test/MC/AMDGPU/hsa.s

   1 // RUN: llvm-mc -triple amdgcn--amdhsa -mcpu=kaveri -mattr=-code-object-v3 -show-encoding %s | FileCheck %s --check-prefix=ASM
   2 // RUN: llvm-mc -filetype=obj -triple amdgcn--amdhsa -mcpu=kaveri -mattr=-code-object-v3 -show-encoding %s | llvm-readobj --symbols -S --sd | FileCheck %s --check-prefix=ELF
   3
   4 // ELF: Section {
   5 // ELF: Name: .text
   6 // ELF: Type: SHT_PROGBITS (0x1)
   7 // ELF: Flags [ (0x6)
   8 // ELF: SHF_ALLOC (0x2)
   9 // ELF: SHF_EXECINSTR (0x4)
  10
  11 // ELF: SHT_NOTE
  12 // ELF: 0000: 04000000 08000000 01000000 414D4400
  13 // ELF: 0010: 02000000 00000000 04000000 1B000000
  14 // ELF: 0020: 03000000 414D4400 04000700 07000000
  15 // ELF: 0030: 00000000 00000000 414D4400 414D4447
  16 // ELF: 0040: 50550000
  17 // We can't check binary representation of metadata note: it is different on
  18 // Windows and Linux because of carriage return on Windows
  19
  20 // ELF: Symbol {
  21 // ELF: Name: amd_kernel_code_t_minimal
  22 // ELF: Type: AMDGPU_HSA_KERNEL (0xA)
  23 // ELF: Section: .text
  24 // ELF: }
  25 // ELF: Symbol {
  26 // ELF: Name: amd_kernel_code_t_test_all
  27 // ELF: Type: AMDGPU_HSA_KERNEL (0xA)
  28 // ELF: Section: .text
  29 // ELF: }
  30
  31 .text
  32 // ASM: .text
  33
  34 .hsa_code_object_version 2,0
  35 // ASM: .hsa_code_object_version 2,0
  36
  37 .hsa_code_object_isa 7,0,0,"AMD","AMDGPU"
  38 // ASM: .hsa_code_object_isa 7,0,0,"AMD","AMDGPU"
  39
  40 .amd_amdgpu_hsa_metadata
  41   Version: [ 3, 0 ]
  42   Kernels:
  43     - Name:       amd_kernel_code_t_test_all
  44       SymbolName: amd_kernel_code_t_test_all@kd
  45     - Name:       amd_kernel_code_t_minimal
  46       SymbolName: amd_kernel_code_t_minimal@kd
  47 .end_amd_amdgpu_hsa_metadata
  48
  49 // ASM: .amd_amdgpu_hsa_metadata
  50 // ASM:    Version: [ 3, 0 ]
  51 // ASM:    Kernels:
  52 // ASM:      - Name:       amd_kernel_code_t_test_all
  53 // ASM:        SymbolName: 'amd_kernel_code_t_test_all@kd'
  54 // ASM:      - Name:       amd_kernel_code_t_minimal
  55 // ASM:        SymbolName: 'amd_kernel_code_t_minimal@kd'
  56 // ASM: .end_amd_amdgpu_hsa_metadata
  57
  58 .amdgpu_hsa_kernel amd_kernel_code_t_test_all
  59 .amdgpu_hsa_kernel amd_kernel_code_t_minimal
  60
  61 amd_kernel_code_t_test_all:
  62 ; Test all amd_kernel_code_t members with non-default values.
  63 .amd_kernel_code_t
  64     kernel_code_version_major = 100
  65     kernel_code_version_minor = 100
  66     machine_kind = 0
  67     machine_version_major = 5
  68     machine_version_minor = 5
  69     machine_version_stepping = 5
  70     kernel_code_entry_byte_offset = 512
  71     kernel_code_prefetch_byte_size = 1
  72     max_scratch_backing_memory_byte_size = 1
  73     compute_pgm_rsrc1_vgprs = 1
  74     compute_pgm_rsrc1_sgprs = 1
  75     compute_pgm_rsrc1_priority = 1
  76     compute_pgm_rsrc1_float_mode = 1
  77     compute_pgm_rsrc1_priv = 1
  78     compute_pgm_rsrc1_dx10_clamp = 1
  79     compute_pgm_rsrc1_debug_mode = 1
  80     compute_pgm_rsrc1_ieee_mode = 1
  81     compute_pgm_rsrc2_scratch_en = 1
  82     compute_pgm_rsrc2_user_sgpr = 1
  83     compute_pgm_rsrc2_tgid_x_en = 1
  84     compute_pgm_rsrc2_tgid_y_en = 1
  85     compute_pgm_rsrc2_tgid_z_en = 1
  86     compute_pgm_rsrc2_tg_size_en = 1
  87     compute_pgm_rsrc2_tidig_comp_cnt = 1
  88     compute_pgm_rsrc2_excp_en_msb = 1
  89     compute_pgm_rsrc2_lds_size = 1
  90     compute_pgm_rsrc2_excp_en = 1
  91     enable_sgpr_private_segment_buffer = 1
  92     enable_sgpr_dispatch_ptr = 1
