xref: /aosp_15_r20/external/mesa3d/src/amd/common/amd_kernel_code_t.h (revision 6104692788411f58d303aa86923a9ff6ecaded22)
1 /*
2  * Copyright 2015,2016 Advanced Micro Devices, Inc.
3  *
4  * SPDX-License-Identifier: MIT
5  */
6 
7 #ifndef AMDKERNELCODET_H
8 #define AMDKERNELCODET_H
9 
10 //---------------------------------------------------------------------------//
11 // AMD Kernel Code, and its dependencies                                     //
12 //---------------------------------------------------------------------------//
13 
14 // Sets val bits for specified mask in specified dst packed instance.
15 #define AMD_HSA_BITS_SET(dst, mask, val)                                                           \
16    dst &= (~(1 << mask##_SHIFT) & ~mask);                                                          \
17    dst |= (((val) << mask##_SHIFT) & mask)
18 
19 // Gets bits for specified mask from specified src packed instance.
20 #define AMD_HSA_BITS_GET(src, mask) ((src & mask) >> mask##_SHIFT)
21 
22 /* Every amd_*_code_t has the following properties, which are composed of
23  * a number of bit fields. Every bit field has a mask (AMD_CODE_PROPERTY_*),
24  * bit width (AMD_CODE_PROPERTY_*_WIDTH, and bit shift amount
25  * (AMD_CODE_PROPERTY_*_SHIFT) for convenient access. Unused bits must be 0.
26  *
27  * (Note that bit fields cannot be used as their layout is
28  * implementation defined in the C standard and so cannot be used to
29  * specify an ABI)
30  */
31 enum amd_code_property_mask_t
32 {
33 
34    /* Enable the setup of the SGPR user data registers
35     * (AMD_CODE_PROPERTY_ENABLE_SGPR_*), see documentation of amd_kernel_code_t
36     * for initial register state.
37     *
38     * The total number of SGPRuser data registers requested must not
39     * exceed 16. Any requests beyond 16 will be ignored.
40     *
41     * Used to set COMPUTE_PGM_RSRC2.USER_SGPR (set to total count of
42     * SGPR user data registers enabled up to 16).
43     */
44 
45    AMD_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_BUFFER_SHIFT = 0,
46    AMD_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_BUFFER_WIDTH = 1,
47    AMD_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_BUFFER =
48       ((1 << AMD_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_BUFFER_WIDTH) - 1)
49       << AMD_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_BUFFER_SHIFT,
50 
51    AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_PTR_SHIFT = 1,
52    AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_PTR_WIDTH = 1,
53    AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_PTR =
54       ((1 << AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_PTR_WIDTH) - 1)
55       << AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_PTR_SHIFT,
56 
57    AMD_CODE_PROPERTY_ENABLE_SGPR_QUEUE_PTR_SHIFT = 2,
58    AMD_CODE_PROPERTY_ENABLE_SGPR_QUEUE_PTR_WIDTH = 1,
59    AMD_CODE_PROPERTY_ENABLE_SGPR_QUEUE_PTR =
60       ((1 << AMD_CODE_PROPERTY_ENABLE_SGPR_QUEUE_PTR_WIDTH) - 1)
61       << AMD_CODE_PROPERTY_ENABLE_SGPR_QUEUE_PTR_SHIFT,
62 
63    AMD_CODE_PROPERTY_ENABLE_SGPR_KERNARG_SEGMENT_PTR_SHIFT = 3,
64    AMD_CODE_PROPERTY_ENABLE_SGPR_KERNARG_SEGMENT_PTR_WIDTH = 1,
65    AMD_CODE_PROPERTY_ENABLE_SGPR_KERNARG_SEGMENT_PTR =
66       ((1 << AMD_CODE_PROPERTY_ENABLE_SGPR_KERNARG_SEGMENT_PTR_WIDTH) - 1)
67       << AMD_CODE_PROPERTY_ENABLE_SGPR_KERNARG_SEGMENT_PTR_SHIFT,
68 
69    AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_ID_SHIFT = 4,
70    AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_ID_WIDTH = 1,
71    AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_ID =
72       ((1 << AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_ID_WIDTH) - 1)
73       << AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_ID_SHIFT,
74 
75    AMD_CODE_PROPERTY_ENABLE_SGPR_FLAT_SCRATCH_INIT_SHIFT = 5,
76    AMD_CODE_PROPERTY_ENABLE_SGPR_FLAT_SCRATCH_INIT_WIDTH = 1,
77    AMD_CODE_PROPERTY_ENABLE_SGPR_FLAT_SCRATCH_INIT =
