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