1 //===-- AMDGPUKernelCodeT.h - Print AMDGPU assembly code ---------*- C++ -*-===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 /// \file AMDKernelCodeT.h 10 //===----------------------------------------------------------------------===// 11 12 #ifndef AMDKERNELCODET_H 13 #define AMDKERNELCODET_H 14 15 #include "llvm/MC/SubtargetFeature.h" 16 17 #include <cstddef> 18 #include <cstdint> 19 20 #include "llvm/Support/Debug.h" 21 //---------------------------------------------------------------------------// 22 // AMD Kernel Code, and its dependencies // 23 //---------------------------------------------------------------------------// 24 25 typedef uint8_t hsa_powertwo8_t; 26 typedef uint32_t hsa_ext_code_kind_t; 27 typedef uint8_t hsa_ext_brig_profile8_t; 28 typedef uint8_t hsa_ext_brig_machine_model8_t; 29 typedef uint64_t hsa_ext_control_directive_present64_t; 30 typedef uint16_t hsa_ext_exception_kind16_t; 31 typedef uint32_t hsa_ext_code_kind32_t; 32 33 typedef struct hsa_dim3_s { 34 uint32_t x; 35 uint32_t y; 36 uint32_t z; 37 } hsa_dim3_t; 38 39 /// The version of the amd_*_code_t struct. Minor versions must be 40 /// backward compatible. 41 typedef uint32_t amd_code_version32_t; 42 enum amd_code_version_t { 43 AMD_CODE_VERSION_MAJOR = 0, 44 AMD_CODE_VERSION_MINOR = 1 45 }; 46 47 /// The values used to define the number of bytes to use for the 48 /// swizzle element size. 49 enum amd_element_byte_size_t { 50 AMD_ELEMENT_2_BYTES = 0, 51 AMD_ELEMENT_4_BYTES = 1, 52 AMD_ELEMENT_8_BYTES = 2, 53 AMD_ELEMENT_16_BYTES = 3 54 }; 55 56 /// Shader program settings for CS. Contains COMPUTE_PGM_RSRC1 and 57 /// COMPUTE_PGM_RSRC2 registers. 58 typedef uint64_t amd_compute_pgm_resource_register64_t; 59 60 /// Every amd_*_code_t has the following properties, which are composed of 61 /// a number of bit fields. Every bit field has a mask (AMD_CODE_PROPERTY_*), 62 /// bit width (AMD_CODE_PROPERTY_*_WIDTH, and bit shift amount 63 /// (AMD_CODE_PROPERTY_*_SHIFT) for convenient access. Unused bits must be 0. 64 /// 65 /// (Note that bit fields cannot be used as their layout is 66 /// implementation defined in the C standard and so cannot be used to 67 /// specify an ABI) 68 typedef uint32_t amd_code_property32_t; 69 enum amd_code_property_mask_t { 70 71 /// Enable the setup of the SGPR user data registers 72 /// (AMD_CODE_PROPERTY_ENABLE_SGPR_*), see documentation of amd_kernel_code_t 73 /// for initial register state. 74 /// 75 /// The total number of SGPRuser data registers requested must not 76 /// exceed 16. Any requests beyond 16 will be ignored. 77 /// 78 /// Used to set COMPUTE_PGM_RSRC2.USER_SGPR (set to total count of 79 /// SGPR user data registers enabled up to 16). 80 81 AMD_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_BUFFER_SHIFT = 0, 82 AMD_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_BUFFER_WIDTH = 1, 83 AMD_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_BUFFER = ((1 << AMD_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_BUFFER_WIDTH) - 1) << AMD_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_BUFFER_SHIFT, 84 85 AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_PTR_SHIFT = 1, 86 AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_PTR_WIDTH = 1, 87 AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_PTR = ((1 << AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_PTR_WIDTH) - 1) << AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_PTR_SHIFT, 88 89 AMD_CODE_PROPERTY_ENABLE_SGPR_QUEUE_PTR_SHIFT = 2, 90 AMD_CODE_PROPERTY_ENABLE_SGPR_QUEUE_PTR_WIDTH = 1, 91 AMD_CODE_PROPERTY_ENABLE_SGPR_QUEUE_PTR = ((1 << AMD_CODE_PROPERTY_ENABLE_SGPR_QUEUE_PTR_WIDTH) - 1) << AMD_CODE_PROPERTY_ENABLE_SGPR_QUEUE_PTR_SHIFT, 92 93 AMD_CODE_PROPERTY_ENABLE_SGPR_KERNARG_SEGMENT_PTR_SHIFT = 3, 94 AMD_CODE_PROPERTY_ENABLE_SGPR_KERNARG_SEGMENT_PTR_WIDTH = 1, 95 AMD_CODE_PROPERTY_ENABLE_SGPR_KERNARG_SEGMENT_PTR = ((1 << AMD_CODE_PROPERTY_ENABLE_SGPR_KERNARG_SEGMENT_PTR_WIDTH) - 1) << AMD_CODE_PROPERTY_ENABLE_SGPR_KERNARG_SEGMENT_PTR_SHIFT, 96 97 AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_ID_SHIFT = 4, 98 AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_ID_WIDTH = 1, 99 AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_ID = ((1 << AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_ID_WIDTH) - 1) << AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_ID_SHIFT, 100 101 AMD_CODE_PROPERTY_ENABLE_SGPR_FLAT_SCRATCH_INIT_SHIFT = 5, 102 AMD_CODE_PROPERTY_ENABLE_SGPR_FLAT_SCRATCH_INIT_WIDTH = 1, 103 AMD_CODE_PROPERTY_ENABLE_SGPR_FLAT_SCRATCH_INIT = ((1 << AMD_CODE_PROPERTY_ENABLE_SGPR_FLAT_SCRATCH_INIT_WIDTH) - 1) << AMD_CODE_PROPERTY_ENABLE_SGPR_FLAT_SCRATCH_INIT_SHIFT, 104 105 AMD_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_SIZE_SHIFT = 6, 106 AMD_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_SIZE_WIDTH = 1, 107 AMD_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_SIZE = ((1 << AMD_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_SIZE_WIDTH) - 1) << AMD_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_SIZE_SHIFT, 108 109 AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_X_SHIFT = 7, 110 AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_X_WIDTH = 1, 111 AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_X = ((1 << AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_X_WIDTH) - 1) << AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_X_SHIFT, 112 113 AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_Y_SHIFT = 8, 114 AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_Y_WIDTH = 1, 115 AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_Y = ((1 << AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_Y_WIDTH) - 1) << AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_Y_SHIFT, 116 117 AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_Z_SHIFT = 9, 118 AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_Z_WIDTH = 1, 119 AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_Z = ((1 << AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_Z_WIDTH) - 1) << AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_Z_SHIFT, 120 121 /// Control wave ID base counter for GDS ordered-append. Used to set 122 /// COMPUTE_DISPATCH_INITIATOR.ORDERED_APPEND_ENBL. (Not sure if 123 /// ORDERED_APPEND_MODE also needs to be settable) 124 AMD_CODE_PROPERTY_ENABLE_ORDERED_APPEND_GDS_SHIFT = 10, 125 AMD_CODE_PROPERTY_ENABLE_ORDERED_APPEND_GDS_WIDTH = 1, 126 AMD_CODE_PROPERTY_ENABLE_ORDERED_APPEND_GDS = ((1 << AMD_CODE_PROPERTY_ENABLE_ORDERED_APPEND_GDS_WIDTH) - 1) << AMD_CODE_PROPERTY_ENABLE_ORDERED_APPEND_GDS_SHIFT, 127 128 /// The interleave (swizzle) element size in bytes required by the 129 /// code for private memory. This must be 2, 4, 8 or 16. This value 130 /// is provided to the finalizer when it is invoked and is recorded 131 /// here. The hardware will interleave the memory requests of each 132 /// lane of a wavefront by this element size to ensure each 133 /// work-item gets a distinct memory memory location. Therefore, the 134 /// finalizer ensures that all load and store operations done to 135 /// private memory do not exceed this size. For example, if the 136 /// element size is 4 (32-bits or dword) and a 64-bit value must be 137 /// loaded, the finalizer will generate two 32-bit loads. This 138 /// ensures that the interleaving will get the work-item 139 /// specific dword for both halves of the 64-bit value. If it just 140 /// did a 64-bit load then it would get one dword which belonged to 141 /// its own work-item, but the second dword would belong to the 142 /// adjacent lane work-item since the interleaving is in dwords. 143 /// 144 /// The value used must match the value that the runtime configures 145 /// the GPU flat scratch (SH_STATIC_MEM_CONFIG.ELEMENT_SIZE). This 146 /// is generally DWORD. 147 /// 148 /// uSE VALUES FROM THE AMD_ELEMENT_BYTE_SIZE_T ENUM. 149 AMD_CODE_PROPERTY_PRIVATE_ELEMENT_SIZE_SHIFT = 11, 150 AMD_CODE_PROPERTY_PRIVATE_ELEMENT_SIZE_WIDTH = 2, 151 AMD_CODE_PROPERTY_PRIVATE_ELEMENT_SIZE = ((1 << AMD_CODE_PROPERTY_PRIVATE_ELEMENT_SIZE_WIDTH) - 1) << AMD_CODE_PROPERTY_PRIVATE_ELEMENT_SIZE_SHIFT, 152 153 /// Are global memory addresses 64 bits. Must match 154 /// amd_kernel_code_t.hsail_machine_model == 155 /// HSA_MACHINE_LARGE. Must also match 156 /// SH_MEM_CONFIG.PTR32 (GFX6 (SI)/GFX7 (CI)), 157 /// SH_MEM_CONFIG.ADDRESS_MODE (GFX8 (VI)+). 