| blob | c60d60af5a82cf1204fb9854816f2762ff8c2ecc |
1 =============================
2 User Guide for AMDGPU Backend
3 =============================
5 .. contents::
6 :local:
8 Introduction
9 ============
11 The AMDGPU backend provides ISA code generation for AMD GPUs, starting with the
12 R600 family up until the current GCN families. It lives in the
13 ``lib/Target/AMDGPU`` directory.
15 LLVM
16 ====
18 .. _amdgpu-target-triples:
20 Target Triples
21 --------------
23 Use the ``clang -target <Architecture>-<Vendor>-<OS>-<Environment>`` option to
24 specify the target triple:
26 .. table:: AMDGPU Architectures
27 :name: amdgpu-architecture-table
29 ============ ==============================================================
30 Architecture Description
31 ============ ==============================================================
32 ``r600`` AMD GPUs HD2XXX-HD6XXX for graphics and compute shaders.
33 ``amdgcn`` AMD GPUs GCN GFX6 onwards for graphics and compute shaders.
34 ============ ==============================================================
36 .. table:: AMDGPU Vendors
37 :name: amdgpu-vendor-table
39 ============ ==============================================================
40 Vendor Description
41 ============ ==============================================================
42 ``amd`` Can be used for all AMD GPU usage.
43 ``mesa3d`` Can be used if the OS is ``mesa3d``.
44 ============ ==============================================================
46 .. table:: AMDGPU Operating Systems
47 :name: amdgpu-os-table
49 ============== ============================================================
50 OS Description
51 ============== ============================================================
52 *<empty>* Defaults to the *unknown* OS.
53 ``amdhsa`` Compute kernels executed on HSA [HSA]_ compatible runtimes
54 such as AMD's ROCm [AMD-ROCm]_.
55 ``amdpal`` Graphic shaders and compute kernels executed on AMD PAL
56 runtime.
57 ``mesa3d`` Graphic shaders and compute kernels executed on Mesa 3D
58 runtime.
59 ============== ============================================================
61 .. table:: AMDGPU Environments
62 :name: amdgpu-environment-table
64 ============ ==============================================================
65 Environment Description
66 ============ ==============================================================
67 *<empty>* Default.
68 ============ ==============================================================
70 .. _amdgpu-processors:
72 Processors
73 ----------
75 Use the ``clang -mcpu <Processor>`` option to specify the AMD GPU processor. The
76 names from both the *Processor* and *Alternative Processor* can be used.
78 .. table:: AMDGPU Processors
79 :name: amdgpu-processor-table
81 =========== =============== ============ ===== ========== ======= ======================
82 Processor Alternative Target dGPU/ Target ROCm Example
83 Processor Triple APU Features Support Products
84 Architecture Supported
85 [Default]
86 =========== =============== ============ ===== ========== ======= ======================
87 **Radeon HD 2000/3000 Series (R600)** [AMD-RADEON-HD-2000-3000]_
88 ----------------------------------------------------------------------------------------
89 ``r600`` ``r600`` dGPU
90 ``r630`` ``r600`` dGPU
91 ``rs880`` ``r600`` dGPU
92 ``rv670`` ``r600`` dGPU
93 **Radeon HD 4000 Series (R700)** [AMD-RADEON-HD-4000]_
94 ----------------------------------------------------------------------------------------
95 ``rv710`` ``r600`` dGPU
96 ``rv730`` ``r600`` dGPU
97 ``rv770`` ``r600`` dGPU
98 **Radeon HD 5000 Series (Evergreen)** [AMD-RADEON-HD-5000]_
99 ----------------------------------------------------------------------------------------
100 ``cedar`` ``r600`` dGPU
101 ``cypress`` ``r600`` dGPU
102 ``juniper`` ``r600`` dGPU
103 ``redwood`` ``r600`` dGPU
104 ``sumo`` ``r600`` dGPU
105 **Radeon HD 6000 Series (Northern Islands)** [AMD-RADEON-HD-6000]_
106 ----------------------------------------------------------------------------------------
107 ``barts`` ``r600`` dGPU