78       ((1 << AMD_CODE_PROPERTY_ENABLE_SGPR_FLAT_SCRATCH_INIT_WIDTH) - 1)
79       << AMD_CODE_PROPERTY_ENABLE_SGPR_FLAT_SCRATCH_INIT_SHIFT,
80 
81    AMD_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_SIZE_SHIFT = 6,
82    AMD_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_SIZE_WIDTH = 1,
83    AMD_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_SIZE =
84       ((1 << AMD_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_SIZE_WIDTH) - 1)
85       << AMD_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_SIZE_SHIFT,
86 
87    AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_X_SHIFT = 7,
88    AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_X_WIDTH = 1,
89    AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_X =
90       ((1 << AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_X_WIDTH) - 1)
91       << AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_X_SHIFT,
92 
93    AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_Y_SHIFT = 8,
94    AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_Y_WIDTH = 1,
95    AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_Y =
96       ((1 << AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_Y_WIDTH) - 1)
97       << AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_Y_SHIFT,
98 
99    AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_Z_SHIFT = 9,
100    AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_Z_WIDTH = 1,
101    AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_Z =
102       ((1 << AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_Z_WIDTH) - 1)
103       << AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_Z_SHIFT,
104 
105    AMD_CODE_PROPERTY_ENABLE_WAVEFRONT_SIZE32_SHIFT = 10,
106    AMD_CODE_PROPERTY_ENABLE_WAVEFRONT_SIZE32_WIDTH = 1,
107    AMD_CODE_PROPERTY_ENABLE_WAVEFRONT_SIZE32 =
108       ((1 << AMD_CODE_PROPERTY_ENABLE_WAVEFRONT_SIZE32_WIDTH) - 1)
109       << AMD_CODE_PROPERTY_ENABLE_WAVEFRONT_SIZE32_SHIFT,
110 
111    AMD_CODE_PROPERTY_RESERVED1_SHIFT = 11,
112    AMD_CODE_PROPERTY_RESERVED1_WIDTH = 5,
113    AMD_CODE_PROPERTY_RESERVED1 = ((1 << AMD_CODE_PROPERTY_RESERVED1_WIDTH) - 1)
114       << AMD_CODE_PROPERTY_RESERVED1_SHIFT,
115 
116    /* Control wave ID base counter for GDS ordered-append. Used to set
117     * COMPUTE_DISPATCH_INITIATOR.ORDERED_APPEND_ENBL. (Not sure if
118     * ORDERED_APPEND_MODE also needs to be settable)
119     */
120    AMD_CODE_PROPERTY_ENABLE_ORDERED_APPEND_GDS_SHIFT = 16,
121    AMD_CODE_PROPERTY_ENABLE_ORDERED_APPEND_GDS_WIDTH = 1,
122    AMD_CODE_PROPERTY_ENABLE_ORDERED_APPEND_GDS =
123       ((1 << AMD_CODE_PROPERTY_ENABLE_ORDERED_APPEND_GDS_WIDTH) - 1)
124       << AMD_CODE_PROPERTY_ENABLE_ORDERED_APPEND_GDS_SHIFT,
125 
126    /* The interleave (swizzle) element size in bytes required by the
127     * code for private memory. This must be 2, 4, 8 or 16. This value
128     * is provided to the finalizer when it is invoked and is recorded
129     * here. The hardware will interleave the memory requests of each
130     * lane of a wavefront by this element size to ensure each
131     * work-item gets a distinct memory memory location. Therefore, the
132     * finalizer ensures that all load and store operations done to
133     * private memory do not exceed this size. For example, if the
134     * element size is 4 (32-bits or dword) and a 64-bit value must be
135     * loaded, the finalizer will generate two 32-bit loads. This
136     * ensures that the interleaving will get the work-item
137     * specific dword for both halves of the 64-bit value. If it just
138     * did a 64-bit load then it would get one dword which belonged to
139     * its own work-item, but the second dword would belong to the
140     * adjacent lane work-item since the interleaving is in dwords.
141     *
142     * The value used must match the value that the runtime configures
143     * the GPU flat scratch (SH_STATIC_MEM_CONFIG.ELEMENT_SIZE). This
144     * is generally DWORD.