158 AMD_CODE_PROPERTY_IS_PTR64_SHIFT = 13, 159 AMD_CODE_PROPERTY_IS_PTR64_WIDTH = 1, 160 AMD_CODE_PROPERTY_IS_PTR64 = ((1 << AMD_CODE_PROPERTY_IS_PTR64_WIDTH) - 1) << AMD_CODE_PROPERTY_IS_PTR64_SHIFT, 161 162 /// Indicate if the generated ISA is using a dynamically sized call 163 /// stack. This can happen if calls are implemented using a call 164 /// stack and recursion, alloca or calls to indirect functions are 165 /// present. In these cases the Finalizer cannot compute the total 166 /// private segment size at compile time. In this case the 167 /// workitem_private_segment_byte_size only specifies the statically 168 /// know private segment size, and additional space must be added 169 /// for the call stack. 170 AMD_CODE_PROPERTY_IS_DYNAMIC_CALLSTACK_SHIFT = 14, 171 AMD_CODE_PROPERTY_IS_DYNAMIC_CALLSTACK_WIDTH = 1, 172 AMD_CODE_PROPERTY_IS_DYNAMIC_CALLSTACK = ((1 << AMD_CODE_PROPERTY_IS_DYNAMIC_CALLSTACK_WIDTH) - 1) << AMD_CODE_PROPERTY_IS_DYNAMIC_CALLSTACK_SHIFT, 173 174 /// Indicate if code generated has support for debugging. 175 AMD_CODE_PROPERTY_IS_DEBUG_SUPPORTED_SHIFT = 15, 176 AMD_CODE_PROPERTY_IS_DEBUG_SUPPORTED_WIDTH = 1, 177 AMD_CODE_PROPERTY_IS_DEBUG_SUPPORTED = ((1 << AMD_CODE_PROPERTY_IS_DEBUG_SUPPORTED_WIDTH) - 1) << AMD_CODE_PROPERTY_IS_DEBUG_SUPPORTED_SHIFT, 178 179 AMD_CODE_PROPERTY_IS_XNACK_SUPPORTED_SHIFT = 15, 180 AMD_CODE_PROPERTY_IS_XNACK_SUPPORTED_WIDTH = 1, 181 AMD_CODE_PROPERTY_IS_XNACK_SUPPORTED = ((1 << AMD_CODE_PROPERTY_IS_XNACK_SUPPORTED_WIDTH) - 1) << AMD_CODE_PROPERTY_IS_XNACK_SUPPORTED_SHIFT 182 }; 183 184 /// @brief The hsa_ext_control_directives_t specifies the values for the HSAIL 185 /// control directives. These control how the finalizer generates code. This 186 /// struct is used both as an argument to hsaFinalizeKernel to specify values for 187 /// the control directives, and is used in HsaKernelCode to record the values of 188 /// the control directives that the finalize used when generating the code which 189 /// either came from the finalizer argument or explicit HSAIL control 190 /// directives. See the definition of the control directives in HSA Programmer's 191 /// Reference Manual which also defines how the values specified as finalizer 192 /// arguments have to agree with the control directives in the HSAIL code. 193 typedef struct hsa_ext_control_directives_s { 194 /// This is a bit set indicating which control directives have been 195 /// specified. If the value is 0 then there are no control directives specified 196 /// and the rest of the fields can be ignored. The bits are accessed using the 197 /// hsa_ext_control_directives_present_mask_t. Any control directive that is not 198 /// enabled in this bit set must have the value of all 0s. 199 hsa_ext_control_directive_present64_t enabled_control_directives; 200 201 /// If enableBreakExceptions is not enabled then must be 0, otherwise must be 202 /// non-0 and specifies the set of HSAIL exceptions that must have the BREAK 203 /// policy enabled. If this set is not empty then the generated code may have 204 /// lower performance than if the set is empty. If the kernel being finalized 205 /// has any enablebreakexceptions control directives, then the values specified 206 /// by this argument are unioned with the values in these control 207 /// directives. If any of the functions the kernel calls have an 208 /// enablebreakexceptions control directive, then they must be equal or a 209 /// subset of, this union. 210 hsa_ext_exception_kind16_t enable_break_exceptions; 211 212 /// If enableDetectExceptions is not enabled then must be 0, otherwise must be 213 /// non-0 and specifies the set of HSAIL exceptions that must have the DETECT 214 /// policy enabled. If this set is not empty then the generated code may have 215 /// lower performance than if the set is empty. However, an implementation 216 /// should endeavour to make the performance impact small. If the kernel being 217 /// finalized has any enabledetectexceptions control directives, then the 218 /// values specified by this argument are unioned with the values in these 219 /// control directives. If any of the functions the kernel calls have an 220 /// enabledetectexceptions control directive, then they must be equal or a 221 /// subset of, this union. 