108 ``caicos`` ``r600`` dGPU
109 ``cayman`` ``r600`` dGPU
110 ``turks`` ``r600`` dGPU
111 **GCN GFX6 (Southern Islands (SI))** [AMD-GCN-GFX6]_
112 ----------------------------------------------------------------------------------------
113 ``gfx600`` - ``tahiti`` ``amdgcn`` dGPU
114 ``gfx601`` - ``hainan`` ``amdgcn`` dGPU
115 - ``oland``
116 - ``pitcairn``
117 - ``verde``
118 **GCN GFX7 (Sea Islands (CI))** [AMD-GCN-GFX7]_
119 ----------------------------------------------------------------------------------------
120 ``gfx700`` - ``kaveri`` ``amdgcn`` APU - A6-7000
121 - A6 Pro-7050B
122 - A8-7100
123 - A8 Pro-7150B
124 - A10-7300
125 - A10 Pro-7350B
126 - FX-7500
127 - A8-7200P
128 - A10-7400P
129 - FX-7600P
130 ``gfx701`` - ``hawaii`` ``amdgcn`` dGPU ROCm - FirePro W8100
131 - FirePro W9100
132 - FirePro S9150
133 - FirePro S9170
134 ``gfx702`` ``amdgcn`` dGPU ROCm - Radeon R9 290
135 - Radeon R9 290x
136 - Radeon R390
137 - Radeon R390x
138 ``gfx703`` - ``kabini`` ``amdgcn`` APU - E1-2100
139 - ``mullins`` - E1-2200
140 - E1-2500
141 - E2-3000
142 - E2-3800
143 - A4-5000
144 - A4-5100
145 - A6-5200
146 - A4 Pro-3340B
147 ``gfx704`` - ``bonaire`` ``amdgcn`` dGPU - Radeon HD 7790
148 - Radeon HD 8770
149 - R7 260
150 - R7 260X
151 **GCN GFX8 (Volcanic Islands (VI))** [AMD-GCN-GFX8]_
152 ----------------------------------------------------------------------------------------
153 ``gfx801`` - ``carrizo`` ``amdgcn`` APU - xnack - A6-8500P
154 [on] - Pro A6-8500B
155 - A8-8600P
156 - Pro A8-8600B
157 - FX-8800P
158 - Pro A12-8800B
159 \ ``amdgcn`` APU - xnack ROCm - A10-8700P
160 [on] - Pro A10-8700B
161 - A10-8780P
162 \ ``amdgcn`` APU - xnack - A10-9600P
163 [on] - A10-9630P
164 - A12-9700P
165 - A12-9730P
166 - FX-9800P
167 - FX-9830P
168 \ ``amdgcn`` APU - xnack - E2-9010
169 [on] - A6-9210
170 - A9-9410
171 ``gfx802`` - ``iceland`` ``amdgcn`` dGPU - xnack ROCm - FirePro S7150
172 - ``tonga`` [off] - FirePro S7100
173 - FirePro W7100
174 - Radeon R285
175 - Radeon R9 380
176 - Radeon R9 385
177 - Mobile FirePro
178 M7170
179 ``gfx803`` - ``fiji`` ``amdgcn`` dGPU - xnack ROCm - Radeon R9 Nano
180 [off] - Radeon R9 Fury
181 - Radeon R9 FuryX
182 - Radeon Pro Duo
183 - FirePro S9300x2
184 - Radeon Instinct MI8
185 \ - ``polaris10`` ``amdgcn`` dGPU - xnack ROCm - Radeon RX 470
186 [off] - Radeon RX 480
187 - Radeon Instinct MI6
188 \ - ``polaris11`` ``amdgcn`` dGPU - xnack ROCm - Radeon RX 460
189 [off]
190 ``gfx810`` - ``stoney`` ``amdgcn`` APU - xnack
191 [on]
192 **GCN GFX9** [AMD-GCN-GFX9]_
193 ----------------------------------------------------------------------------------------
194 ``gfx900`` ``amdgcn`` dGPU - xnack ROCm - Radeon Vega
195 [off] Frontier Edition
196 - Radeon RX Vega 56
197 - Radeon RX Vega 64
198 - Radeon RX Vega 64
199 Liquid
200 - Radeon Instinct MI25
201 ``gfx902`` ``amdgcn`` APU - xnack - Ryzen 3 2200G
202 [on] - Ryzen 5 2400G
203 ``gfx904`` ``amdgcn`` dGPU - xnack *TBA*
204 [off]
205 .. TODO
206 Add product
207 names.
208 ``gfx906`` ``amdgcn`` dGPU - xnack - Radeon Instinct MI50
209 [off] - Radeon Instinct MI60
210 sram-ecc
211 [on]
212 ``gfx909`` ``amdgcn`` APU - xnack *TBA* (Raven Ridge 2)
213 [on]
214 .. TODO
215 Add product
216 names.
217 =========== =============== ============ ===== ========== ======= ======================
219 .. _amdgpu-target-features:
221 Target Features
222 ---------------
224 Target features control how code is generated to support certain
225 processor specific features. Not all target features are supported by
226 all processors. The runtime must ensure that the features supported by
227 the device used to execute the code match the features enabled when
228 generating the code. A mismatch of features may result in incorrect
229 execution, or a reduction in performance.
231 The target features supported by each processor, and the default value
232 used if not specified explicitly, is listed in
233 :ref:`amdgpu-processor-table`.
235 Use the ``clang -m[no-]<TargetFeature>`` option to specify the AMD GPU
236 target features.