145     *
146     * USE VALUES FROM THE AMD_ELEMENT_BYTE_SIZE_T ENUM.
147     */
148    AMD_CODE_PROPERTY_PRIVATE_ELEMENT_SIZE_SHIFT = 17,
149    AMD_CODE_PROPERTY_PRIVATE_ELEMENT_SIZE_WIDTH = 2,
150    AMD_CODE_PROPERTY_PRIVATE_ELEMENT_SIZE =
151       ((1 << AMD_CODE_PROPERTY_PRIVATE_ELEMENT_SIZE_WIDTH) - 1)
152       << AMD_CODE_PROPERTY_PRIVATE_ELEMENT_SIZE_SHIFT,
153 
154    /* Are global memory addresses 64 bits. Must match
155     * amd_kernel_code_t.hsail_machine_model ==
156     * HSA_MACHINE_LARGE. Must also match
157     * SH_MEM_CONFIG.PTR32 (GFX6 (SI)/GFX7 (CI)),
158     * SH_MEM_CONFIG.ADDRESS_MODE (GFX8 (VI)+).
159     */
160    AMD_CODE_PROPERTY_IS_PTR64_SHIFT = 19,
161    AMD_CODE_PROPERTY_IS_PTR64_WIDTH = 1,
162    AMD_CODE_PROPERTY_IS_PTR64 = ((1 << AMD_CODE_PROPERTY_IS_PTR64_WIDTH) - 1)
163                                 << AMD_CODE_PROPERTY_IS_PTR64_SHIFT,
164 
165    /* Indicate if the generated ISA is using a dynamically sized call
166     * stack. This can happen if calls are implemented using a call
167     * stack and recursion, alloca or calls to indirect functions are
168     * present. In these cases the Finalizer cannot compute the total
169     * private segment size at compile time. In this case the
170     * workitem_private_segment_byte_size only specifies the statically
171     * know private segment size, and additional space must be added
172     * for the call stack.
173     */
174    AMD_CODE_PROPERTY_IS_DYNAMIC_CALLSTACK_SHIFT = 20,
175    AMD_CODE_PROPERTY_IS_DYNAMIC_CALLSTACK_WIDTH = 1,
176    AMD_CODE_PROPERTY_IS_DYNAMIC_CALLSTACK =
177       ((1 << AMD_CODE_PROPERTY_IS_DYNAMIC_CALLSTACK_WIDTH) - 1)
178       << AMD_CODE_PROPERTY_IS_DYNAMIC_CALLSTACK_SHIFT,
179 
180    /* Indicate if code generated has support for debugging. */
181    AMD_CODE_PROPERTY_IS_DEBUG_SUPPORTED_SHIFT = 21,
182    AMD_CODE_PROPERTY_IS_DEBUG_SUPPORTED_WIDTH = 1,
183    AMD_CODE_PROPERTY_IS_DEBUG_SUPPORTED = ((1 << AMD_CODE_PROPERTY_IS_DEBUG_SUPPORTED_WIDTH) - 1)
184                                           << AMD_CODE_PROPERTY_IS_DEBUG_SUPPORTED_SHIFT,
185 
186    AMD_CODE_PROPERTY_IS_XNACK_SUPPORTED_SHIFT = 22,
187    AMD_CODE_PROPERTY_IS_XNACK_SUPPORTED_WIDTH = 1,
188    AMD_CODE_PROPERTY_IS_XNACK_SUPPORTED = ((1 << AMD_CODE_PROPERTY_IS_XNACK_SUPPORTED_WIDTH) - 1)
189                                           << AMD_CODE_PROPERTY_IS_XNACK_SUPPORTED_SHIFT,
190 
191    AMD_CODE_PROPERTY_RESERVED2_SHIFT = 23,
192    AMD_CODE_PROPERTY_RESERVED2_WIDTH = 9,
193    AMD_CODE_PROPERTY_RESERVED2 = ((1 << AMD_CODE_PROPERTY_RESERVED2_WIDTH) - 1)
194                                  << AMD_CODE_PROPERTY_RESERVED2_SHIFT
195 };
196 
197 /* AMD Kernel Code Object (amd_kernel_code_t). GPU CP uses the AMD Kernel
198  * Code Object to set up the hardware to execute the kernel dispatch.
199  *
200  * Initial Kernel Register State.
