xref: /freebsd/contrib/llvm-project/llvm/lib/Target/AMDGPU/AMDKernelCodeT.h (revision dd41de95a84d979615a2ef11df6850622bf6184e)
1 //===-- AMDGPUKernelCodeT.h - Print AMDGPU assembly code ---------*- C++ -*-===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 /// \file AMDKernelCodeT.h
9 //===----------------------------------------------------------------------===//
10 
11 #ifndef AMDKERNELCODET_H
12 #define AMDKERNELCODET_H
13 
14 #include "llvm/MC/SubtargetFeature.h"
15 
16 #include <cstddef>
17 #include <cstdint>
18 
19 #include "llvm/Support/Debug.h"
20 //---------------------------------------------------------------------------//
21 // AMD Kernel Code, and its dependencies                                     //
22 //---------------------------------------------------------------------------//
23 
24 typedef uint8_t hsa_powertwo8_t;
25 typedef uint32_t hsa_ext_code_kind_t;
26 typedef uint8_t hsa_ext_brig_profile8_t;
27 typedef uint8_t hsa_ext_brig_machine_model8_t;
28 typedef uint64_t hsa_ext_control_directive_present64_t;
29 typedef uint16_t hsa_ext_exception_kind16_t;
30 typedef uint32_t hsa_ext_code_kind32_t;
31 
32 typedef struct hsa_dim3_s {
33   uint32_t x;
34   uint32_t y;
35   uint32_t z;
36 } hsa_dim3_t;
37 
38 /// The version of the amd_*_code_t struct. Minor versions must be
39 /// backward compatible.
40 typedef uint32_t amd_code_version32_t;
41 enum amd_code_version_t {
42   AMD_CODE_VERSION_MAJOR = 0,
43   AMD_CODE_VERSION_MINOR = 1
44 };
45 
46 // Sets val bits for specified mask in specified dst packed instance.
47 #define AMD_HSA_BITS_SET(dst, mask, val)                                       \
48   dst &= (~(1 << mask ## _SHIFT) & ~mask);                                     \
49   dst |= (((val) << mask ## _SHIFT) & mask)
50 
51 // Gets bits for specified mask from specified src packed instance.
52 #define AMD_HSA_BITS_GET(src, mask)                                            \
53   ((src & mask) >> mask ## _SHIFT)                                             \
54 
55 /// The values used to define the number of bytes to use for the
56 /// swizzle element size.
57 enum amd_element_byte_size_t {
58   AMD_ELEMENT_2_BYTES = 0,
59   AMD_ELEMENT_4_BYTES = 1,
60   AMD_ELEMENT_8_BYTES = 2,
61   AMD_ELEMENT_16_BYTES = 3
62 };
63 
64 /// Shader program settings for CS. Contains COMPUTE_PGM_RSRC1 and
65 /// COMPUTE_PGM_RSRC2 registers.
66 typedef uint64_t amd_compute_pgm_resource_register64_t;
67 
68 /// Every amd_*_code_t has the following properties, which are composed of
69 /// a number of bit fields. Every bit field has a mask (AMD_CODE_PROPERTY_*),
70 /// bit width (AMD_CODE_PROPERTY_*_WIDTH, and bit shift amount
71 /// (AMD_CODE_PROPERTY_*_SHIFT) for convenient access. Unused bits must be 0.
72 ///
73 /// (Note that bit fields cannot be used as their layout is
74 /// implementation defined in the C standard and so cannot be used to
75 /// specify an ABI)
76 typedef uint32_t amd_code_property32_t;
77 enum amd_code_property_mask_t {
78 
79   /// Enable the setup of the SGPR user data registers
80   /// (AMD_CODE_PROPERTY_ENABLE_SGPR_*), see documentation of amd_kernel_code_t
81   /// for initial register state.
82   ///
83   /// The total number of SGPRuser data registers requested must not
84   /// exceed 16. Any requests beyond 16 will be ignored.
85   ///
86   /// Used to set COMPUTE_PGM_RSRC2.USER_SGPR (set to total count of
87   /// SGPR user data registers enabled up to 16).
