu-flush-denormals-to-zero -fno-gpu-flush-denormals-to-zero \
  60 // RUN:   --rocm-path=%S/Inputs/rocm \
  61 // RUN:   %S/Inputs/hip_multiple_inputs/b.hip \
  62 // RUN: 2>&1 | FileCheck %s --check-prefixes=ALL,NOFLUSHD,ROCMDIR
  63
  64
  65 // RUN: %clang -### --target=x86_64-linux-gnu \
  66 // RUN:   --cuda-gpu-arch=gfx900 \
  67 // RUN:   -fgpu-flush-denormals-to-zero -fno-gpu-flush-denormals-to-zero \
  68 // RUN:   --rocm-path=%S/Inputs/rocm   \
  69 // RUN:   %S/Inputs/hip_multiple_inputs/b.hip \
  70 // RUN: 2>&1 | FileCheck %s --check-prefixes=ALL,NOFLUSHD,ROCMDIR
  71
  72
  73 // RUN: %clang -### --target=x86_64-linux-gnu \
  74 // RUN:   --cuda-gpu-arch=gfx900 \
  75 // RUN:   -fno-gpu-flush-denormals-to-zero -fgpu-flush-denormals-to-zero \
  76 // RUN:   --rocm-path=%S/Inputs/rocm   \
  77 // RUN:   %S/Inputs/hip_multiple_inputs/b.hip \
  78 // RUN: 2>&1 | FileCheck %s --check-prefixes=ALL,FLUSHD,ROCMDIR
  79
  80
  81 // RUN: %clang -### --target=x86_64-linux-gnu \
  82 // RUN:   --cuda-gpu-arch=gfx803 \
  83 // RUN:   -fno-gpu-flush-denormals-to-zero -fgpu-flush-denormals-to-zero \
  84 // RUN:   --rocm-path=%S/Inputs/rocm \
  85 // RUN:   %S/Inputs/hip_multiple_inputs/b.hip \
  86 // RUN: 2>&1 | FileCheck %s --check-prefixes=ALL,FLUSHD,ROCMDIR
  87
  88 // Test finding device lib in resource dir
  89 // RUN: %clang -### --target=x86_64-linux-gnu \
  90 // RUN:   --offload-arch=gfx803 -nogpuinc \
  91 // RUN:   -resource-dir=%S/Inputs/rocm_resource_dir \
  92 // RUN:   %S/Inputs/hip_multiple_inputs/b.hip \
  93 // RUN: 2>&1 | FileCheck %s --check-prefixes=ALL,FLUSHD,RESDIR
  94
  95 // Test --hip-device-lib-path flag
  96 // RUN: %clang -### --target=x86_64-linux-gnu \
  97 // RUN:   --cuda-gpu-arch=gfx803 -nogpuinc \
  98 // RUN:   --hip-device-lib-path=%S/Inputs/rocm/amdgcn/bitcode   \
  99 // RUN:   %S/Inputs/hip_multiple_inputs/b.hip \
 100 // RUN: 2>&1 | FileCheck %s --check-prefixes=ALL,FLUSHD,ROCMDIR
 101
 102 // Test --hip-device-lib-path wins over -resource-dir
 103 // RUN: %clang -### --target=x86_64-linux-gnu \
 104 // RUN:   --cuda-gpu-arch=gfx803 -nogpuinc \
 105 // RUN:   --hip-device-lib-path=%S/Inputs/rocm/amdgcn/bitcode   \
 106 // RUN:   -resource-dir=%S/Inputs/rocm_resource_dir \
 107 // RUN:   %S/Inputs/hip_multiple_inputs/b.hip \
 108 // RUN: 2>&1 | FileCheck %s --check-prefixes=ALL,FLUSHD,ROCMDIR
 109