201  *
202  * Initial kernel register state will be set up by CP/SPI prior to the start
203  * of execution of every wavefront. This is limited by the constraints of the
204  * current hardware.
205  *
206  * The order of the SGPR registers is defined, but the Finalizer can specify
207  * which ones are actually setup in the amd_kernel_code_t object using the
208  * enable_sgpr_* bit fields. The register numbers used for enabled registers
209  * are dense starting at SGPR0: the first enabled register is SGPR0, the next
210  * enabled register is SGPR1 etc.; disabled registers do not have an SGPR
211  * number.
212  *
213  * The initial SGPRs comprise up to 16 User SRGPs that are set up by CP and
214  * apply to all waves of the grid. It is possible to specify more than 16 User
215  * SGPRs using the enable_sgpr_* bit fields, in which case only the first 16
216  * are actually initialized. These are then immediately followed by the System
217  * SGPRs that are set up by ADC/SPI and can have different values for each wave
218  * of the grid dispatch.
219  *
220  * SGPR register initial state is defined as follows:
221  *
222  * Private Segment Buffer (enable_sgpr_private_segment_buffer):
223  *   Number of User SGPR registers: 4. V# that can be used, together with
224  *   Scratch Wave Offset as an offset, to access the Private/Spill/Arg
225  *   segments using a segment address. It must be set as follows:
226  *     - Base address: of the scratch memory area used by the dispatch. It
227  *       does not include the scratch wave offset. It will be the per process
228  *       SH_HIDDEN_PRIVATE_BASE_VMID plus any offset from this dispatch (for
229  *       example there may be a per pipe offset, or per AQL Queue offset).
230  *     - Stride + data_format: Element Size * Index Stride (???)
231  *     - Cache swizzle: ???
232  *     - Swizzle enable: SH_STATIC_MEM_CONFIG.SWIZZLE_ENABLE (must be 1 for
233  *       scratch)
234  *     - Num records: Flat Scratch Work Item Size / Element Size (???)
235  *     - Dst_sel_*: ???
236  *     - Num_format: ???
237  *     - Element_size: SH_STATIC_MEM_CONFIG.ELEMENT_SIZE (will be DWORD, must
238  *       agree with amd_kernel_code_t.privateElementSize)
239  *     - Index_stride: SH_STATIC_MEM_CONFIG.INDEX_STRIDE (will be 64 as must
240  *       be number of wavefront lanes for scratch, must agree with
241  *       amd_kernel_code_t.wavefrontSize)
242  *     - Add tid enable: 1
243  *     - ATC: from SH_MEM_CONFIG.PRIVATE_ATC,
244  *     - Hash_enable: ???
245  *     - Heap: ???
246  *     - Mtype: from SH_STATIC_MEM_CONFIG.PRIVATE_MTYPE
247  *     - Type: 0 (a buffer) (???)
248  *
249  * Dispatch Ptr (enable_sgpr_dispatch_ptr):
250  *   Number of User SGPR registers: 2. 64 bit address of AQL dispatch packet
251  *   for kernel actually executing.
252  *
253  * Queue Ptr (enable_sgpr_queue_ptr):
254  *   Number of User SGPR registers: 2. 64 bit address of AmdQueue object for
255  *   AQL queue on which the dispatch packet was queued.
256  *
257  * Kernarg Segment Ptr (enable_sgpr_kernarg_segment_ptr):
258  *   Number of User SGPR registers: 2. 64 bit address of Kernarg segment. This
259  *   is directly copied from the kernargPtr in the dispatch packet. Having CP
260  *   load it once avoids loading it at the beginning of every wavefront.
261  *
262  * Dispatch Id (enable_sgpr_dispatch_id):
263  *   Number of User SGPR registers: 2. 64 bit Dispatch ID of the dispatch
264  *   packet being executed.
265  *
266  * Flat Scratch Init (enable_sgpr_flat_scratch_init):
267  *   Number of User SGPR registers: 2. This is 2 SGPRs.
268  *
269  *   For CI/VI:
270  *     The first SGPR is a 32 bit byte offset from SH_MEM_HIDDEN_PRIVATE_BASE
271  *     to base of memory for scratch for this dispatch. This is the same offset
272  *     used in computing the Scratch Segment Buffer base address. The value of
273  *     Scratch Wave Offset must be added by the kernel code and moved to
274  *     SGPRn-4 for use as the FLAT SCRATCH BASE in flat memory instructions.
