clang/test/Driver/rocm-device-libs.cl

   1 // REQUIRES: !system-windows
   2
   3 // Test flush-denormals-to-zero enabled uses oclc_daz_opt_on
   4
   5 // RUN: %clang -### -target amdgcn-amd-amdhsa \
   6 // RUN:   -x cl -mcpu=gfx900 \
   7 // RUN:   --rocm-path=%S/Inputs/rocm \
   8 // RUN:   %s \
   9 // RUN: 2>&1 | FileCheck  --check-prefixes=COMMON,COMMON-DEFAULT,GFX900-DEFAULT,GFX900,WAVE64 %s
  10
  11
  12
  13 // Make sure the different denormal default is respected for gfx8
  14 // RUN: %clang -### -target amdgcn-amd-amdhsa \
  15 // RUN:   -x cl -mcpu=gfx803 \
  16 // RUN:   --rocm-path=%S/Inputs/rocm \
  17 // RUN:   %s \
  18 // RUN: 2>&1 | FileCheck  --check-prefixes=COMMON,COMMON-DEFAULT,GFX803-DEFAULT,GFX803,WAVE64 %s
  19
  20
  21
  22 // Make sure the non-canonical name works
  23 // RUN: %clang -### -target amdgcn-amd-amdhsa \
  24 // RUN:   -x cl -mcpu=fiji \
  25 // RUN:   --rocm-path=%S/Inputs/rocm \
  26 // RUN:   %s \
  27 // RUN: 2>&1 | FileCheck  --check-prefixes=COMMON,COMMON-DEFAULT,GFX803-DEFAULT,GFX803,WAVE64 %s
  28
  29
  30
  31 // RUN: %clang -### -target amdgcn-amd-amdhsa \
  32 // RUN:   -x cl -mcpu=gfx900 \
  33 // RUN:   -cl-denorms-are-zero \
  34 // RUN:   --rocm-path=%S/Inputs/rocm \
  35 // RUN:   %s \
  36 // RUN: 2>&1 | FileCheck  --check-prefixes=COMMON,COMMON-DAZ,GFX900,WAVE64 %s
  37
  38
  39 // RUN: %clang -### -target amdgcn-amd-amdhsa \
  40 // RUN:   -x cl -mcpu=gfx803 \
  41 // RUN:   -cl-denorms-are-zero \
  42 // RUN:   --rocm-path=%S/Inputs/rocm \
  43 // RUN:   %s \
  44 // RUN: 2>&1 | FileCheck  --check-prefixes=COMMON,COMMON-DAZ,GFX803,WAVE64 %s
  45
  46
  47
  48 // RUN: %clang -### -target amdgcn-amd-amdhsa \
  49 // RUN:   -x cl -mcpu=gfx803 \
  50 // RUN:   -cl-finite-math-only \
  51 // RUN:   --rocm-path=%S/Inputs/rocm \
  52 // RUN:   %s \
  53 // RUN: 2>&1 | FileCheck  --check-prefixes=COMMON,COMMON-FINITE-ONLY,GFX803,WAVE64 %s
  54
  55
  56
  57 // RUN: %clang -### -target amdgcn-amd-amdhsa    \
  58 // RUN:   -x cl -mcpu=gfx803                     \
  59 // RUN:   -cl-fp32-correctly-rounded-divide-sqrt \
  60 // RUN:   --rocm-path=%S/Inputs/rocm \
  61 // RUN:   %s \
  62 // RUN: 2>&1 | FileCheck  --check-prefixes=COMMON,COMMON-CORRECT-SQRT,GFX803,WAVE64 %s
  63
  64
  65
  66 // RUN: %clang -### -target amdgcn-amd-amdhsa    \
  67 // RUN:   -x cl -mcpu=gfx803                     \
  68 // RUN:   -cl-fast-relaxed-math \
  69 // RUN:   --rocm-path=%S/Inputs/rocm \
  70 // RUN:   %s \
  71 // RUN: 2>&1 | FileCheck  --check-prefixes=COMMON,COMMON-FAST-RELAXED,GFX803,WAVE64 %s
  72
  73
  74
  75 // RUN: %clang -### -target amdgcn-amd-amdhsa    \
  76 // RUN:   -x cl -mcpu=gfx803                     \
  77 // RUN:   -cl-unsafe-math-optimizations \
  78 // RUN:   --rocm-path=%S/Inputs/rocm \
  79 // RUN:   %s 2>&1 | \