 110 // Test environment variable HIP_DEVICE_LIB_PATH
 111 // RUN: env HIP_DEVICE_LIB_PATH=%S/Inputs/rocm/amdgcn/bitcode \
 112 // RUN:   %clang -### --target=x86_64-linux-gnu \
 113 // RUN:   --cuda-gpu-arch=gfx900 -nogpuinc \
 114 // RUN:   %S/Inputs/hip_multiple_inputs/b.hip \
 115 // RUN: 2>&1 | FileCheck %s --check-prefixes=ALL,ROCMDIR
 116
 117 // Test environment variable HIP_DEVICE_LIB_PATH wins over -resource-dir
 118 // RUN: env HIP_DEVICE_LIB_PATH=%S/Inputs/rocm/amdgcn/bitcode \
 119 // RUN:   %clang -### --target=x86_64-linux-gnu \
 120 // RUN:   --cuda-gpu-arch=gfx900 -nogpuinc \
 121 // RUN:   -resource-dir=%S/Inputs/rocm_resource_dir \
 122 // RUN:   %S/Inputs/hip_multiple_inputs/b.hip \
 123 // RUN: 2>&1 | FileCheck %s --check-prefixes=ALL,ROCMDIR
 124
 125 // Test --gpu-instrument-lib
 126 // RUN: %clang -### --target=x86_64-linux-gnu \
 127 // RUN:   --cuda-gpu-arch=gfx900 \
 128 // RUN:   --rocm-path=%S/Inputs/rocm \
 129 // RUN:   --gpu-instrument-lib=%S/Inputs/hip_multiple_inputs/instrument.bc \
 130 // RUN:   %S/Inputs/hip_multiple_inputs/b.hip \
 131 // RUN: 2>&1 | FileCheck %s --check-prefixes=ALL,INST,ROCMDIR
 132
 133 // Test -fast-math
 134 // RUN: %clang -### --target=x86_64-linux-gnu --offload-arch=gfx900 \
 135 // RUN:   -ffast-math --rocm-path=%S/Inputs/rocm \
 136 // RUN:   %S/Inputs/hip_multiple_inputs/b.hip \
 137 // RUN: 2>&1 | FileCheck %s --check-prefixes=FAST
 138
 139 // Test -ffinite-math-only
 140 // RUN: %clang -### --target=x86_64-linux-gnu --offload-arch=gfx900 \
 141 // RUN:   -ffinite-math-only --rocm-path=%S/Inputs/rocm \
 142 // RUN:   %S/Inputs/hip_multiple_inputs/b.hip \
 143 // RUN: 2>&1 | FileCheck %s --check-prefixes=FINITE
 144
 145 // Test -funsafe-math-optimizations
 146 // RUN: %clang -### --target=x86_64-linux-gnu --offload-arch=gfx900 \
 147 // RUN:   -funsafe-math-optimizations --rocm-path=%S/Inputs/rocm \
 148 // RUN:   %S/Inputs/hip_multiple_inputs/b.hip \
 149 // RUN: 2>&1 | FileCheck %s --check-prefixes=UNSAFE
 150
 151 // Test -fno-hip-fp32-correctly-rounded-divide-sqrt
 152 // RUN: %clang -### --target=x86_64-linux-gnu --offload-arch=gfx900 \
 153 // RUN:   -fno-hip-fp32-correctly-rounded-divide-sqrt \
 154 // RUN:   --rocm-path=%S/Inputs/rocm %S/Inputs/hip_multiple_inputs/b.hip \
 155 // RUN: 2>&1 | FileCheck %s --check-prefixes=DIVSQRT
 156
 157 // Test default code object version.