222 hsa_ext_exception_kind16_t enable_detect_exceptions; 223 224 /// If maxDynamicGroupSize is not enabled then must be 0, and any amount of 225 /// dynamic group segment can be allocated for a dispatch, otherwise the value 226 /// specifies the maximum number of bytes of dynamic group segment that can be 227 /// allocated for a dispatch. If the kernel being finalized has any 228 /// maxdynamicsize control directives, then the values must be the same, and 229 /// must be the same as this argument if it is enabled. This value can be used 230 /// by the finalizer to determine the maximum number of bytes of group memory 231 /// used by each work-group by adding this value to the group memory required 232 /// for all group segment variables used by the kernel and all functions it 233 /// calls, and group memory used to implement other HSAIL features such as 234 /// fbarriers and the detect exception operations. This can allow the finalizer 235 /// to determine the expected number of work-groups that can be executed by a 236 /// compute unit and allow more resources to be allocated to the work-items if 237 /// it is known that fewer work-groups can be executed due to group memory 238 /// limitations. 239 uint32_t max_dynamic_group_size; 240 241 /// If maxFlatGridSize is not enabled then must be 0, otherwise must be greater 242 /// than 0. See HSA Programmer's Reference Manual description of 243 /// maxflatgridsize control directive. 244 uint32_t max_flat_grid_size; 245 246 /// If maxFlatWorkgroupSize is not enabled then must be 0, otherwise must be 247 /// greater than 0. See HSA Programmer's Reference Manual description of 248 /// maxflatworkgroupsize control directive. 249 uint32_t max_flat_workgroup_size; 250 251 /// If requestedWorkgroupsPerCu is not enabled then must be 0, and the 252 /// finalizer is free to generate ISA that may result in any number of 253 /// work-groups executing on a single compute unit. Otherwise, the finalizer 254 /// should attempt to generate ISA that will allow the specified number of 255 /// work-groups to execute on a single compute unit. This is only a hint and 256 /// can be ignored by the finalizer. If the kernel being finalized, or any of 257 /// the functions it calls, has a requested control directive, then the values 258 /// must be the same. This can be used to determine the number of resources 259 /// that should be allocated to a single work-group and work-item. For example, 260 /// a low value may allow more resources to be allocated, resulting in higher 261 /// per work-item performance, as it is known there will never be more than the 262 /// specified number of work-groups actually executing on the compute 263 /// unit. Conversely, a high value may allocate fewer resources, resulting in 264 /// lower per work-item performance, which is offset by the fact it allows more 265 /// work-groups to actually execute on the compute unit. 266 uint32_t requested_workgroups_per_cu; 267 268 /// If not enabled then all elements for Dim3 must be 0, otherwise every 269 /// element must be greater than 0. See HSA Programmer's Reference Manual 270 /// description of requiredgridsize control directive. 271 hsa_dim3_t required_grid_size; 272 273 /// If requiredWorkgroupSize is not enabled then all elements for Dim3 must be 274 /// 0, and the produced code can be dispatched with any legal work-group range 275 /// consistent with the dispatch dimensions. Otherwise, the code produced must 276 /// always be dispatched with the specified work-group range. No element of the 277 /// specified range must be 0. It must be consistent with required_dimensions 278 /// and max_flat_workgroup_size. If the kernel being finalized, or any of the 279 /// functions it calls, has a requiredworkgroupsize control directive, then the 280 /// values must be the same. Specifying a value can allow the finalizer to 281 /// optimize work-group id operations, and if the number of work-items in the 282 /// work-group is less than the WAVESIZE then barrier operations can be 283 /// optimized to just a memory fence. 284 hsa_dim3_t required_workgroup_size; 285 286 /// If requiredDim is not enabled then must be 0 and the produced kernel code 287 /// can be dispatched with 1, 2 or 3 dimensions. If enabled then the value is 288 /// 1..3 and the code produced must only be dispatched with a dimension that 289 /// matches. Other values are illegal. If the kernel being finalized, or any of 290 /// the functions it calls, has a requireddimsize control directive, then the 291 /// values must be the same. This can be used to optimize the code generated to 292 /// compute the absolute and flat work-group and work-item id, and the dim 293 /// HSAIL operations. 