  93     enable_sgpr_queue_ptr = 1
  94     enable_sgpr_kernarg_segment_ptr = 1
  95     enable_sgpr_dispatch_id = 1
  96     enable_sgpr_flat_scratch_init = 1
  97     enable_sgpr_private_segment_size = 1
  98     enable_sgpr_grid_workgroup_count_x = 1
  99     enable_sgpr_grid_workgroup_count_y = 1
 100     enable_sgpr_grid_workgroup_count_z = 1
 101     enable_ordered_append_gds = 1
 102     private_element_size = 1
 103     is_ptr64 = 1
 104     is_dynamic_callstack = 1
 105     is_debug_enabled = 1
 106     is_xnack_enabled = 1
 107     workitem_private_segment_byte_size = 1
 108     workgroup_group_segment_byte_size = 1
 109     gds_segment_byte_size = 1
 110     kernarg_segment_byte_size = 1
 111     workgroup_fbarrier_count = 1
 112     wavefront_sgpr_count = 1
 113     workitem_vgpr_count = 1
 114     reserved_vgpr_first = 1
 115     reserved_vgpr_count = 1
 116     reserved_sgpr_first = 1
 117     reserved_sgpr_count = 1
 118     debug_wavefront_private_segment_offset_sgpr = 1
 119     debug_private_segment_buffer_sgpr = 1
 120     kernarg_segment_alignment = 5
 121     group_segment_alignment = 5
 122     private_segment_alignment = 5
 123     wavefront_size = 6
 124     call_convention = 1
 125     runtime_loader_kernel_symbol = 1
 126 .end_amd_kernel_code_t
 127
 128 // ASM-LABEL: {{^}}amd_kernel_code_t_test_all:
 129 // ASM: .amd_kernel_code_t
 130 // ASM: amd_code_version_major = 100
 131 // ASM: amd_code_version_minor = 100
 132 // ASM: amd_machine_kind = 0
 133 // ASM: amd_machine_version_major = 5
 134 // ASM: amd_machine_version_minor = 5
 135 // ASM: amd_machine_version_stepping = 5
 136 // ASM: kernel_code_entry_byte_offset = 512
 137 // ASM: kernel_code_prefetch_byte_size = 1
 138 // ASM: granulated_workitem_vgpr_count = 1
 139 // ASM: granulated_wavefront_sgpr_count = 1
 140 // ASM: priority = 1
 141 // ASM: float_mode = 1
 142 // ASM: priv = 1
 143 // ASM: enable_dx10_clamp = 1
 144 // ASM: debug_mode = 1
 145 // ASM: enable_ieee_mode = 1
 146 // ASM: enable_sgpr_private_segment_wave_byte_offset = 1
 147 // ASM: user_sgpr_count = 1
 148 // ASM: enable_sgpr_workgroup_id_x = 1
 149 // ASM: enable_sgpr_workgroup_id_y = 1
 150 // ASM: enable_sgpr_workgroup_id_z = 1
 151 // ASM: enable_sgpr_workgroup_info = 1
 152 // ASM: enable_vgpr_workitem_id = 1
 153 // ASM: enable_exception_msb = 1
 154 // ASM: granulated_lds_size = 1
 155 // ASM: enable_exception = 1
 156 // ASM: enable_sgpr_private_segment_buffer = 1
 157 // ASM: enable_sgpr_dispatch_ptr = 1
 158 // ASM: enable_sgpr_queue_ptr = 1
 159 // ASM: enable_sgpr_kernarg_segment_ptr = 1
 160 // ASM: enable_sgpr_dispatch_id = 1
 161 // ASM: enable_sgpr_flat_scratch_init = 1
 162 // ASM: enable_sgpr_private_segment_size = 1
 163 // ASM: enable_sgpr_grid_workgroup_count_x = 1
 164 // ASM: enable_sgpr_grid_workgroup_count_y = 1
 165 // ASM: enable_sgpr_grid_workgroup_count_z = 1
 166 // ASM: enable_ordered_append_gds = 1
 167 // ASM: private_element_size = 1
 168 // ASM: is_ptr64 = 1
 169 // ASM: is_dynamic_callstack = 1
 170 // ASM: is_debug_enabled = 1
 171 // ASM: is_xnack_enabled = 1
 172 // ASM: workitem_private_segment_byte_size = 1
 173 // ASM: workgroup_group_segment_byte_size = 1
 174 // ASM: gds_segment_byte_size = 1
 175 // ASM: kernarg_segment_byte_size = 1
 176 // ASM: workgroup_fbarrier_count = 1
 177 // ASM: wavefront_sgpr_count = 1
 178 // ASM: workitem_vgpr_count = 1
 179 // ASM: reserved_vgpr_first = 1
 180 // ASM: reserved_vgpr_count = 1
 181 // ASM: reserved_sgpr_first = 1
 182 // ASM: reserved_sgpr_count = 1
 183 // ASM: debug_wavefront_private_segment_offset_sgpr = 1