238 For example:
240 ``-mxnack``
241 Enable the ``xnack`` feature.
242 ``-mno-xnack``
243 Disable the ``xnack`` feature.
245 .. table:: AMDGPU Target Features
246 :name: amdgpu-target-feature-table
248 =============== ==================================================
249 Target Feature Description
250 =============== ==================================================
251 -m[no-]xnack Enable/disable generating code that has
252 memory clauses that are compatible with
253 having XNACK replay enabled.
255 This is used for demand paging and page
256 migration. If XNACK replay is enabled in
257 the device, then if a page fault occurs
258 the code may execute incorrectly if the
259 ``xnack`` feature is not enabled. Executing
260 code that has the feature enabled on a
261 device that does not have XNACK replay
262 enabled will execute correctly, but may
263 be less performant than code with the
264 feature disabled.
265 -m[no-]sram-ecc Enable/disable generating code that assumes SRAM
266 ECC is enabled/disabled.
267 =============== ==================================================
269 .. _amdgpu-address-spaces:
271 Address Spaces
272 --------------
274 The AMDGPU backend uses the following address space mappings.
276 The memory space names used in the table, aside from the region memory space, is
277 from the OpenCL standard.
279 LLVM Address Space number is used throughout LLVM (for example, in LLVM IR).
281 .. table:: Address Space Mapping
282 :name: amdgpu-address-space-mapping-table
284 ================== =================
285 LLVM Address Space Memory Space
286 ================== =================
287 0 Generic (Flat)
288 1 Global
289 2 Region (GDS)
290 3 Local (group/LDS)
291 4 Constant
292 5 Private (Scratch)
293 6 Constant 32-bit
294 ================== =================
296 .. _amdgpu-memory-scopes:
298 Memory Scopes
299 -------------
301 This section provides LLVM memory synchronization scopes supported by the AMDGPU
302 backend memory model when the target triple OS is ``amdhsa`` (see
303 :ref:`amdgpu-amdhsa-memory-model` and :ref:`amdgpu-target-triples`).
305 The memory model supported is based on the HSA memory model [HSA]_ which is
306 based in turn on HRF-indirect with scope inclusion [HRF]_. The happens-before
307 relation is transitive over the synchonizes-with relation independent of scope,
308 and synchonizes-with allows the memory scope instances to be inclusive (see
309 table :ref:`amdgpu-amdhsa-llvm-sync-scopes-table`).
311 This is different to the OpenCL [OpenCL]_ memory model which does not have scope
312 inclusion and requires the memory scopes to exactly match. However, this
313 is conservatively correct for OpenCL.
315 .. table:: AMDHSA LLVM Sync Scopes
316 :name: amdgpu-amdhsa-llvm-sync-scopes-table
318 ================ ==========================================================
319 LLVM Sync Scope Description
320 ================ ==========================================================
321 *none* The default: ``system``.
323 Synchronizes with, and participates in modification and
324 seq_cst total orderings with, other operations (except
325 image operations) for all address spaces (except private,
326 or generic that accesses private) provided the other
327 operation's sync scope is:
329 - ``system``.
330 - ``agent`` and executed by a thread on the same agent.
331 - ``workgroup`` and executed by a thread in the same
332 workgroup.
333 - ``wavefront`` and executed by a thread in the same
334 wavefront.
336 ``agent`` Synchronizes with, and participates in modification and
337 seq_cst total orderings with, other operations (except
338 image operations) for all address spaces (except private,
339 or generic that accesses private) provided the other
340 operation's sync scope is:
342 - ``system`` or ``agent`` and executed by a thread on the
343 same agent.
344 - ``workgroup`` and executed by a thread in the same
345 workgroup.
346 - ``wavefront`` and executed by a thread in the same
347 wavefront.
349 ``workgroup`` Synchronizes with, and participates in modification and
350 seq_cst total orderings with, other operations (except
351 image operations) for all address spaces (except private,
352 or generic that accesses private) provided the other
353 operation's sync scope is:
355 - ``system``, ``agent`` or ``workgroup`` and executed by a
356 thread in the same workgroup.
357 - ``wavefront`` and executed by a thread in the same
358 wavefront.
360 ``wavefront`` Synchronizes with, and participates in modification and
361 seq_cst total orderings with, other operations (except
362 image operations) for all address spaces (except private,
363 or generic that accesses private) provided the other
364 operation's sync scope is:
366 - ``system``, ``agent``, ``workgroup`` or ``wavefront``
367 and executed by a thread in the same wavefront.
369 ``singlethread`` Only synchronizes with, and participates in modification
370 and seq_cst total orderings with, other operations (except
371 image operations) running in the same thread for all
372 address spaces (for example, in signal handlers).