88 
89   AMD_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_BUFFER_SHIFT = 0,
90   AMD_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_BUFFER_WIDTH = 1,
91   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,
92 
93   AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_PTR_SHIFT = 1,
94   AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_PTR_WIDTH = 1,
95   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,
96 
97   AMD_CODE_PROPERTY_ENABLE_SGPR_QUEUE_PTR_SHIFT = 2,
98   AMD_CODE_PROPERTY_ENABLE_SGPR_QUEUE_PTR_WIDTH = 1,
99   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,
100 
101   AMD_CODE_PROPERTY_ENABLE_SGPR_KERNARG_SEGMENT_PTR_SHIFT = 3,
102   AMD_CODE_PROPERTY_ENABLE_SGPR_KERNARG_SEGMENT_PTR_WIDTH = 1,
103   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,
104 
105   AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_ID_SHIFT = 4,
106   AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_ID_WIDTH = 1,
107   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,
108 
109   AMD_CODE_PROPERTY_ENABLE_SGPR_FLAT_SCRATCH_INIT_SHIFT = 5,
110   AMD_CODE_PROPERTY_ENABLE_SGPR_FLAT_SCRATCH_INIT_WIDTH = 1,
111   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,
112 
113   AMD_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_SIZE_SHIFT = 6,
114   AMD_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_SIZE_WIDTH = 1,
115   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,
116 
117   AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_X_SHIFT = 7,
118   AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_X_WIDTH = 1,
119   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,
120 
121   AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_Y_SHIFT = 8,
122   AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_Y_WIDTH = 1,
123   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,
124 
125   AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_Z_SHIFT = 9,
126   AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_Z_WIDTH = 1,
127   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,
128 
129   AMD_CODE_PROPERTY_ENABLE_WAVEFRONT_SIZE32_SHIFT = 10,
130   AMD_CODE_PROPERTY_ENABLE_WAVEFRONT_SIZE32_WIDTH = 1,
131   AMD_CODE_PROPERTY_ENABLE_WAVEFRONT_SIZE32 = ((1 << AMD_CODE_PROPERTY_ENABLE_WAVEFRONT_SIZE32_WIDTH) - 1) << AMD_CODE_PROPERTY_ENABLE_WAVEFRONT_SIZE32_SHIFT,
132 
133   AMD_CODE_PROPERTY_RESERVED1_SHIFT = 11,
134   AMD_CODE_PROPERTY_RESERVED1_WIDTH = 5,
135   AMD_CODE_PROPERTY_RESERVED1 = ((1 << AMD_CODE_PROPERTY_RESERVED1_WIDTH) - 1) << AMD_CODE_PROPERTY_RESERVED1_SHIFT,
136 
137   /// Control wave ID base counter for GDS ordered-append. Used to set
138   /// COMPUTE_DISPATCH_INITIATOR.ORDERED_APPEND_ENBL. (Not sure if
139   /// ORDERED_APPEND_MODE also needs to be settable)
140   AMD_CODE_PROPERTY_ENABLE_ORDERED_APPEND_GDS_SHIFT = 16,
141   AMD_CODE_PROPERTY_ENABLE_ORDERED_APPEND_GDS_WIDTH = 1,
142   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,
143 
144   /// The interleave (swizzle) element size in bytes required by the
145   /// code for private memory. This must be 2, 4, 8 or 16. This value
146   /// is provided to the finalizer when it is invoked and is recorded
147   /// here. The hardware will interleave the memory requests of each
148   /// lane of a wavefront by this element size to ensure each
149   /// work-item gets a distinct memory memory location. Therefore, the
150   /// finalizer ensures that all load and store operations done to
151   /// private memory do not exceed this size. For example, if the
152   /// element size is 4 (32-bits or dword) and a 64-bit value must be
153   /// loaded, the finalizer will generate two 32-bit loads. This
154   /// ensures that the interleaving will get the work-item
155   /// specific dword for both halves of the 64-bit value. If it just
156   /// did a 64-bit load then it would get one dword which belonged to
157   /// its own work-item, but the second dword would belong to the
158   /// adjacent lane work-item since the interleaving is in dwords.
159   ///
160   /// The value used must match the value that the runtime configures
161   /// the GPU flat scratch (SH_STATIC_MEM_CONFIG.ELEMENT_SIZE). This
162   /// is generally DWORD.
163   ///
164   /// uSE VALUES FROM THE AMD_ELEMENT_BYTE_SIZE_T ENUM.
165   AMD_CODE_PROPERTY_PRIVATE_ELEMENT_SIZE_SHIFT = 17,
166   AMD_CODE_PROPERTY_PRIVATE_ELEMENT_SIZE_WIDTH = 2,
167   AMD_CODE_PROPERTY_PRIVATE_ELEMENT_SIZE = ((1 << AMD_CODE_PROPERTY_PRIVATE_ELEMENT_SIZE_WIDTH) - 1) << AMD_CODE_PROPERTY_PRIVATE_ELEMENT_SIZE_SHIFT,
168 
169   /// Are global memory addresses 64 bits. Must match
170   /// amd_kernel_code_t.hsail_machine_model ==
171   /// HSA_MACHINE_LARGE. Must also match
172   /// SH_MEM_CONFIG.PTR32 (GFX6 (SI)/GFX7 (CI)),
173   /// SH_MEM_CONFIG.ADDRESS_MODE (GFX8 (VI)+).