275  *
276  *     The second SGPR is 32 bit byte size of a single work-item's scratch
277  *     memory usage. This is directly loaded from the dispatch packet Private
278  *     Segment Byte Size and rounded up to a multiple of DWORD.
279  *
280  *     \todo [Does CP need to round this to >4 byte alignment?]
281  *
282  *     The kernel code must move to SGPRn-3 for use as the FLAT SCRATCH SIZE in
283  *     flat memory instructions. Having CP load it once avoids loading it at
284  *     the beginning of every wavefront.
285  *
286  * Private Segment Size (enable_sgpr_private_segment_size):
287  *   Number of User SGPR registers: 1. The 32 bit byte size of a single
288  *   work-item's scratch memory allocation. This is the value from the dispatch
289  *   packet. Private Segment Byte Size rounded up by CP to a multiple of DWORD.
290  *
291  *   \todo [Does CP need to round this to >4 byte alignment?]
292  *
293  *   Having CP load it once avoids loading it at the beginning of every
294  *   wavefront.
295  *
296  *   \todo [This will not be used for CI/VI since it is the same value as
297  *   the second SGPR of Flat Scratch Init.
298  *
299  * Grid Work-Group Count X (enable_sgpr_grid_workgroup_count_x):
300  *   Number of User SGPR registers: 1. 32 bit count of the number of
301  *   work-groups in the X dimension for the grid being executed. Computed from
302  *   the fields in the HsaDispatchPacket as
303  *   ((gridSize.x+workgroupSize.x-1)/workgroupSize.x).
304  *
305  * Grid Work-Group Count Y (enable_sgpr_grid_workgroup_count_y):
306  *   Number of User SGPR registers: 1. 32 bit count of the number of
307  *   work-groups in the Y dimension for the grid being executed. Computed from
308  *   the fields in the HsaDispatchPacket as
309  *   ((gridSize.y+workgroupSize.y-1)/workgroupSize.y).
310  *
311  *   Only initialized if <16 previous SGPRs initialized.
312  *
313  * Grid Work-Group Count Z (enable_sgpr_grid_workgroup_count_z):
314  *   Number of User SGPR registers: 1. 32 bit count of the number of
315  *   work-groups in the Z dimension for the grid being executed. Computed
316  *   from the fields in the HsaDispatchPacket as
317  *   ((gridSize.z+workgroupSize.z-1)/workgroupSize.z).
318  *
319  *   Only initialized if <16 previous SGPRs initialized.
320  *
321  * Work-Group Id X (enable_sgpr_workgroup_id_x):
322  *   Number of System SGPR registers: 1. 32 bit work group id in X dimension
323  *   of grid for wavefront. Always present.
324  *
325  * Work-Group Id Y (enable_sgpr_workgroup_id_y):
326  *   Number of System SGPR registers: 1. 32 bit work group id in Y dimension
327  *   of grid for wavefront.
328  *
329  * Work-Group Id Z (enable_sgpr_workgroup_id_z):
330  *   Number of System SGPR registers: 1. 32 bit work group id in Z dimension
331  *   of grid for wavefront. If present then Work-group Id Y will also be
332  *   present
333  *
334  * Work-Group Info (enable_sgpr_workgroup_info):
335  *   Number of System SGPR registers: 1. {first_wave, 14'b0000,
336  *   ordered_append_term[10:0], threadgroup_size_in_waves[5:0]}
337  *
338  * Private Segment Wave Byte Offset
339  * (enable_sgpr_private_segment_wave_byte_offset):
340  *   Number of System SGPR registers: 1. 32 bit byte offset from base of
341  *   dispatch scratch base. Must be used as an offset with Private/Spill/Arg
342  *   segment address when using Scratch Segment Buffer. It must be added to
343  *   Flat Scratch Offset if setting up FLAT SCRATCH for flat addressing.
344  *
345  *
346  * The order of the VGPR registers is defined, but the Finalizer can specify
347  * which ones are actually setup in the amd_kernel_code_t object using the
348  * enableVgpr*  bit fields. The register numbers used for enabled registers
349  * are dense starting at VGPR0: the first enabled register is VGPR0, the next
350  * enabled register is VGPR1 etc.; disabled registers do not have an VGPR
351  * number.
352  *
353  * VGPR register initial state is defined as follows:
354  *
355  * Work-Item Id X (always initialized):
356  *   Number of registers: 1. 32 bit work item id in X dimension of work-group
357  *   for wavefront lane.