373 ================ ==========================================================
375 AMDGPU Intrinsics
376 -----------------
378 The AMDGPU backend implements the following LLVM IR intrinsics.
380 *This section is WIP.*
382 .. TODO
383 List AMDGPU intrinsics
385 AMDGPU Attributes
386 -----------------
388 The AMDGPU backend supports the following LLVM IR attributes.
390 .. table:: AMDGPU LLVM IR Attributes
391 :name: amdgpu-llvm-ir-attributes-table
393 ======================================= ==========================================================
394 LLVM Attribute Description
395 ======================================= ==========================================================
396 "amdgpu-flat-work-group-size"="min,max" Specify the minimum and maximum flat work group sizes that
397 will be specified when the kernel is dispatched. Generated
398 by the ``amdgpu_flat_work_group_size`` CLANG attribute [CLANG-ATTR]_.
399 "amdgpu-implicitarg-num-bytes"="n" Number of kernel argument bytes to add to the kernel
400 argument block size for the implicit arguments. This
401 varies by OS and language (for OpenCL see
402 :ref:`opencl-kernel-implicit-arguments-appended-for-amdhsa-os-table`).
403 "amdgpu-max-work-group-size"="n" Specify the maximum work-group size that will be specifed
404 when the kernel is dispatched.
405 "amdgpu-num-sgpr"="n" Specifies the number of SGPRs to use. Generated by
406 the ``amdgpu_num_sgpr`` CLANG attribute [CLANG-ATTR]_.
407 "amdgpu-num-vgpr"="n" Specifies the number of VGPRs to use. Generated by the
408 ``amdgpu_num_vgpr`` CLANG attribute [CLANG-ATTR]_.
409 "amdgpu-waves-per-eu"="m,n" Specify the minimum and maximum number of waves per
410 execution unit. Generated by the ``amdgpu_waves_per_eu``
411 CLANG attribute [CLANG-ATTR]_.
412 ======================================= ==========================================================
414 Code Object
415 ===========
417 The AMDGPU backend generates a standard ELF [ELF]_ relocatable code object that
418 can be linked by ``lld`` to produce a standard ELF shared code object which can
419 be loaded and executed on an AMDGPU target.
421 Header
422 ------
424 The AMDGPU backend uses the following ELF header:
426 .. table:: AMDGPU ELF Header
427 :name: amdgpu-elf-header-table
429 ========================== ===============================
430 Field Value
431 ========================== ===============================
432 ``e_ident[EI_CLASS]`` ``ELFCLASS64``
433 ``e_ident[EI_DATA]`` ``ELFDATA2LSB``
434 ``e_ident[EI_OSABI]`` - ``ELFOSABI_NONE``
435 - ``ELFOSABI_AMDGPU_HSA``
436 - ``ELFOSABI_AMDGPU_PAL``
437 - ``ELFOSABI_AMDGPU_MESA3D``
438 ``e_ident[EI_ABIVERSION]`` - ``ELFABIVERSION_AMDGPU_HSA``
439 - ``ELFABIVERSION_AMDGPU_PAL``
440 - ``ELFABIVERSION_AMDGPU_MESA3D``
441 ``e_type`` - ``ET_REL``
442 - ``ET_DYN``
443 ``e_machine`` ``EM_AMDGPU``
444 ``e_entry`` 0
445 ``e_flags`` See :ref:`amdgpu-elf-header-e_flags-table`
446 ========================== ===============================
448 ..
450 .. table:: AMDGPU ELF Header Enumeration Values
451 :name: amdgpu-elf-header-enumeration-values-table
453 =============================== =====
454 Name Value
455 =============================== =====
456 ``EM_AMDGPU`` 224
457 ``ELFOSABI_NONE`` 0
458 ``ELFOSABI_AMDGPU_HSA`` 64
459 ``ELFOSABI_AMDGPU_PAL`` 65
460 ``ELFOSABI_AMDGPU_MESA3D`` 66
461 ``ELFABIVERSION_AMDGPU_HSA`` 1
462 ``ELFABIVERSION_AMDGPU_PAL`` 0
463 ``ELFABIVERSION_AMDGPU_MESA3D`` 0
464 =============================== =====
466 ``e_ident[EI_CLASS]``
467 The ELF class is:
469 * ``ELFCLASS32`` for ``r600`` architecture.
471 * ``ELFCLASS64`` for ``amdgcn`` architecture which only supports 64
472 bit applications.
474 ``e_ident[EI_DATA]``
475 All AMDGPU targets use ``ELFDATA2LSB`` for little-endian byte ordering.
477 ``e_ident[EI_OSABI]``
478 One of the following AMD GPU architecture specific OS ABIs
479 (see :ref:`amdgpu-os-table`):
481 * ``ELFOSABI_NONE`` for *unknown* OS.
483 * ``ELFOSABI_AMDGPU_HSA`` for ``amdhsa`` OS.