294 uint8_t required_dim; 295 296 /// Reserved. Must be 0. 297 uint8_t reserved[75]; 298 } hsa_ext_control_directives_t; 299 300 /// AMD Kernel Code Object (amd_kernel_code_t). GPU CP uses the AMD Kernel 301 /// Code Object to set up the hardware to execute the kernel dispatch. 302 /// 303 /// Initial Kernel Register State. 304 /// 305 /// Initial kernel register state will be set up by CP/SPI prior to the start 306 /// of execution of every wavefront. This is limited by the constraints of the 307 /// current hardware. 308 /// 309 /// The order of the SGPR registers is defined, but the Finalizer can specify 310 /// which ones are actually setup in the amd_kernel_code_t object using the 311 /// enable_sgpr_* bit fields. The register numbers used for enabled registers 312 /// are dense starting at SGPR0: the first enabled register is SGPR0, the next 313 /// enabled register is SGPR1 etc.; disabled registers do not have an SGPR 314 /// number. 315 /// 316 /// The initial SGPRs comprise up to 16 User SRGPs that are set up by CP and 317 /// apply to all waves of the grid. It is possible to specify more than 16 User 318 /// SGPRs using the enable_sgpr_* bit fields, in which case only the first 16 319 /// are actually initialized. These are then immediately followed by the System 320 /// SGPRs that are set up by ADC/SPI and can have different values for each wave 321 /// of the grid dispatch. 322 /// 323 /// SGPR register initial state is defined as follows: 324 /// 325 /// Private Segment Buffer (enable_sgpr_private_segment_buffer): 326 /// Number of User SGPR registers: 4. V# that can be used, together with 327 /// Scratch Wave Offset as an offset, to access the Private/Spill/Arg 328 /// segments using a segment address. It must be set as follows: 329 /// - Base address: of the scratch memory area used by the dispatch. It 330 /// does not include the scratch wave offset. It will be the per process 331 /// SH_HIDDEN_PRIVATE_BASE_VMID plus any offset from this dispatch (for 332 /// example there may be a per pipe offset, or per AQL Queue offset). 333 /// - Stride + data_format: Element Size * Index Stride (???) 334 /// - Cache swizzle: ??? 335 /// - Swizzle enable: SH_STATIC_MEM_CONFIG.SWIZZLE_ENABLE (must be 1 for 336 /// scratch) 337 /// - Num records: Flat Scratch Work Item Size / Element Size (???) 338 /// - Dst_sel_*: ??? 339 /// - Num_format: ??? 340 /// - Element_size: SH_STATIC_MEM_CONFIG.ELEMENT_SIZE (will be DWORD, must 341 /// agree with amd_kernel_code_t.privateElementSize) 342 /// - Index_stride: SH_STATIC_MEM_CONFIG.INDEX_STRIDE (will be 64 as must 343 /// be number of wavefront lanes for scratch, must agree with 344 /// amd_kernel_code_t.wavefrontSize) 345 /// - Add tid enable: 1 346 /// - ATC: from SH_MEM_CONFIG.PRIVATE_ATC, 347 /// - Hash_enable: ??? 348 /// - Heap: ??? 349 /// - Mtype: from SH_STATIC_MEM_CONFIG.PRIVATE_MTYPE 350 /// - Type: 0 (a buffer) (???) 351 /// 352 /// Dispatch Ptr (enable_sgpr_dispatch_ptr): 353 /// Number of User SGPR registers: 2. 64 bit address of AQL dispatch packet 354 /// for kernel actually executing. 355 /// 356 /// Queue Ptr (enable_sgpr_queue_ptr): 357 /// Number of User SGPR registers: 2. 64 bit address of AmdQueue object for 358 /// AQL queue on which the dispatch packet was queued. 359 /// 360 /// Kernarg Segment Ptr (enable_sgpr_kernarg_segment_ptr): 361 /// Number of User SGPR registers: 2. 64 bit address of Kernarg segment. This 362 /// is directly copied from the kernargPtr in the dispatch packet. Having CP 363 /// load it once avoids loading it at the beginning of every wavefront. 364 /// 365 /// Dispatch Id (enable_sgpr_dispatch_id): 366 /// Number of User SGPR registers: 2. 64 bit Dispatch ID of the dispatch 367 /// packet being executed. 368 /// 369 /// Flat Scratch Init (enable_sgpr_flat_scratch_init): 370 /// Number of User SGPR registers: 2. This is 2 SGPRs. 