  80 // RUN:   FileCheck  --check-prefixes=COMMON,GFX803,WAVE64 %s
  81
  82 // RUN: %clang -### -target amdgcn-amd-amdhsa    \
  83 // RUN:   -x cl -mcpu=gfx1010                    \
  84 // RUN:   --rocm-path=%S/Inputs/rocm \
  85 // RUN:   %s 2>&1 | \
  86 // RUN:   FileCheck  --check-prefix=WAVE32 %s
  87
  88 // RUN: %clang -### -target amdgcn-amd-amdhsa    \
  89 // RUN:   -x cl -mcpu=gfx1011                    \
  90 // RUN:   --rocm-path=%S/Inputs/rocm \
  91 // RUN:   %s 2>&1 | \
  92 // RUN:   FileCheck  --check-prefix=WAVE32 %s
  93
  94 // RUN: %clang -### -target amdgcn-amd-amdhsa    \
  95 // RUN:   -x cl -mcpu=gfx1012                    \
  96 // RUN:   --rocm-path=%S/Inputs/rocm \
  97 // RUN:   %s 2>&1 | \
  98 // RUN:   FileCheck  --check-prefix=WAVE32 %s
  99
 100
 101 // RUN: %clang -### -target amdgcn-amd-amdhsa    \
 102 // RUN:   -x cl -mcpu=gfx1010 -mwavefrontsize64  \
 103 // RUN:   --rocm-path=%S/Inputs/rocm \
 104 // RUN:   %s 2>&1 | \
 105 // RUN:   FileCheck  --check-prefix=WAVE64 %s
 106
 107 // RUN: %clang -### -target amdgcn-amd-amdhsa    \
 108 // RUN:   -x cl -mcpu=gfx1010 -mwavefrontsize64 -mno-wavefrontsize64  \
 109 // RUN:   --rocm-path=%S/Inputs/rocm \
 110 // RUN:   %s 2>&1 | \
 111 // RUN:   FileCheck  --check-prefix=WAVE32 %s
 112
 113 // Ignore -mno-wavefrontsize64 without wave32 support
 114 // RUN: %clang -### -target amdgcn-amd-amdhsa       \
 115 // RUN:   -x cl -mcpu=gfx803  -mno-wavefrontsize64  \
 116 // RUN:   --rocm-path=%S/Inputs/rocm    \
 117 // RUN:   %s 2>&1 | \
 118 // RUN:   FileCheck  --check-prefixes=GFX803,WAVE64 %s
 119
 120
 121
 122 // Test --hip-device-lib-path format
 123 // RUN: %clang -### -target amdgcn-amd-amdhsa \
 124 // RUN:   -x cl -mcpu=gfx900 \
 125 // RUN:   --hip-device-lib-path=%S/Inputs/rocm/amdgcn/bitcode \
 126 // RUN:   %S/opencl.cl \
 127 // RUN: 2>&1 | FileCheck  --check-prefixes=COMMON,COMMON-DEFAULT,GFX900-DEFAULT,GFX900,WAVE64 %s
 128
 129 // Test environment variable HIP_DEVICE_LIB_PATH
 130 // RUN: env HIP_DEVICE_LIB_PATH=%S/Inputs/rocm/amdgcn/bitcode %clang -### -target amdgcn-amd-amdhsa \
 131 // RUN:   -x cl -mcpu=gfx900 \
 132 // RUN:   %S/opencl.cl \
 133 // RUN: 2>&1 | FileCheck  --check-prefixes=COMMON,COMMON-DEFAULT,GFX900-DEFAULT,GFX900,WAVE64 %s
 134
 135 // RUN: %clang -### -target amdgcn-amd-amdhsa \
 136 // RUN:   -x cl -mcpu=gfx908:xnack+ -fsanitize=address \
 137 // RUN:   --rocm-path=%S/Inputs/rocm \
 138 // RUN:   %s \
 139 // RUN: 2>&1 | FileCheck  --check-prefixes=ASAN,COMMON %s
 140
 141 // RUN: %clang -### -target amdgcn-amd-amdhsa \
 142 // RUN:   -x cl -mcpu=gfx908:xnack+ \
 143 // RUN:   --rocm-path=%S/Inputs/rocm \
 144 // RUN:   %s \
 145 // RUN: 2>&1 | FileCheck  --check-prefixes=NOASAN %s
 146
 147 // COMMON: "-triple" "amdgcn-amd-amdhsa"
 148 // ASAN-SAME: "-mlink-bitcode-file" "{{.*}}/amdgcn/bitcode/asanrtl.bc"