485 * ``ELFOSABI_AMDGPU_PAL`` for ``amdpal`` OS.
487 * ``ELFOSABI_AMDGPU_MESA3D`` for ``mesa3D`` OS.
489 ``e_ident[EI_ABIVERSION]``
490 The ABI version of the AMD GPU architecture specific OS ABI to which the code
491 object conforms:
493 * ``ELFABIVERSION_AMDGPU_HSA`` is used to specify the version of AMD HSA
494 runtime ABI.
496 * ``ELFABIVERSION_AMDGPU_PAL`` is used to specify the version of AMD PAL
497 runtime ABI.
499 * ``ELFABIVERSION_AMDGPU_MESA3D`` is used to specify the version of AMD MESA
500 3D runtime ABI.
502 ``e_type``
503 Can be one of the following values:
506 ``ET_REL``
507 The type produced by the AMD GPU backend compiler as it is relocatable code
508 object.
510 ``ET_DYN``
511 The type produced by the linker as it is a shared code object.
513 The AMD HSA runtime loader requires a ``ET_DYN`` code object.
515 ``e_machine``
516 The value ``EM_AMDGPU`` is used for the machine for all processors supported
517 by the ``r600`` and ``amdgcn`` architectures (see
518 :ref:`amdgpu-processor-table`). The specific processor is specified in the
519 ``EF_AMDGPU_MACH`` bit field of the ``e_flags`` (see
520 :ref:`amdgpu-elf-header-e_flags-table`).
522 ``e_entry``
523 The entry point is 0 as the entry points for individual kernels must be
524 selected in order to invoke them through AQL packets.
526 ``e_flags``
527 The AMDGPU backend uses the following ELF header flags:
529 .. table:: AMDGPU ELF Header ``e_flags``
530 :name: amdgpu-elf-header-e_flags-table
532 ================================= ========== =============================
533 Name Value Description
534 ================================= ========== =============================
535 **AMDGPU Processor Flag** See :ref:`amdgpu-processor-table`.
536 -------------------------------------------- -----------------------------
537 ``EF_AMDGPU_MACH`` 0x000000ff AMDGPU processor selection
538 mask for
539 ``EF_AMDGPU_MACH_xxx`` values
540 defined in
541 :ref:`amdgpu-ef-amdgpu-mach-table`.
542 ``EF_AMDGPU_XNACK`` 0x00000100 Indicates if the ``xnack``
543 target feature is
544 enabled for all code
545 contained in the code object.
546 If the processor
547 does not support the
548 ``xnack`` target
549 feature then must
550 be 0.
551 See
552 :ref:`amdgpu-target-features`.
553 ``EF_AMDGPU_SRAM_ECC`` 0x00000200 Indicates if the ``sram-ecc``
554 target feature is
555 enabled for all code
556 contained in the code object.
557 If the processor
558 does not support the
559 ``sram-ecc`` target
560 feature then must
561 be 0.
562 See
563 :ref:`amdgpu-target-features`.
564 ================================= ========== =============================
566 .. table:: AMDGPU ``EF_AMDGPU_MACH`` Values
567 :name: amdgpu-ef-amdgpu-mach-table
569 ================================= ========== =============================
570 Name Value Description (see
571 :ref:`amdgpu-processor-table`)
572 ================================= ========== =============================
573 ``EF_AMDGPU_MACH_NONE`` 0x000 *not specified*
574 ``EF_AMDGPU_MACH_R600_R600`` 0x001 ``r600``
575 ``EF_AMDGPU_MACH_R600_R630`` 0x002 ``r630``
576 ``EF_AMDGPU_MACH_R600_RS880`` 0x003 ``rs880``
577 ``EF_AMDGPU_MACH_R600_RV670`` 0x004 ``rv670``
578 ``EF_AMDGPU_MACH_R600_RV710`` 0x005 ``rv710``
579 ``EF_AMDGPU_MACH_R600_RV730`` 0x006 ``rv730``
580 ``EF_AMDGPU_MACH_R600_RV770`` 0x007 ``rv770``
581 ``EF_AMDGPU_MACH_R600_CEDAR`` 0x008 ``cedar``
582 ``EF_AMDGPU_MACH_R600_CYPRESS`` 0x009 ``cypress``
583 ``EF_AMDGPU_MACH_R600_JUNIPER`` 0x00a ``juniper``
584 ``EF_AMDGPU_MACH_R600_REDWOOD`` 0x00b ``redwood``
585 ``EF_AMDGPU_MACH_R600_SUMO`` 0x00c ``sumo``
586 ``EF_AMDGPU_MACH_R600_BARTS`` 0x00d ``barts``
587 ``EF_AMDGPU_MACH_R600_CAICOS`` 0x00e ``caicos``
588 ``EF_AMDGPU_MACH_R600_CAYMAN`` 0x00f ``cayman``
589 ``EF_AMDGPU_MACH_R600_TURKS`` 0x010 ``turks``
590 *reserved* 0x011 - Reserved for ``r600``
591 0x01f architecture processors.