371 /// 372 /// For CI/VI: 373 /// The first SGPR is a 32 bit byte offset from SH_MEM_HIDDEN_PRIVATE_BASE 374 /// to base of memory for scratch for this dispatch. This is the same offset 375 /// used in computing the Scratch Segment Buffer base address. The value of 376 /// Scratch Wave Offset must be added by the kernel code and moved to 377 /// SGPRn-4 for use as the FLAT SCRATCH BASE in flat memory instructions. 378 /// 379 /// The second SGPR is 32 bit byte size of a single work-item's scratch 380 /// memory usage. This is directly loaded from the dispatch packet Private 381 /// Segment Byte Size and rounded up to a multiple of DWORD. 382 /// 383 /// \todo [Does CP need to round this to >4 byte alignment?] 384 /// 385 /// The kernel code must move to SGPRn-3 for use as the FLAT SCRATCH SIZE in 386 /// flat memory instructions. Having CP load it once avoids loading it at 387 /// the beginning of every wavefront. 388 /// 389 /// For PI: 390 /// This is the 64 bit base address of the scratch backing memory for 391 /// allocated by CP for this dispatch. 392 /// 393 /// Private Segment Size (enable_sgpr_private_segment_size): 394 /// Number of User SGPR registers: 1. The 32 bit byte size of a single 395 /// work-item's scratch memory allocation. This is the value from the dispatch 396 /// packet. Private Segment Byte Size rounded up by CP to a multiple of DWORD. 397 /// 398 /// \todo [Does CP need to round this to >4 byte alignment?] 399 /// 400 /// Having CP load it once avoids loading it at the beginning of every 401 /// wavefront. 402 /// 403 /// \todo [This will not be used for CI/VI since it is the same value as 404 /// the second SGPR of Flat Scratch Init. However, it is need for PI which 405 /// changes meaning of Flat Scratchg Init..] 406 /// 407 /// Grid Work-Group Count X (enable_sgpr_grid_workgroup_count_x): 408 /// Number of User SGPR registers: 1. 32 bit count of the number of 409 /// work-groups in the X dimension for the grid being executed. Computed from 410 /// the fields in the HsaDispatchPacket as 411 /// ((gridSize.x+workgroupSize.x-1)/workgroupSize.x). 412 /// 413 /// Grid Work-Group Count Y (enable_sgpr_grid_workgroup_count_y): 414 /// Number of User SGPR registers: 1. 32 bit count of the number of 415 /// work-groups in the Y dimension for the grid being executed. Computed from 416 /// the fields in the HsaDispatchPacket as 417 /// ((gridSize.y+workgroupSize.y-1)/workgroupSize.y). 418 /// 419 /// Only initialized if <16 previous SGPRs initialized. 420 /// 421 /// Grid Work-Group Count Z (enable_sgpr_grid_workgroup_count_z): 422 /// Number of User SGPR registers: 1. 32 bit count of the number of 423 /// work-groups in the Z dimension for the grid being executed. Computed 424 /// from the fields in the HsaDispatchPacket as 425 /// ((gridSize.z+workgroupSize.z-1)/workgroupSize.z). 426 /// 427 /// Only initialized if <16 previous SGPRs initialized. 428 /// 429 /// Work-Group Id X (enable_sgpr_workgroup_id_x): 430 /// Number of System SGPR registers: 1. 32 bit work group id in X dimension 431 /// of grid for wavefront. Always present. 432 /// 433 /// Work-Group Id Y (enable_sgpr_workgroup_id_y): 434 /// Number of System SGPR registers: 1. 32 bit work group id in Y dimension 435 /// of grid for wavefront. 436 /// 437 /// Work-Group Id Z (enable_sgpr_workgroup_id_z): 438 /// Number of System SGPR registers: 1. 32 bit work group id in Z dimension 439 /// of grid for wavefront. If present then Work-group Id Y will also be 440 /// present 441 /// 442 /// Work-Group Info (enable_sgpr_workgroup_info): 443 /// Number of System SGPR registers: 1. {first_wave, 14'b0000, 444 /// ordered_append_term[10:0], threadgroup_size_in_waves[5:0]} 445 /// 446 /// Private Segment Wave Byte Offset 447 /// (enable_sgpr_private_segment_wave_byte_offset): 448 /// Number of System SGPR registers: 1. 32 bit byte offset from base of 449 /// dispatch scratch base. Must be used as an offset with Private/Spill/Arg 450 /// segment address when using Scratch Segment Buffer. It must be added to 451 /// Flat Scratch Offset if setting up FLAT SCRATCH for flat addressing. 452 /// 453 /// 454 /// The order of the VGPR registers is defined, but the Finalizer can specify 455 /// which ones are actually setup in the amd_kernel_code_t object using the 456 /// enableVgpr* bit fields. The register numbers used for enabled registers 457 /// are dense starting at VGPR0: the first enabled register is VGPR0, the next 458 /// enabled register is VGPR1 etc.; disabled registers do not have an VGPR 459 /// number. 