 149 // COMMON-SAME: "-mlink-builtin-bitcode" "{{.*}}/amdgcn/bitcode/opencl.bc"
 150 // COMMON-SAME: "-mlink-builtin-bitcode" "{{.*}}/amdgcn/bitcode/ocml.bc"
 151 // COMMON-SAME: "-mlink-builtin-bitcode" "{{.*}}/amdgcn/bitcode/ockl.bc"
 152
 153 // GFX900-DEFAULT-SAME: "-mlink-builtin-bitcode" "{{.*}}/amdgcn/bitcode/oclc_daz_opt_off.bc"
 154 // GFX803-DEFAULT-SAME: "-mlink-builtin-bitcode" "{{.*}}/amdgcn/bitcode/oclc_daz_opt_on.bc"
 155 // GFX700-DEFAULT-SAME: "-mlink-builtin-bitcode" "{{.*}}/amdgcn/bitcode/oclc_daz_opt_on.bc"
 156 // COMMON-DAZ-SAME: "-mlink-builtin-bitcode" "{{.*}}/amdgcn/bitcode/oclc_daz_opt_on.bc"
 157
 158
 159 // COMMON-DEFAULT-SAME: "-mlink-builtin-bitcode" "{{.*}}/amdgcn/bitcode/oclc_unsafe_math_off.bc"
 160 // COMMON-DEFAULT-SAME: "-mlink-builtin-bitcode" "{{.*}}/amdgcn/bitcode/oclc_finite_only_off.bc"
 161 // COMMON-DEFAULT-SAME: "-mlink-builtin-bitcode" "{{.*}}/amdgcn/bitcode/oclc_correctly_rounded_sqrt_off.bc"
 162
 163
 164 // COMMON-FINITE-ONLY-SAME: "-mlink-builtin-bitcode" "{{.*}}/amdgcn/bitcode/oclc_unsafe_math_off.bc"
 165 // COMMON-FINITE-ONLY-SAME: "-mlink-builtin-bitcode" "{{.*}}/amdgcn/bitcode/oclc_finite_only_on.bc"
 166 // COMMON-FINITE-ONLY-SAME: "-mlink-builtin-bitcode" "{{.*}}/amdgcn/bitcode/oclc_correctly_rounded_sqrt_off.bc"
 167
 168
 169 // COMMON-CORRECT-SQRT-SAME: "-mlink-builtin-bitcode" "{{.*}}/amdgcn/bitcode/oclc_unsafe_math_off.bc"
 170 // COMMON-CORRECT-SQRT-SAME: "-mlink-builtin-bitcode" "{{.*}}/amdgcn/bitcode/oclc_finite_only_off.bc"
 171 // COMMON-CORRECT-SQRT-SAME: "-mlink-builtin-bitcode" "{{.*}}/amdgcn/bitcode/oclc_correctly_rounded_sqrt_on.bc"
 172
 173
 174 // COMMON-FAST-RELAXED-SAME: "-mlink-builtin-bitcode" "{{.*}}/amdgcn/bitcode/oclc_unsafe_math_on.bc"
 175 // COMMON-FAST-RELAXED-SAME: "-mlink-builtin-bitcode" "{{.*}}/amdgcn/bitcode/oclc_finite_only_on.bc"
 176 // COMMON-FAST-RELAXED-SAME: "-mlink-builtin-bitcode" "{{.*}}/amdgcn/bitcode/oclc_correctly_rounded_sqrt_off.bc"
 177
 178
 179 // COMMON-UNSAFE-MATH-SAME: "-mlink-builtin-bitcode" "{{.*}}/amdgcn/bitcode/oclc_unsafe_math_on.bc"
 180 // COMMON-UNSAFE-MATH-SAME: "-mlink-builtin-bitcode" "{{.*}}/amdgcn/bitcode/oclc_finite_only_off.bc"
 181 // COMMON-UNSAFE-MATH-SAME: "-mlink-builtin-bitcode" "{{.*}}/amdgcn/bitcode/oclc_correctly_rounded_sqrt_off.bc"
 182
 183 // ASAN-SAME: "-fsanitize=address"
 184
 185 // NOASAN-NOT: "-fsanitize=address"
 186 // NOASAN-NOT: amdgcn/bitcode/asanrtl.bc
 187
 188 // WAVE64: "-mlink-builtin-bitcode" "{{.*}}/amdgcn/bitcode/oclc_wavefrontsize64_on.bc"
 189 // WAVE32: "-mlink-builtin-bitcode" "{{.*}}/amdgcn/bitcode/oclc_wavefrontsize64_off.bc"
 190
 191 // GFX900: "-mlink-builtin-bitcode" "{{.*}}/amdgcn/bitcode/oclc_isa_version_900.bc"
 192 // GFX803: "-mlink-builtin-bitcode" "{{.*}}/amdgcn/bitcode/oclc_isa_version_803.bc"
 193
 194 kernel void func(void);