358  *
359  * Work-Item Id X (enable_vgpr_workitem_id > 0):
360  *   Number of registers: 1. 32 bit work item id in Y dimension of work-group
361  *   for wavefront lane.
362  *
363  * Work-Item Id X (enable_vgpr_workitem_id > 0):
364  *   Number of registers: 1. 32 bit work item id in Z dimension of work-group
365  *   for wavefront lane.
366  *
367  *
368  * The setting of registers is being done by existing GPU hardware as follows:
369  *   1) SGPRs before the Work-Group Ids are set by CP using the 16 User Data
370  *      registers.
371  *   2) Work-group Id registers X, Y, Z are set by SPI which supports any
372  *      combination including none.
373  *   3) Scratch Wave Offset is also set by SPI which is why its value cannot
374  *      be added into the value Flat Scratch Offset which would avoid the
375  *      Finalizer generated prolog having to do the add.
376  *   4) The VGPRs are set by SPI which only supports specifying either (X),
377  *      (X, Y) or (X, Y, Z).
378  *
379  * Flat Scratch Dispatch Offset and Flat Scratch Size are adjacent SGRRs so
380  * they can be moved as a 64 bit value to the hardware required SGPRn-3 and
381  * SGPRn-4 respectively using the Finalizer ?FLAT_SCRATCH? Register.
382  *
383  * The global segment can be accessed either using flat operations or buffer
384  * operations. If buffer operations are used then the Global Buffer used to
385  * access HSAIL Global/Readonly/Kernarg (which are combine) segments using a
386  * segment address is not passed into the kernel code by CP since its base
387  * address is always 0. Instead the Finalizer generates prolog code to
388  * initialize 4 SGPRs with a V# that has the following properties, and then
389  * uses that in the buffer instructions:
390  *   - base address of 0
391  *   - no swizzle
392  *   - ATC=1
393  *   - MTYPE set to support memory coherence specified in
394  *     amd_kernel_code_t.globalMemoryCoherence
395  *
396  * When the Global Buffer is used to access the Kernarg segment, must add the
397  * dispatch packet kernArgPtr to a kernarg segment address before using this V#.
398  * Alternatively scalar loads can be used if the kernarg offset is uniform, as
399  * the kernarg segment is constant for the duration of the kernel execution.
400  */
401 
402 typedef struct amd_kernel_code_s {
403    uint32_t amd_kernel_code_version_major;
404    uint32_t amd_kernel_code_version_minor;
405    uint16_t amd_machine_kind;
406    uint16_t amd_machine_version_major;
407    uint16_t amd_machine_version_minor;
408    uint16_t amd_machine_version_stepping;
409 
410    /* Byte offset (possibly negative) from start of amd_kernel_code_t
411     * object to kernel's entry point instruction. The actual code for
412     * the kernel is required to be 256 byte aligned to match hardware
413     * requirements (SQ cache line is 16). The code must be position
414     * independent code (PIC) for AMD devices to give runtime the
415     * option of copying code to discrete GPU memory or APU L2
416     * cache. The Finalizer should endeavour to allocate all kernel
417     * machine code in contiguous memory pages so that a device
418     * pre-fetcher will tend to only pre-fetch Kernel Code objects,
419     * improving cache performance.
420     */
421    int64_t kernel_code_entry_byte_offset;
422 
423    /* Range of bytes to consider prefetching expressed as an offset
424     * and size. The offset is from the start (possibly negative) of
425     * amd_kernel_code_t object. Set both to 0 if no prefetch
426     * information is available.
427     */
428    int64_t kernel_code_prefetch_byte_offset;
429    uint64_t kernel_code_prefetch_byte_size;
430 
431    /* Number of bytes of scratch backing memory required for full
432     * occupancy of target chip. This takes into account the number of
433     * bytes of scratch per work-item, the wavefront size, the maximum
434     * number of wavefronts per CU, and the number of CUs. This is an
435     * upper limit on scratch. If the grid being dispatched is small it
436     * may only need less than this. If the kernel uses no scratch, or
437     * the Finalizer has not computed this value, it must be 0.
438     */
439    uint64_t max_scratch_backing_memory_byte_size;
440 
441    /* Shader program settings for CS. Contains COMPUTE_PGM_RSRC1 and
442     * COMPUTE_PGM_RSRC2 registers.