174   AMD_CODE_PROPERTY_IS_PTR64_SHIFT = 19,
175   AMD_CODE_PROPERTY_IS_PTR64_WIDTH = 1,
176   AMD_CODE_PROPERTY_IS_PTR64 = ((1 << AMD_CODE_PROPERTY_IS_PTR64_WIDTH) - 1) << AMD_CODE_PROPERTY_IS_PTR64_SHIFT,
177 
178   /// Indicate if the generated ISA is using a dynamically sized call
179   /// stack. This can happen if calls are implemented using a call
180   /// stack and recursion, alloca or calls to indirect functions are
181   /// present. In these cases the Finalizer cannot compute the total
182   /// private segment size at compile time. In this case the
183   /// workitem_private_segment_byte_size only specifies the statically
184   /// know private segment size, and additional space must be added
185   /// for the call stack.
186   AMD_CODE_PROPERTY_IS_DYNAMIC_CALLSTACK_SHIFT = 20,
187   AMD_CODE_PROPERTY_IS_DYNAMIC_CALLSTACK_WIDTH = 1,
188   AMD_CODE_PROPERTY_IS_DYNAMIC_CALLSTACK = ((1 << AMD_CODE_PROPERTY_IS_DYNAMIC_CALLSTACK_WIDTH) - 1) << AMD_CODE_PROPERTY_IS_DYNAMIC_CALLSTACK_SHIFT,
189 
190   /// Indicate if code generated has support for debugging.
191   AMD_CODE_PROPERTY_IS_DEBUG_SUPPORTED_SHIFT = 21,
192   AMD_CODE_PROPERTY_IS_DEBUG_SUPPORTED_WIDTH = 1,
193   AMD_CODE_PROPERTY_IS_DEBUG_SUPPORTED = ((1 << AMD_CODE_PROPERTY_IS_DEBUG_SUPPORTED_WIDTH) - 1) << AMD_CODE_PROPERTY_IS_DEBUG_SUPPORTED_SHIFT,
194 
195   AMD_CODE_PROPERTY_IS_XNACK_SUPPORTED_SHIFT = 22,
196   AMD_CODE_PROPERTY_IS_XNACK_SUPPORTED_WIDTH = 1,
197   AMD_CODE_PROPERTY_IS_XNACK_SUPPORTED = ((1 << AMD_CODE_PROPERTY_IS_XNACK_SUPPORTED_WIDTH) - 1) << AMD_CODE_PROPERTY_IS_XNACK_SUPPORTED_SHIFT,
198 
199   AMD_CODE_PROPERTY_RESERVED2_SHIFT = 23,
200   AMD_CODE_PROPERTY_RESERVED2_WIDTH = 9,
201   AMD_CODE_PROPERTY_RESERVED2 = ((1 << AMD_CODE_PROPERTY_RESERVED2_WIDTH) - 1) << AMD_CODE_PROPERTY_RESERVED2_SHIFT
202 };
203 
204 /// The hsa_ext_control_directives_t specifies the values for the HSAIL
205 /// control directives. These control how the finalizer generates code. This
206 /// struct is used both as an argument to hsaFinalizeKernel to specify values for
207 /// the control directives, and is used in HsaKernelCode to record the values of
208 /// the control directives that the finalize used when generating the code which
209 /// either came from the finalizer argument or explicit HSAIL control
210 /// directives. See the definition of the control directives in HSA Programmer's
211 /// Reference Manual which also defines how the values specified as finalizer
212 /// arguments have to agree with the control directives in the HSAIL code.
213 typedef struct hsa_ext_control_directives_s {
214   /// This is a bit set indicating which control directives have been
215   /// specified. If the value is 0 then there are no control directives specified
216   /// and the rest of the fields can be ignored. The bits are accessed using the
217   /// hsa_ext_control_directives_present_mask_t. Any control directive that is not
218   /// enabled in this bit set must have the value of all 0s.
219   hsa_ext_control_directive_present64_t enabled_control_directives;
220 
221   /// If enableBreakExceptions is not enabled then must be 0, otherwise must be
222   /// non-0 and specifies the set of HSAIL exceptions that must have the BREAK
223   /// policy enabled. If this set is not empty then the generated code may have
224   /// lower performance than if the set is empty. If the kernel being finalized
225   /// has any enablebreakexceptions control directives, then the values specified
226   /// by this argument are unioned with the values in these control
227   /// directives. If any of the functions the kernel calls have an
228   /// enablebreakexceptions control directive, then they must be equal or a
229   /// subset of, this union.