592 ``EF_AMDGPU_MACH_AMDGCN_GFX600`` 0x020 ``gfx600``
593 ``EF_AMDGPU_MACH_AMDGCN_GFX601`` 0x021 ``gfx601``
594 ``EF_AMDGPU_MACH_AMDGCN_GFX700`` 0x022 ``gfx700``
595 ``EF_AMDGPU_MACH_AMDGCN_GFX701`` 0x023 ``gfx701``
596 ``EF_AMDGPU_MACH_AMDGCN_GFX702`` 0x024 ``gfx702``
597 ``EF_AMDGPU_MACH_AMDGCN_GFX703`` 0x025 ``gfx703``
598 ``EF_AMDGPU_MACH_AMDGCN_GFX704`` 0x026 ``gfx704``
599 *reserved* 0x027 Reserved.
600 ``EF_AMDGPU_MACH_AMDGCN_GFX801`` 0x028 ``gfx801``
601 ``EF_AMDGPU_MACH_AMDGCN_GFX802`` 0x029 ``gfx802``
602 ``EF_AMDGPU_MACH_AMDGCN_GFX803`` 0x02a ``gfx803``
603 ``EF_AMDGPU_MACH_AMDGCN_GFX810`` 0x02b ``gfx810``
604 ``EF_AMDGPU_MACH_AMDGCN_GFX900`` 0x02c ``gfx900``
605 ``EF_AMDGPU_MACH_AMDGCN_GFX902`` 0x02d ``gfx902``
606 ``EF_AMDGPU_MACH_AMDGCN_GFX904`` 0x02e ``gfx904``
607 ``EF_AMDGPU_MACH_AMDGCN_GFX906`` 0x02f ``gfx906``
608 *reserved* 0x030 Reserved.
609 ``EF_AMDGPU_MACH_AMDGCN_GFX909`` 0x031 ``gfx909``
610 ================================= ========== =============================
612 Sections
613 --------
615 An AMDGPU target ELF code object has the standard ELF sections which include:
617 .. table:: AMDGPU ELF Sections
618 :name: amdgpu-elf-sections-table
620 ================== ================ =================================
621 Name Type Attributes
622 ================== ================ =================================
623 ``.bss`` ``SHT_NOBITS`` ``SHF_ALLOC`` + ``SHF_WRITE``
624 ``.data`` ``SHT_PROGBITS`` ``SHF_ALLOC`` + ``SHF_WRITE``
625 ``.debug_``\ *\** ``SHT_PROGBITS`` *none*
626 ``.dynamic`` ``SHT_DYNAMIC`` ``SHF_ALLOC``
627 ``.dynstr`` ``SHT_PROGBITS`` ``SHF_ALLOC``
628 ``.dynsym`` ``SHT_PROGBITS`` ``SHF_ALLOC``
629 ``.got`` ``SHT_PROGBITS`` ``SHF_ALLOC`` + ``SHF_WRITE``
630 ``.hash`` ``SHT_HASH`` ``SHF_ALLOC``
631 ``.note`` ``SHT_NOTE`` *none*
632 ``.rela``\ *name* ``SHT_RELA`` *none*
633 ``.rela.dyn`` ``SHT_RELA`` *none*
634 ``.rodata`` ``SHT_PROGBITS`` ``SHF_ALLOC``
635 ``.shstrtab`` ``SHT_STRTAB`` *none*
636 ``.strtab`` ``SHT_STRTAB`` *none*
637 ``.symtab`` ``SHT_SYMTAB`` *none*
638 ``.text`` ``SHT_PROGBITS`` ``SHF_ALLOC`` + ``SHF_EXECINSTR``
639 ================== ================ =================================
641 These sections have their standard meanings (see [ELF]_) and are only generated
642 if needed.
644 ``.debug``\ *\**
645 The standard DWARF sections. See :ref:`amdgpu-dwarf` for information on the
646 DWARF produced by the AMDGPU backend.
648 ``.dynamic``, ``.dynstr``, ``.dynsym``, ``.hash``
649 The standard sections used by a dynamic loader.
651 ``.note``
652 See :ref:`amdgpu-note-records` for the note records supported by the AMDGPU
653 backend.
655 ``.rela``\ *name*, ``.rela.dyn``
656 For relocatable code objects, *name* is the name of the section that the
657 relocation records apply. For example, ``.rela.text`` is the section name for
658 relocation records associated with the ``.text`` section.
660 For linked shared code objects, ``.rela.dyn`` contains all the relocation
661 records from each of the relocatable code object's ``.rela``\ *name* sections.