443     */
444    uint64_t compute_pgm_resource_registers;
445 
446    /* Code properties. See amd_code_property_mask_t for a full list of
447     * properties.
448     */
449    uint32_t code_properties;
450 
451    /* The amount of memory required for the combined private, spill
452     * and arg segments for a work-item in bytes. If
453     * is_dynamic_callstack is 1 then additional space must be added to
454     * this value for the call stack.
455     */
456    uint32_t workitem_private_segment_byte_size;
457 
458    /* The amount of group segment memory required by a work-group in
459     * bytes. This does not include any dynamically allocated group
460     * segment memory that may be added when the kernel is
461     * dispatched.
462     */
463    uint32_t workgroup_group_segment_byte_size;
464 
465    /* Number of byte of GDS required by kernel dispatch. Must be 0 if
466     * not using GDS.
467     */
468    uint32_t gds_segment_byte_size;
469 
470    /* The size in bytes of the kernarg segment that holds the values
471     * of the arguments to the kernel. This could be used by CP to
472     * prefetch the kernarg segment pointed to by the dispatch packet.
473     */
474    uint64_t kernarg_segment_byte_size;
475 
476    /* Number of fbarrier's used in the kernel and all functions it
477     * calls. If the implementation uses group memory to allocate the
478     * fbarriers then that amount must already be included in the
479     * workgroup_group_segment_byte_size total.
480     */
481    uint32_t workgroup_fbarrier_count;
482 
483    /* Number of scalar registers used by a wavefront. This includes
484     * the special SGPRs for VCC, Flat Scratch Base, Flat Scratch Size
485     * and XNACK (for GFX8 (VI)). It does not include the 16 SGPR added if a
486     * trap handler is enabled. Used to set COMPUTE_PGM_RSRC1.SGPRS.
487     */
488    uint16_t wavefront_sgpr_count;
489 
490    /* Number of vector registers used by each work-item. Used to set
491     * COMPUTE_PGM_RSRC1.VGPRS.
492     */
493    uint16_t workitem_vgpr_count;
494 
495    /* If reserved_vgpr_count is 0 then must be 0. Otherwise, this is the
496     * first fixed VGPR number reserved.
497     */
498    uint16_t reserved_vgpr_first;
499 
500    /* The number of consecutive VGPRs reserved by the client. If
501     * is_debug_supported then this count includes VGPRs reserved
502     * for debugger use.
503     */
504    uint16_t reserved_vgpr_count;
505 
506    /* If reserved_sgpr_count is 0 then must be 0. Otherwise, this is the
507     * first fixed SGPR number reserved.
508     */
509    uint16_t reserved_sgpr_first;
510 
511    /* The number of consecutive SGPRs reserved by the client. If
512     * is_debug_supported then this count includes SGPRs reserved
513     * for debugger use.
514     */
515    uint16_t reserved_sgpr_count;
516 
517    /* If is_debug_supported is 0 then must be 0. Otherwise, this is the
518     * fixed SGPR number used to hold the wave scratch offset for the
519     * entire kernel execution, or uint16_t(-1) if the register is not
520     * used or not known.
521     */
522    uint16_t debug_wavefront_private_segment_offset_sgpr;
523 
524    /* If is_debug_supported is 0 then must be 0. Otherwise, this is the
525     * fixed SGPR number of the first of 4 SGPRs used to hold the
526     * scratch V# used for the entire kernel execution, or uint16_t(-1)
527     * if the registers are not used or not known.
528     */
529    uint16_t debug_private_segment_buffer_sgpr;
530 
531    /* The maximum byte alignment of variables used by the kernel in
532     * the specified memory segment. Expressed as a power of two. Must
533     * be at least HSA_POWERTWO_16.
534     */
535    uint8_t kernarg_segment_alignment;
536    uint8_t group_segment_alignment;
537    uint8_t private_segment_alignment;
538 
539    /* Wavefront size expressed as a power of two. Must be a power of 2
540     * in range 1..64 inclusive. Used to support runtime query that
541     * obtains wavefront size, which may be used by application to
542     * allocated dynamic group memory and set the dispatch work-group
543     * size.
544     */
545    uint8_t wavefront_size;
546 
547    int32_t call_convention;
548    uint8_t reserved3[12];
549    uint64_t runtime_loader_kernel_symbol;
550    uint64_t control_directives[16];
551 } amd_kernel_code_t;
552 
553 #endif // AMDKERNELCODET_H
554