230   hsa_ext_exception_kind16_t enable_break_exceptions;
231 
232   /// If enableDetectExceptions is not enabled then must be 0, otherwise must be
233   /// non-0 and specifies the set of HSAIL exceptions that must have the DETECT
234   /// policy enabled. If this set is not empty then the generated code may have
235   /// lower performance than if the set is empty. However, an implementation
236   /// should endeavour to make the performance impact small. If the kernel being
237   /// finalized has any enabledetectexceptions control directives, then the
238   /// values specified by this argument are unioned with the values in these
239   /// control directives. If any of the functions the kernel calls have an
240   /// enabledetectexceptions control directive, then they must be equal or a
241   /// subset of, this union.
242   hsa_ext_exception_kind16_t enable_detect_exceptions;
243 
244   /// If maxDynamicGroupSize is not enabled then must be 0, and any amount of
245   /// dynamic group segment can be allocated for a dispatch, otherwise the value
246   /// specifies the maximum number of bytes of dynamic group segment that can be
247   /// allocated for a dispatch. If the kernel being finalized has any
248   /// maxdynamicsize control directives, then the values must be the same, and
249   /// must be the same as this argument if it is enabled. This value can be used
250   /// by the finalizer to determine the maximum number of bytes of group memory
251   /// used by each work-group by adding this value to the group memory required
252   /// for all group segment variables used by the kernel and all functions it
253   /// calls, and group memory used to implement other HSAIL features such as
254   /// fbarriers and the detect exception operations. This can allow the finalizer
255   /// to determine the expected number of work-groups that can be executed by a
256   /// compute unit and allow more resources to be allocated to the work-items if
257   /// it is known that fewer work-groups can be executed due to group memory
258   /// limitations.
259   uint32_t max_dynamic_group_size;
260 
261   /// If maxFlatGridSize is not enabled then must be 0, otherwise must be greater
262   /// than 0. See HSA Programmer's Reference Manual description of
263   /// maxflatgridsize control directive.
264   uint32_t max_flat_grid_size;
265 
266   /// If maxFlatWorkgroupSize is not enabled then must be 0, otherwise must be
267   /// greater than 0. See HSA Programmer's Reference Manual description of
268   /// maxflatworkgroupsize control directive.
269   uint32_t max_flat_workgroup_size;
270 
271   /// If requestedWorkgroupsPerCu is not enabled then must be 0, and the
272   /// finalizer is free to generate ISA that may result in any number of
273   /// work-groups executing on a single compute unit. Otherwise, the finalizer
274   /// should attempt to generate ISA that will allow the specified number of
275   /// work-groups to execute on a single compute unit. This is only a hint and
276   /// can be ignored by the finalizer. If the kernel being finalized, or any of
277   /// the functions it calls, has a requested control directive, then the values
278   /// must be the same. This can be used to determine the number of resources
279   /// that should be allocated to a single work-group and work-item. For example,
280   /// a low value may allow more resources to be allocated, resulting in higher
281   /// per work-item performance, as it is known there will never be more than the
282   /// specified number of work-groups actually executing on the compute
283   /// unit. Conversely, a high value may allocate fewer resources, resulting in
284   /// lower per work-item performance, which is offset by the fact it allows more
285   /// work-groups to actually execute on the compute unit.
286   uint32_t requested_workgroups_per_cu;
287 
288   /// If not enabled then all elements for Dim3 must be 0, otherwise every
289   /// element must be greater than 0. See HSA Programmer's Reference Manual
290   /// description of requiredgridsize control directive.
291   hsa_dim3_t required_grid_size;
292 
293   /// If requiredWorkgroupSize is not enabled then all elements for Dim3 must be
294   /// 0, and the produced code can be dispatched with any legal work-group range
295   /// consistent with the dispatch dimensions. Otherwise, the code produced must
296   /// always be dispatched with the specified work-group range. No element of the
297   /// specified range must be 0. It must be consistent with required_dimensions
298   /// and max_flat_workgroup_size. If the kernel being finalized, or any of the
299   /// functions it calls, has a requiredworkgroupsize control directive, then the
300   /// values must be the same. Specifying a value can allow the finalizer to
301   /// optimize work-group id operations, and if the number of work-items in the
302   /// work-group is less than the WAVESIZE then barrier operations can be
303   /// optimized to just a memory fence.