 184 // ASM: debug_private_segment_buffer_sgpr = 1
 185 // ASM: kernarg_segment_alignment = 5
 186 // ASM: group_segment_alignment = 5
 187 // ASM: private_segment_alignment = 5
 188 // ASM: wavefront_size = 6
 189 // ASM: call_convention = 1
 190 // ASM: runtime_loader_kernel_symbol = 1
 191 // ASM: .end_amd_kernel_code_t
 192
 193 amd_kernel_code_t_minimal:
 194 .amd_kernel_code_t
 195         enable_sgpr_kernarg_segment_ptr = 1
 196         is_ptr64 = 1
 197         granulated_workitem_vgpr_count = 1
 198         granulated_wavefront_sgpr_count = 1
 199         user_sgpr_count = 2
 200         kernarg_segment_byte_size = 16
 201         wavefront_sgpr_count = 8
 202 //      wavefront_sgpr_count = 7
 203 ;       wavefront_sgpr_count = 7
 204 // Make sure a blank line won't break anything:
 205
 206 // Make sure a line with whitespace won't break anything:
 207
 208         workitem_vgpr_count = 16
 209 .end_amd_kernel_code_t
 210
 211 // ASM-LABEL: {{^}}amd_kernel_code_t_minimal:
 212 // ASM: .amd_kernel_code_t
 213 // ASM: amd_code_version_major = 1
 214 // ASM: amd_code_version_minor = 2
 215 // ASM: amd_machine_kind = 1
 216 // ASM: amd_machine_version_major = 7
 217 // ASM: amd_machine_version_minor = 0
 218 // ASM: amd_machine_version_stepping = 0
 219 // ASM: kernel_code_entry_byte_offset = 256
 220 // ASM: kernel_code_prefetch_byte_size = 0
 221 // ASM: granulated_workitem_vgpr_count = 1
 222 // ASM: granulated_wavefront_sgpr_count = 1
 223 // ASM: priority = 0
 224 // ASM: float_mode = 0
 225 // ASM: priv = 0
 226 // ASM: enable_dx10_clamp = 0
 227 // ASM: debug_mode = 0
 228 // ASM: enable_ieee_mode = 0
 229 // ASM: enable_sgpr_private_segment_wave_byte_offset = 0
 230 // ASM: user_sgpr_count = 2
 231 // ASM: enable_sgpr_workgroup_id_x = 0
 232 // ASM: enable_sgpr_workgroup_id_y = 0
 233 // ASM: enable_sgpr_workgroup_id_z = 0
 234 // ASM: enable_sgpr_workgroup_info = 0
 235 // ASM: enable_vgpr_workitem_id = 0
 236 // ASM: enable_exception_msb = 0
 237 // ASM: granulated_lds_size = 0
 238 // ASM: enable_exception = 0
 239 // ASM: enable_sgpr_private_segment_buffer = 0
 240 // ASM: enable_sgpr_dispatch_ptr = 0
 241 // ASM: enable_sgpr_queue_ptr = 0
 242 // ASM: enable_sgpr_kernarg_segment_ptr = 1
 243 // ASM: enable_sgpr_dispatch_id = 0
 244 // ASM: enable_sgpr_flat_scratch_init = 0
 245 // ASM: enable_sgpr_private_segment_size = 0
 246 // ASM: enable_sgpr_grid_workgroup_count_x = 0
 247 // ASM: enable_sgpr_grid_workgroup_count_y = 0
 248 // ASM: enable_sgpr_grid_workgroup_count_z = 0
 249 // ASM: enable_ordered_append_gds = 0
 250 // ASM: private_element_size = 0
 251 // ASM: is_ptr64 = 1
 252 // ASM: is_dynamic_callstack = 0
 253 // ASM: is_debug_enabled = 0
 254 // ASM: is_xnack_enabled = 0
 255 // ASM: workitem_private_segment_byte_size = 0
 256 // ASM: workgroup_group_segment_byte_size = 0
 257 // ASM: gds_segment_byte_size = 0
 258 // ASM: kernarg_segment_byte_size = 16
 259 // ASM: workgroup_fbarrier_count = 0
 260 // ASM: wavefront_sgpr_count = 8
 261 // ASM: workitem_vgpr_count = 16
 262 // ASM: reserved_vgpr_first = 0
 263 // ASM: reserved_vgpr_count = 0
 264 // ASM: reserved_sgpr_first = 0
 265 // ASM: reserved_sgpr_count = 0
 266 // ASM: debug_wavefront_private_segment_offset_sgpr = 0
 267 // ASM: debug_private_segment_buffer_sgpr = 0
 268 // ASM: kernarg_segment_alignment = 4
 269 // ASM: group_segment_alignment = 4
 270 // ASM: private_segment_alignment = 4
 271 // ASM: wavefront_size = 6
 272 // ASM: call_convention = -1
 273 // ASM: runtime_loader_kernel_symbol = 0
 274 // ASM: .end_amd_kernel_code_t