460 /// 461 /// VGPR register initial state is defined as follows: 462 /// 463 /// Work-Item Id X (always initialized): 464 /// Number of registers: 1. 32 bit work item id in X dimension of work-group 465 /// for wavefront lane. 466 /// 467 /// Work-Item Id X (enable_vgpr_workitem_id > 0): 468 /// Number of registers: 1. 32 bit work item id in Y dimension of work-group 469 /// for wavefront lane. 470 /// 471 /// Work-Item Id X (enable_vgpr_workitem_id > 0): 472 /// Number of registers: 1. 32 bit work item id in Z dimension of work-group 473 /// for wavefront lane. 474 /// 475 /// 476 /// The setting of registers is being done by existing GPU hardware as follows: 477 /// 1) SGPRs before the Work-Group Ids are set by CP using the 16 User Data 478 /// registers. 479 /// 2) Work-group Id registers X, Y, Z are set by SPI which supports any 480 /// combination including none. 481 /// 3) Scratch Wave Offset is also set by SPI which is why its value cannot 482 /// be added into the value Flat Scratch Offset which would avoid the 483 /// Finalizer generated prolog having to do the add. 484 /// 4) The VGPRs are set by SPI which only supports specifying either (X), 485 /// (X, Y) or (X, Y, Z). 486 /// 487 /// Flat Scratch Dispatch Offset and Flat Scratch Size are adjacent SGRRs so 488 /// they can be moved as a 64 bit value to the hardware required SGPRn-3 and 489 /// SGPRn-4 respectively using the Finalizer ?FLAT_SCRATCH? Register. 490 /// 491 /// The global segment can be accessed either using flat operations or buffer 492 /// operations. If buffer operations are used then the Global Buffer used to 493 /// access HSAIL Global/Readonly/Kernarg (which are combine) segments using a 494 /// segment address is not passed into the kernel code by CP since its base 495 /// address is always 0. Instead the Finalizer generates prolog code to 496 /// initialize 4 SGPRs with a V# that has the following properties, and then 497 /// uses that in the buffer instructions: 498 /// - base address of 0 499 /// - no swizzle 500 /// - ATC=1 501 /// - MTYPE set to support memory coherence specified in 502 /// amd_kernel_code_t.globalMemoryCoherence 503 /// 504 /// When the Global Buffer is used to access the Kernarg segment, must add the 505 /// dispatch packet kernArgPtr to a kernarg segment address before using this V#. 506 /// Alternatively scalar loads can be used if the kernarg offset is uniform, as 507 /// the kernarg segment is constant for the duration of the kernel execution. 508 /// 509 510 typedef struct amd_kernel_code_s { 511 uint32_t amd_kernel_code_version_major; 512 uint32_t amd_kernel_code_version_minor; 513 uint16_t amd_machine_kind; 514 uint16_t amd_machine_version_major; 515 uint16_t amd_machine_version_minor; 516 uint16_t amd_machine_version_stepping; 517 518 /// Byte offset (possibly negative) from start of amd_kernel_code_t 519 /// object to kernel's entry point instruction. The actual code for 520 /// the kernel is required to be 256 byte aligned to match hardware 521 /// requirements (SQ cache line is 16). The code must be position 522 /// independent code (PIC) for AMD devices to give runtime the 523 /// option of copying code to discrete GPU memory or APU L2 524 /// cache. The Finalizer should endeavour to allocate all kernel 525 /// machine code in contiguous memory pages so that a device 526 /// pre-fetcher will tend to only pre-fetch Kernel Code objects, 527 /// improving cache performance. 528 int64_t kernel_code_entry_byte_offset; 529 530 /// Range of bytes to consider prefetching expressed as an offset 531 /// and size. The offset is from the start (possibly negative) of 532 /// amd_kernel_code_t object. Set both to 0 if no prefetch 533 /// information is available. 534 int64_t kernel_code_prefetch_byte_offset; 535 uint64_t kernel_code_prefetch_byte_size; 536 537 /// Number of bytes of scratch backing memory required for full 538 /// occupancy of target chip. This takes into account the number of 539 /// bytes of scratch per work-item, the wavefront size, the maximum 540 /// number of wavefronts per CU, and the number of CUs. This is an 541 /// upper limit on scratch. If the grid being dispatched is small it 542 /// may only need less than this. If the kernel uses no scratch, or 543 /// the Finalizer has not computed this value, it must be 0. 