304   hsa_dim3_t required_workgroup_size;
305 
306   /// If requiredDim is not enabled then must be 0 and the produced kernel code
307   /// can be dispatched with 1, 2 or 3 dimensions. If enabled then the value is
308   /// 1..3 and the code produced must only be dispatched with a dimension that
309   /// matches. Other values are illegal. If the kernel being finalized, or any of
310   /// the functions it calls, has a requireddimsize control directive, then the
311   /// values must be the same. This can be used to optimize the code generated to
312   /// compute the absolute and flat work-group and work-item id, and the dim
313   /// HSAIL operations.
314   uint8_t required_dim;
315 
316   /// Reserved. Must be 0.
317   uint8_t reserved[75];
318 } hsa_ext_control_directives_t;
319 
320 /// AMD Kernel Code Object (amd_kernel_code_t). GPU CP uses the AMD Kernel
321 /// Code Object to set up the hardware to execute the kernel dispatch.
322 ///
323 /// Initial Kernel Register State.
324 ///
325 /// Initial kernel register state will be set up by CP/SPI prior to the start
326 /// of execution of every wavefront. This is limited by the constraints of the
327 /// current hardware.
328 ///
329 /// The order of the SGPR registers is defined, but the Finalizer can specify
330 /// which ones are actually setup in the amd_kernel_code_t object using the
331 /// enable_sgpr_* bit fields. The register numbers used for enabled registers
332 /// are dense starting at SGPR0: the first enabled register is SGPR0, the next
333 /// enabled register is SGPR1 etc.; disabled registers do not have an SGPR
334 /// number.
335 ///
336 /// The initial SGPRs comprise up to 16 User SRGPs that are set up by CP and
337 /// apply to all waves of the grid. It is possible to specify more than 16 User
338 /// SGPRs using the enable_sgpr_* bit fields, in which case only the first 16
339 /// are actually initialized. These are then immediately followed by the System
340 /// SGPRs that are set up by ADC/SPI and can have different values for each wave
341 /// of the grid dispatch.
342 ///
343 /// SGPR register initial state is defined as follows:
344 ///
345 /// Private Segment Buffer (enable_sgpr_private_segment_buffer):
346 ///   Number of User SGPR registers: 4. V# that can be used, together with
347 ///   Scratch Wave Offset as an offset, to access the Private/Spill/Arg
348 ///   segments using a segment address. It must be set as follows:
349 ///     - Base address: of the scratch memory area used by the dispatch. It
350 ///       does not include the scratch wave offset. It will be the per process
351 ///       SH_HIDDEN_PRIVATE_BASE_VMID plus any offset from this dispatch (for
352 ///       example there may be a per pipe offset, or per AQL Queue offset).
353 ///     - Stride + data_format: Element Size * Index Stride (???)
354 ///     - Cache swizzle: ???
355 ///     - Swizzle enable: SH_STATIC_MEM_CONFIG.SWIZZLE_ENABLE (must be 1 for
356 ///       scratch)
357 ///     - Num records: Flat Scratch Work Item Size / Element Size (???)
358 ///     - Dst_sel_*: ???
359 ///     - Num_format: ???
360 ///     - Element_size: SH_STATIC_MEM_CONFIG.ELEMENT_SIZE (will be DWORD, must
361 ///       agree with amd_kernel_code_t.privateElementSize)
362 ///     - Index_stride: SH_STATIC_MEM_CONFIG.INDEX_STRIDE (will be 64 as must
363 ///       be number of wavefront lanes for scratch, must agree with
364 ///       amd_kernel_code_t.wavefrontSize)
365 ///     - Add tid enable: 1
366 ///     - ATC: from SH_MEM_CONFIG.PRIVATE_ATC,
367 ///     - Hash_enable: ???
368 ///     - Heap: ???
369 ///     - Mtype: from SH_STATIC_MEM_CONFIG.PRIVATE_MTYPE
370 ///     - Type: 0 (a buffer) (???)
371 ///
372 /// Dispatch Ptr (enable_sgpr_dispatch_ptr):
373 ///   Number of User SGPR registers: 2. 64 bit address of AQL dispatch packet
374 ///   for kernel actually executing.
375 ///
376 /// Queue Ptr (enable_sgpr_queue_ptr):
377 ///   Number of User SGPR registers: 2. 64 bit address of AmdQueue object for
378 ///   AQL queue on which the dispatch packet was queued.
379 ///
380 /// Kernarg Segment Ptr (enable_sgpr_kernarg_segment_ptr):
381 ///   Number of User SGPR registers: 2. 64 bit address of Kernarg segment. This
382 ///   is directly copied from the kernargPtr in the dispatch packet. Having CP
383 ///   load it once avoids loading it at the beginning of every wavefront.
384 ///
385 /// Dispatch Id (enable_sgpr_dispatch_id):
386 ///   Number of User SGPR registers: 2. 64 bit Dispatch ID of the dispatch
387 ///   packet being executed.