 158 // RUN: %clang -### --target=x86_64-linux-gnu --offload-arch=gfx900 \
 159 // RUN:   --rocm-path=%S/Inputs/rocm %S/Inputs/hip_multiple_inputs/b.hip \
 160 // RUN: 2>&1 | FileCheck %s --check-prefixes=ABI5
 161
 162 // Test default code object version with old device library without abi_version_400.bc
 163 // RUN: %clang -### --target=x86_64-linux-gnu --offload-arch=gfx900 \
 164 // RUN:   -mcode-object-version=4 \
 165 // RUN:   --hip-device-lib-path=%S/Inputs/rocm/amdgcn/bitcode-no-abi-ver \
 166 // RUN:   --rocm-path=%S/Inputs/rocm %S/Inputs/hip_multiple_inputs/b.hip \
 167 // RUN: 2>&1 | FileCheck %s --check-prefixes=NOABI4
 168
 169 // Test -mcode-object-version=4
 170 // RUN: %clang -### --target=x86_64-linux-gnu --offload-arch=gfx900 \
 171 // RUN:   -mcode-object-version=4 \
 172 // RUN:   --rocm-path=%S/Inputs/rocm %S/Inputs/hip_multiple_inputs/b.hip \
 173 // RUN: 2>&1 | FileCheck %s --check-prefixes=ABI4
 174
 175 // Test -mcode-object-version=4 with old device library without abi_version_400.bc
 176 // RUN: %clang -### --target=x86_64-linux-gnu --offload-arch=gfx900 \
 177 // RUN:   -mcode-object-version=4 \
 178 // RUN:   --hip-device-lib-path=%S/Inputs/rocm/amdgcn/bitcode-no-abi-ver   \
 179 // RUN:   --rocm-path=%S/Inputs/rocm %S/Inputs/hip_multiple_inputs/b.hip \
 180 // RUN: 2>&1 | FileCheck %s --check-prefixes=NOABI4
 181
 182 // Test -mcode-object-version=5
 183 // RUN: %clang -### --target=x86_64-linux-gnu --offload-arch=gfx900 \
 184 // RUN:   -mcode-object-version=5 \
 185 // RUN:   --rocm-path=%S/Inputs/rocm %S/Inputs/hip_multiple_inputs/b.hip \
 186 // RUN: 2>&1 | FileCheck %s --check-prefixes=ABI5
 187
 188 // Test -mcode-object-version=5 with old device library without abi_version_500.bc
 189 // RUN: not %clang -### --target=x86_64-linux-gnu --offload-arch=gfx900 \
 190 // RUN:   -mcode-object-version=5 \
 191 // RUN:   --hip-device-lib-path=%S/Inputs/rocm/amdgcn/bitcode-no-abi-ver   \
 192 // RUN:   --rocm-path=%S/Inputs/rocm %S/Inputs/hip_multiple_inputs/b.hip \
 193 // RUN: 2>&1 | FileCheck %s --check-prefixes=NOABI5
 194
 195 // Test -mcode-object-version=6
 196 // RUN: %clang -### --target=x86_64-linux-gnu --offload-arch=gfx900 \
 197 // RUN:   -mcode-object-version=6 \
 198 // RUN:   --rocm-path=%S/Inputs/rocm %S/Inputs/hip_multiple_inputs/b.hip \
 199 // RUN: 2>&1 | FileCheck %s --check-prefixes=ABI6
 200
 201 // Test -mcode-object-version=6 with old device library without abi_version_600.bc
 202 // RUN: not %clang -### --target=x86_64-linux-gnu --offload-arch=gfx900 \
 203 // RUN:   -mcode-object-version=6 \
 204 // RUN:   --hip-device-lib-path=%S/Inputs/rocm/amdgcn/bitcode-no-abi-ver   \
 205 // RUN:   --rocm-path=%S/Inputs/rocm %S/Inputs/hip_multiple_inputs/b.hip \
 206 // RUN: 2>&1 | FileCheck %s --check-prefixes=NOABI6
 207
 208 // ALL-NOT: error:
 209 // ALL: {{"[^"]*clang[^"]*"}}
 210
 211 // RESDIR-SAME: "-mlink-builtin-bitcode" "[[DEVICELIB_DIR:[^"]+(/|\\\\)rocm_resource_dir(/|\\\\)lib(64)?(/|\\\\)amdgcn(/|\\\\).*]]hip.bc"