544 uint64_t max_scratch_backing_memory_byte_size; 545 546 /// Shader program settings for CS. Contains COMPUTE_PGM_RSRC1 and 547 /// COMPUTE_PGM_RSRC2 registers. 548 uint64_t compute_pgm_resource_registers; 549 550 /// Code properties. See amd_code_property_mask_t for a full list of 551 /// properties. 552 uint32_t code_properties; 553 554 /// The amount of memory required for the combined private, spill 555 /// and arg segments for a work-item in bytes. If 556 /// is_dynamic_callstack is 1 then additional space must be added to 557 /// this value for the call stack. 558 uint32_t workitem_private_segment_byte_size; 559 560 /// The amount of group segment memory required by a work-group in 561 /// bytes. This does not include any dynamically allocated group 562 /// segment memory that may be added when the kernel is 563 /// dispatched. 564 uint32_t workgroup_group_segment_byte_size; 565 566 /// Number of byte of GDS required by kernel dispatch. Must be 0 if 567 /// not using GDS. 568 uint32_t gds_segment_byte_size; 569 570 /// The size in bytes of the kernarg segment that holds the values 571 /// of the arguments to the kernel. This could be used by CP to 572 /// prefetch the kernarg segment pointed to by the dispatch packet. 573 uint64_t kernarg_segment_byte_size; 574 575 /// Number of fbarrier's used in the kernel and all functions it 576 /// calls. If the implementation uses group memory to allocate the 577 /// fbarriers then that amount must already be included in the 578 /// workgroup_group_segment_byte_size total. 579 uint32_t workgroup_fbarrier_count; 580 581 /// Number of scalar registers used by a wavefront. This includes 582 /// the special SGPRs for VCC, Flat Scratch Base, Flat Scratch Size 583 /// and XNACK (for GFX8 (VI)). It does not include the 16 SGPR added if a 584 /// trap handler is enabled. Used to set COMPUTE_PGM_RSRC1.SGPRS. 585 uint16_t wavefront_sgpr_count; 586 587 /// Number of vector registers used by each work-item. Used to set 588 /// COMPUTE_PGM_RSRC1.VGPRS. 589 uint16_t workitem_vgpr_count; 590 591 /// If reserved_vgpr_count is 0 then must be 0. Otherwise, this is the 592 /// first fixed VGPR number reserved. 593 uint16_t reserved_vgpr_first; 594 595 /// The number of consecutive VGPRs reserved by the client. If 596 /// is_debug_supported then this count includes VGPRs reserved 597 /// for debugger use. 598 uint16_t reserved_vgpr_count; 599 600 /// If reserved_sgpr_count is 0 then must be 0. Otherwise, this is the 601 /// first fixed SGPR number reserved. 602 uint16_t reserved_sgpr_first; 603 604 /// The number of consecutive SGPRs reserved by the client. If 605 /// is_debug_supported then this count includes SGPRs reserved 606 /// for debugger use. 607 uint16_t reserved_sgpr_count; 608 609 /// If is_debug_supported is 0 then must be 0. Otherwise, this is the 610 /// fixed SGPR number used to hold the wave scratch offset for the 611 /// entire kernel execution, or uint16_t(-1) if the register is not 612 /// used or not known. 613 uint16_t debug_wavefront_private_segment_offset_sgpr; 614 615 /// If is_debug_supported is 0 then must be 0. Otherwise, this is the 616 /// fixed SGPR number of the first of 4 SGPRs used to hold the 617 /// scratch V# used for the entire kernel execution, or uint16_t(-1) 618 /// if the registers are not used or not known. 619 uint16_t debug_private_segment_buffer_sgpr; 620 621 /// The maximum byte alignment of variables used by the kernel in 622 /// the specified memory segment. Expressed as a power of two. Must 623 /// be at least HSA_POWERTWO_16. 624 uint8_t kernarg_segment_alignment; 625 uint8_t group_segment_alignment; 626 uint8_t private_segment_alignment; 627 628 /// Wavefront size expressed as a power of two. Must be a power of 2 629 /// in range 1..64 inclusive. Used to support runtime query that 630 /// obtains wavefront size, which may be used by application to 631 /// allocated dynamic group memory and set the dispatch work-group 632 /// size. 633 uint8_t wavefront_size; 634 635 int32_t call_convention; 636 uint8_t reserved3[12]; 637 uint64_t runtime_loader_kernel_symbol; 638 uint64_t control_directives[16]; 639 } amd_kernel_code_t; 640 641 #endif // AMDKERNELCODET_H 642