388 ///
389 /// Flat Scratch Init (enable_sgpr_flat_scratch_init):
390 ///   Number of User SGPR registers: 2. This is 2 SGPRs.
391 ///
392 ///   For CI/VI:
393 ///     The first SGPR is a 32 bit byte offset from SH_MEM_HIDDEN_PRIVATE_BASE
394 ///     to base of memory for scratch for this dispatch. This is the same offset
395 ///     used in computing the Scratch Segment Buffer base address. The value of
396 ///     Scratch Wave Offset must be added by the kernel code and moved to
397 ///     SGPRn-4 for use as the FLAT SCRATCH BASE in flat memory instructions.
398 ///
399 ///     The second SGPR is 32 bit byte size of a single work-item's scratch
400 ///     memory usage. This is directly loaded from the dispatch packet Private
401 ///     Segment Byte Size and rounded up to a multiple of DWORD.
402 ///
403 ///     \todo [Does CP need to round this to >4 byte alignment?]
404 ///
405 ///     The kernel code must move to SGPRn-3 for use as the FLAT SCRATCH SIZE in
406 ///     flat memory instructions. Having CP load it once avoids loading it at
407 ///     the beginning of every wavefront.
408 ///
409 ///   For PI:
410 ///     This is the 64 bit base address of the scratch backing memory for
411 ///     allocated by CP for this dispatch.
412 ///
413 /// Private Segment Size (enable_sgpr_private_segment_size):
414 ///   Number of User SGPR registers: 1. The 32 bit byte size of a single
415 ///   work-item's scratch memory allocation. This is the value from the dispatch
416 ///   packet. Private Segment Byte Size rounded up by CP to a multiple of DWORD.
417 ///
418 ///   \todo [Does CP need to round this to >4 byte alignment?]
419 ///
420 ///   Having CP load it once avoids loading it at the beginning of every
421 ///   wavefront.
422 ///
423 ///   \todo [This will not be used for CI/VI since it is the same value as
424 ///   the second SGPR of Flat Scratch Init. However, it is need for PI which
425 ///   changes meaning of Flat Scratchg Init..]
426 ///
427 /// Grid Work-Group Count X (enable_sgpr_grid_workgroup_count_x):
428 ///   Number of User SGPR registers: 1. 32 bit count of the number of
429 ///   work-groups in the X dimension for the grid being executed. Computed from
430 ///   the fields in the HsaDispatchPacket as
431 ///   ((gridSize.x+workgroupSize.x-1)/workgroupSize.x).
432 ///
433 /// Grid Work-Group Count Y (enable_sgpr_grid_workgroup_count_y):
434 ///   Number of User SGPR registers: 1. 32 bit count of the number of
435 ///   work-groups in the Y dimension for the grid being executed. Computed from
436 ///   the fields in the HsaDispatchPacket as
437 ///   ((gridSize.y+workgroupSize.y-1)/workgroupSize.y).
438 ///
439 ///   Only initialized if <16 previous SGPRs initialized.
440 ///
441 /// Grid Work-Group Count Z (enable_sgpr_grid_workgroup_count_z):
442 ///   Number of User SGPR registers: 1. 32 bit count of the number of
443 ///   work-groups in the Z dimension for the grid being executed. Computed
444 ///   from the fields in the HsaDispatchPacket as
445 ///   ((gridSize.z+workgroupSize.z-1)/workgroupSize.z).
446 ///
447 ///   Only initialized if <16 previous SGPRs initialized.
448 ///
449 /// Work-Group Id X (enable_sgpr_workgroup_id_x):
450 ///   Number of System SGPR registers: 1. 32 bit work group id in X dimension
451 ///   of grid for wavefront. Always present.
452 ///
453 /// Work-Group Id Y (enable_sgpr_workgroup_id_y):
454 ///   Number of System SGPR registers: 1. 32 bit work group id in Y dimension
455 ///   of grid for wavefront.
456 ///
457 /// Work-Group Id Z (enable_sgpr_workgroup_id_z):
458 ///   Number of System SGPR registers: 1. 32 bit work group id in Z dimension
459 ///   of grid for wavefront. If present then Work-group Id Y will also be
460 ///   present
461 ///
462 /// Work-Group Info (enable_sgpr_workgroup_info):
463 ///   Number of System SGPR registers: 1. {first_wave, 14'b0000,
464 ///   ordered_append_term[10:0], threadgroup_size_in_waves[5:0]}
465 ///
466 /// Private Segment Wave Byte Offset
467 /// (enable_sgpr_private_segment_wave_byte_offset):
468 ///   Number of System SGPR registers: 1. 32 bit byte offset from base of
469 ///   dispatch scratch base. Must be used as an offset with Private/Spill/Arg
470 ///   segment address when using Scratch Segment Buffer. It must be added to
471 ///   Flat Scratch Offset if setting up FLAT SCRATCH for flat addressing.