 212 // ROCMDIR-SAME: "-mlink-builtin-bitcode" "[[DEVICELIB_DIR:[^"]+(/|\\\\)rocm(/|\\\\)amdgcn(/|\\\\).*]]hip.bc"
 213
 214 // ALL-SAME: "-mlink-builtin-bitcode" "[[DEVICELIB_DIR]]ocml.bc"
 215 // ALL-SAME: "-mlink-builtin-bitcode" "[[DEVICELIB_DIR]]ockl.bc"
 216
 217 // FLUSHD-SAME: "-mlink-builtin-bitcode" "[[DEVICELIB_DIR]]oclc_daz_opt_on.bc"
 218 // NOFLUSHD-SAME: "-mlink-builtin-bitcode" "[[DEVICELIB_DIR]]oclc_daz_opt_off.bc"
 219
 220 // ALL-SAME: "-mlink-builtin-bitcode" "[[DEVICELIB_DIR]]oclc_unsafe_math_off.bc"
 221 // ALL-SAME: "-mlink-builtin-bitcode" "[[DEVICELIB_DIR]]oclc_finite_only_off.bc"
 222 // ALL-SAME: "-mlink-builtin-bitcode" "[[DEVICELIB_DIR]]oclc_correctly_rounded_sqrt_on.bc"
 223 // ALL-SAME: "-mlink-builtin-bitcode" "[[DEVICELIB_DIR]]oclc_wavefrontsize64_on.bc"
 224 // ALL-SAME: "-mlink-builtin-bitcode" "[[DEVICELIB_DIR]]oclc_isa_version_{{[0-9]+}}.bc"
 225 // INST-SAME: "-mlink-builtin-bitcode" "{{.*}}instrument.bc"
 226
 227 // FAST: "-mlink-builtin-bitcode" "{{.*}}oclc_daz_opt_off.bc"
 228 // FAST-SAME: "-mlink-builtin-bitcode" "{{.*}}oclc_unsafe_math_on.bc"
 229 // FAST-SAME: "-mlink-builtin-bitcode" "{{.*}}oclc_finite_only_on.bc"
 230 // FAST-SAME: "-mlink-builtin-bitcode" "{{.*}}oclc_correctly_rounded_sqrt_on.bc"
 231
 232 // FINITE: "-mlink-builtin-bitcode" "{{.*}}oclc_daz_opt_off.bc"
 233 // FINITE-SAME: "-mlink-builtin-bitcode" "{{.*}}oclc_unsafe_math_off.bc"
 234 // FINITE-SAME: "-mlink-builtin-bitcode" "{{.*}}oclc_finite_only_on.bc"
 235 // FINITE-SAME: "-mlink-builtin-bitcode" "{{.*}}oclc_correctly_rounded_sqrt_on.bc"
 236
 237 // UNSAFE: "-mlink-builtin-bitcode" "{{.*}}oclc_daz_opt_off.bc"
 238 // UNSAFE-SAME: "-mlink-builtin-bitcode" "{{.*}}oclc_unsafe_math_on.bc"
 239 // UNSAFE-SAME: "-mlink-builtin-bitcode" "{{.*}}oclc_finite_only_off.bc"
 240 // UNSAFE-SAME: "-mlink-builtin-bitcode" "{{.*}}oclc_correctly_rounded_sqrt_on.bc"
 241
 242 // DIVSQRT: "-mlink-builtin-bitcode" "{{.*}}oclc_daz_opt_off.bc"
 243 // DIVSQRT-SAME: "-mlink-builtin-bitcode" "{{.*}}oclc_unsafe_math_off.bc"
 244 // DIVSQRT-SAME: "-mlink-builtin-bitcode" "{{.*}}oclc_finite_only_off.bc"
 245 // DIVSQRT-SAME: "-mlink-builtin-bitcode" "{{.*}}oclc_correctly_rounded_sqrt_off.bc"
 246
 247 // ABI4-NOT: error:
 248 // ABI4: "-mlink-builtin-bitcode" "{{.*}}oclc_abi_version_400.bc"
 249 // ABI5-NOT: error:
 250 // ABI5: "-mlink-builtin-bitcode" "{{.*}}oclc_abi_version_500.bc"
 251 // ABI6-NOT: error:
 252 // ABI6: "-mlink-builtin-bitcode" "{{.*}}oclc_abi_version_600.bc"
 253 // NOABI4-NOT: error:
 254 // NOABI4-NOT: "-mlink-builtin-bitcode" "{{.*}}oclc_abi_version_400.bc"
 255 // NOABI4-NOT: "-mlink-builtin-bitcode" "{{.*}}oclc_abi_version_500.bc"
 256 // NOABI5: error: cannot find ROCm device libraryfor ABI version 5; provide its path via '--rocm-path' or '--rocm-device-lib-path', or pass '-nogpulib' to build without ROCm device library
 257 // NOABI6: error: cannot find ROCm device libraryfor ABI version 6; provide its path via '--rocm-path' or '--rocm-device-lib-path', or pass '-nogpulib' to build without ROCm device library