663 See :ref:`amdgpu-relocation-records` for the relocation records supported by
664 the AMDGPU backend.
666 ``.text``
667 The executable machine code for the kernels and functions they call. Generated
668 as position independent code. See :ref:`amdgpu-code-conventions` for
669 information on conventions used in the isa generation.
671 .. _amdgpu-note-records:
673 Note Records
674 ------------
676 As required by ``ELFCLASS32`` and ``ELFCLASS64``, minimal zero byte padding must
677 be generated after the ``name`` field to ensure the ``desc`` field is 4 byte
678 aligned. In addition, minimal zero byte padding must be generated to ensure the
679 ``desc`` field size is a multiple of 4 bytes. The ``sh_addralign`` field of the
680 ``.note`` section must be at least 4 to indicate at least 8 byte alignment.
682 .. _amdgpu-note-records-v2:
684 Code Object V2 Note Records (-mattr=-code-object-v3)
685 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
687 The AMDGPU backend code object uses the following ELF note record in the
688 ``.note`` section.
690 Additional note records can be present.
692 .. table:: AMDGPU Code Object V2 ELF Note Records
693 :name: amdgpu-elf-note-records-table-v2
695 ===== ============================== ======================================
696 Name Type Description
697 ===== ============================== ======================================
698 "AMD" ``NT_AMD_AMDGPU_HSA_METADATA`` <metadata null terminated string>
699 ===== ============================== ======================================
701 ..
703 .. table:: AMDGPU Code Object V2 ELF Note Record Enumeration Values
704 :name: amdgpu-elf-note-record-enumeration-values-table-v2
706 ============================== =====
707 Name Value
708 ============================== =====
709 *reserved* 0-9
710 ``NT_AMD_AMDGPU_HSA_METADATA`` 10
711 *reserved* 11
712 ============================== =====
714 ``NT_AMD_AMDGPU_HSA_METADATA``
715 Specifies extensible metadata associated with the code objects executed on HSA
716 [HSA]_ compatible runtimes such as AMD's ROCm [AMD-ROCm]_. It is required when
717 the target triple OS is ``amdhsa`` (see :ref:`amdgpu-target-triples`). See
718 :ref:`amdgpu-amdhsa-code-object-metadata-v2` for the syntax of the code
719 object metadata string.
721 .. _amdgpu-note-records-v3:
723 Code Object V3 Note Records (-mattr=+code-object-v3)
724 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
726 The AMDGPU backend code object uses the following ELF note record in the
727 ``.note`` section.
729 Additional note records can be present.
731 .. table:: AMDGPU Code Object V3 ELF Note Records
732 :name: amdgpu-elf-note-records-table-v3
734 ======== ============================== ======================================
735 Name Type Description
736 ======== ============================== ======================================
737 "AMDGPU" ``NT_AMDGPU_METADATA`` Metadata in Message Pack [MsgPack]_
738 binary format.
739 ======== ============================== ======================================
741 ..
743 .. table:: AMDGPU Code Object V3 ELF Note Record Enumeration Values
744 :name: amdgpu-elf-note-record-enumeration-values-table-v3
746 ============================== =====
747 Name Value
748 ============================== =====
749 *reserved* 0-31
750 ``NT_AMDGPU_METADATA`` 32
751 ============================== =====
753 ``NT_AMDGPU_METADATA``
754 Specifies extensible metadata associated with an AMDGPU code
755 object. It is encoded as a map in the Message Pack [MsgPack]_ binary
756 data format. See :ref:`amdgpu-amdhsa-code-object-metadata-v3` for the
757 map keys defined for the ``amdhsa`` OS.
759 .. _amdgpu-symbols:
761 Symbols
762 -------
764 Symbols include the following:
766 .. table:: AMDGPU ELF Symbols
767 :name: amdgpu-elf-symbols-table
769 ===================== ============== ============= ==================
770 Name Type Section Description
771 ===================== ============== ============= ==================
772 *link-name* ``STT_OBJECT`` - ``.data`` Global variable
773 - ``.rodata``
774 - ``.bss``
775 *link-name*\ ``.kd`` ``STT_OBJECT`` - ``.rodata`` Kernel descriptor
776 *link-name* ``STT_FUNC`` - ``.text`` Kernel entry point
777 ===================== ============== ============= ==================
779 Global variable
780 Global variables both used and defined by the compilation unit.
782 If the symbol is defined in the compilation unit then it is allocated in the
783 appropriate section according to if it has initialized data or is readonly.
785 If the symbol is external then its section is ``STN_UNDEF`` and the loader
786 will resolve relocations using the definition provided by another code object
787 or explicitly defined by the runtime.
789 All global symbols, whether defined in the compilation unit or external, are
790 accessed by the machine code indirectly through a GOT table entry. This
791 allows them to be preemptable. The GOT table is only supported when the target
792 triple OS is ``amdhsa`` (see :ref:`amdgpu-target-triples`).