472 ///
473 ///
474 /// The order of the VGPR registers is defined, but the Finalizer can specify
475 /// which ones are actually setup in the amd_kernel_code_t object using the
476 /// enableVgpr*  bit fields. The register numbers used for enabled registers
477 /// are dense starting at VGPR0: the first enabled register is VGPR0, the next
478 /// enabled register is VGPR1 etc.; disabled registers do not have an VGPR
479 /// number.
480 ///
481 /// VGPR register initial state is defined as follows:
482 ///
483 /// Work-Item Id X (always initialized):
484 ///   Number of registers: 1. 32 bit work item id in X dimension of work-group
485 ///   for wavefront lane.
486 ///
487 /// Work-Item Id X (enable_vgpr_workitem_id > 0):
488 ///   Number of registers: 1. 32 bit work item id in Y dimension of work-group
489 ///   for wavefront lane.
490 ///
491 /// Work-Item Id X (enable_vgpr_workitem_id > 0):
492 ///   Number of registers: 1. 32 bit work item id in Z dimension of work-group
493 ///   for wavefront lane.
494 ///
495 ///
496 /// The setting of registers is being done by existing GPU hardware as follows:
497 ///   1) SGPRs before the Work-Group Ids are set by CP using the 16 User Data
498 ///      registers.
499 ///   2) Work-group Id registers X, Y, Z are set by SPI which supports any
500 ///      combination including none.
501 ///   3) Scratch Wave Offset is also set by SPI which is why its value cannot
502 ///      be added into the value Flat Scratch Offset which would avoid the
503 ///      Finalizer generated prolog having to do the add.
504 ///   4) The VGPRs are set by SPI which only supports specifying either (X),
505 ///      (X, Y) or (X, Y, Z).
506 ///
507 /// Flat Scratch Dispatch Offset and Flat Scratch Size are adjacent SGRRs so
508 /// they can be moved as a 64 bit value to the hardware required SGPRn-3 and
509 /// SGPRn-4 respectively using the Finalizer ?FLAT_SCRATCH? Register.
510 ///
511 /// The global segment can be accessed either using flat operations or buffer
512 /// operations. If buffer operations are used then the Global Buffer used to
513 /// access HSAIL Global/Readonly/Kernarg (which are combine) segments using a
514 /// segment address is not passed into the kernel code by CP since its base
515 /// address is always 0. Instead the Finalizer generates prolog code to
516 /// initialize 4 SGPRs with a V# that has the following properties, and then
517 /// uses that in the buffer instructions:
518 ///   - base address of 0
519 ///   - no swizzle
520 ///   - ATC=1
521 ///   - MTYPE set to support memory coherence specified in
522 ///     amd_kernel_code_t.globalMemoryCoherence
523 ///
524 /// When the Global Buffer is used to access the Kernarg segment, must add the
525 /// dispatch packet kernArgPtr to a kernarg segment address before using this V#.
526 /// Alternatively scalar loads can be used if the kernarg offset is uniform, as
527 /// the kernarg segment is constant for the duration of the kernel execution.
528 ///
529 
530 typedef struct amd_kernel_code_s {
531   uint32_t amd_kernel_code_version_major;
532   uint32_t amd_kernel_code_version_minor;
533   uint16_t amd_machine_kind;
534   uint16_t amd_machine_version_major;
535   uint16_t amd_machine_version_minor;
536   uint16_t amd_machine_version_stepping;
537 
538   /// Byte offset (possibly negative) from start of amd_kernel_code_t
539   /// object to kernel's entry point instruction. The actual code for
540   /// the kernel is required to be 256 byte aligned to match hardware
541   /// requirements (SQ cache line is 16). The code must be position
542   /// independent code (PIC) for AMD devices to give runtime the
543   /// option of copying code to discrete GPU memory or APU L2
544   /// cache. The Finalizer should endeavour to allocate all kernel
545   /// machine code in contiguous memory pages so that a device
546   /// pre-fetcher will tend to only pre-fetch Kernel Code objects,
547   /// improving cache performance.
548   int64_t kernel_code_entry_byte_offset;
549 
550   /// Range of bytes to consider prefetching expressed as an offset
551   /// and size. The offset is from the start (possibly negative) of
552   /// amd_kernel_code_t object. Set both to 0 if no prefetch
553   /// information is available.
554   int64_t kernel_code_prefetch_byte_offset;
555   uint64_t kernel_code_prefetch_byte_size;
556 
557   /// Reserved. Must be 0.
558   uint64_t reserved0;
559 
560   /// Shader program settings for CS. Contains COMPUTE_PGM_RSRC1 and
561   /// COMPUTE_PGM_RSRC2 registers.
562   uint64_t compute_pgm_resource_registers;
563 
564   /// Code properties. See amd_code_property_mask_t for a full list of
565   /// properties.
566   uint32_t code_properties;
567 
568   /// The amount of memory required for the combined private, spill
569   /// and arg segments for a work-item in bytes. If
570   /// is_dynamic_callstack is 1 then additional space must be added to
571   /// this value for the call stack.
572   uint32_t workitem_private_segment_byte_size;
573 
574   /// The amount of group segment memory required by a work-group in
575   /// bytes. This does not include any dynamically allocated group
576   /// segment memory that may be added when the kernel is
577   /// dispatched.
578   uint32_t workgroup_group_segment_byte_size;
579 
580   /// Number of byte of GDS required by kernel dispatch. Must be 0 if
581   /// not using GDS.
582   uint32_t gds_segment_byte_size;
583 
584   /// The size in bytes of the kernarg segment that holds the values
585   /// of the arguments to the kernel. This could be used by CP to
586   /// prefetch the kernarg segment pointed to by the dispatch packet.
587   uint64_t kernarg_segment_byte_size;
588 
589   /// Number of fbarrier's used in the kernel and all functions it
590   /// calls. If the implementation uses group memory to allocate the
591   /// fbarriers then that amount must already be included in the
592   /// workgroup_group_segment_byte_size total.
593   uint32_t workgroup_fbarrier_count;
594 
595   /// Number of scalar registers used by a wavefront. This includes
596   /// the special SGPRs for VCC, Flat Scratch Base, Flat Scratch Size
597   /// and XNACK (for GFX8 (VI)). It does not include the 16 SGPR added if a
598   /// trap handler is enabled. Used to set COMPUTE_PGM_RSRC1.SGPRS.
599   uint16_t wavefront_sgpr_count;
600 
601   /// Number of vector registers used by each work-item. Used to set
602   /// COMPUTE_PGM_RSRC1.VGPRS.
603   uint16_t workitem_vgpr_count;
604 
605   /// If reserved_vgpr_count is 0 then must be 0. Otherwise, this is the
606   /// first fixed VGPR number reserved.
607   uint16_t reserved_vgpr_first;
608 
609   /// The number of consecutive VGPRs reserved by the client. If
610   /// is_debug_supported then this count includes VGPRs reserved
611   /// for debugger use.
612   uint16_t reserved_vgpr_count;
613 
614   /// If reserved_sgpr_count is 0 then must be 0. Otherwise, this is the
615   /// first fixed SGPR number reserved.
616   uint16_t reserved_sgpr_first;
617 
618   /// The number of consecutive SGPRs reserved by the client. If
619   /// is_debug_supported then this count includes SGPRs reserved
620   /// for debugger use.
621   uint16_t reserved_sgpr_count;
622 
623   /// If is_debug_supported is 0 then must be 0. Otherwise, this is the
624   /// fixed SGPR number used to hold the wave scratch offset for the
625   /// entire kernel execution, or uint16_t(-1) if the register is not
626   /// used or not known.
627   uint16_t debug_wavefront_private_segment_offset_sgpr;
628 
629   /// If is_debug_supported is 0 then must be 0. Otherwise, this is the
630   /// fixed SGPR number of the first of 4 SGPRs used to hold the
631   /// scratch V# used for the entire kernel execution, or uint16_t(-1)
632   /// if the registers are not used or not known.
633   uint16_t debug_private_segment_buffer_sgpr;
634 
635   /// The maximum byte alignment of variables used by the kernel in
636   /// the specified memory segment. Expressed as a power of two. Must
637   /// be at least HSA_POWERTWO_16.
638   uint8_t kernarg_segment_alignment;
639   uint8_t group_segment_alignment;
640   uint8_t private_segment_alignment;
641 
642   /// Wavefront size expressed as a power of two. Must be a power of 2
643   /// in range 1..64 inclusive. Used to support runtime query that
644   /// obtains wavefront size, which may be used by application to
645   /// allocated dynamic group memory and set the dispatch work-group
646   /// size.
647   uint8_t wavefront_size;
648 
649   int32_t call_convention;
650   uint8_t reserved3[12];
651   uint64_t runtime_loader_kernel_symbol;
652   uint64_t control_directives[16];
653 } amd_kernel_code_t;
654 
655 #endif // AMDKERNELCODET_H
656