794 .. TODO
795 Add description of linked shared object symbols. Seems undefined symbols
796 are marked as STT_NOTYPE.
798 Kernel descriptor
799 Every HSA kernel has an associated kernel descriptor. It is the address of the
800 kernel descriptor that is used in the AQL dispatch packet used to invoke the
801 kernel, not the kernel entry point. The layout of the HSA kernel descriptor is
802 defined in :ref:`amdgpu-amdhsa-kernel-descriptor`.
804 Kernel entry point
805 Every HSA kernel also has a symbol for its machine code entry point.
807 .. _amdgpu-relocation-records:
809 Relocation Records
810 ------------------
812 AMDGPU backend generates ``Elf64_Rela`` relocation records. Supported
813 relocatable fields are:
815 ``word32``
816 This specifies a 32-bit field occupying 4 bytes with arbitrary byte
817 alignment. These values use the same byte order as other word values in the
818 AMD GPU architecture.
820 ``word64``
821 This specifies a 64-bit field occupying 8 bytes with arbitrary byte
822 alignment. These values use the same byte order as other word values in the
823 AMD GPU architecture.
825 Following notations are used for specifying relocation calculations:
827 **A**
828 Represents the addend used to compute the value of the relocatable field.
830 **G**
831 Represents the offset into the global offset table at which the relocation
832 entry's symbol will reside during execution.
834 **GOT**
835 Represents the address of the global offset table.
837 **P**
838 Represents the place (section offset for ``et_rel`` or address for ``et_dyn``)
839 of the storage unit being relocated (computed using ``r_offset``).
841 **S**
842 Represents the value of the symbol whose index resides in the relocation
843 entry. Relocations not using this must specify a symbol index of ``STN_UNDEF``.
845 **B**
846 Represents the base address of a loaded executable or shared object which is
847 the difference between the ELF address and the actual load address. Relocations
848 using this are only valid in executable or shared objects.
850 The following relocation types are supported:
852 .. table:: AMDGPU ELF Relocation Records
853 :name: amdgpu-elf-relocation-records-table
855 ========================== ======= ===== ========== ==============================
856 Relocation Type Kind Value Field Calculation
857 ========================== ======= ===== ========== ==============================
858 ``R_AMDGPU_NONE`` 0 *none* *none*
859 ``R_AMDGPU_ABS32_LO`` Static, 1 ``word32`` (S + A) & 0xFFFFFFFF
860 Dynamic
861 ``R_AMDGPU_ABS32_HI`` Static, 2 ``word32`` (S + A) >> 32
862 Dynamic
863 ``R_AMDGPU_ABS64`` Static, 3 ``word64`` S + A
864 Dynamic
865 ``R_AMDGPU_REL32`` Static 4 ``word32`` S + A - P
866 ``R_AMDGPU_REL64`` Static 5 ``word64`` S + A - P
867 ``R_AMDGPU_ABS32`` Static, 6 ``word32`` S + A
868 Dynamic
869 ``R_AMDGPU_GOTPCREL`` Static 7 ``word32`` G + GOT + A - P
870 ``R_AMDGPU_GOTPCREL32_LO`` Static 8 ``word32`` (G + GOT + A - P) & 0xFFFFFFFF
871 ``R_AMDGPU_GOTPCREL32_HI`` Static 9 ``word32`` (G + GOT + A - P) >> 32
872 ``R_AMDGPU_REL32_LO`` Static 10 ``word32`` (S + A - P) & 0xFFFFFFFF
873 ``R_AMDGPU_REL32_HI`` Static 11 ``word32`` (S + A - P) >> 32
874 *reserved* 12
875 ``R_AMDGPU_RELATIVE64`` Dynamic 13 ``word64`` B + A
876 ========================== ======= ===== ========== ==============================
878 ``R_AMDGPU_ABS32_LO`` and ``R_AMDGPU_ABS32_HI`` are only supported by
879 the ``mesa3d`` OS, which does not support ``R_AMDGPU_ABS64``.
881 There is no current OS loader support for 32 bit programs and so
882 ``R_AMDGPU_ABS32`` is not used.
884 .. _amdgpu-dwarf:
886 DWARF
887 -----
889 Standard DWARF [DWARF]_ Version 5 sections can be generated. These contain
890 information that maps the code object executable code and data to the source
891 language constructs. It can be used by tools such as debuggers and profilers.
893 Address Space Mapping
894 ~~~~~~~~~~~~~~~~~~~~~
896 The following address space mapping is used:
898 .. table:: AMDGPU DWARF Address Space Mapping
899 :name: amdgpu-dwarf-address-space-mapping-table
901 =================== =================
902 DWARF Address Space Memory Space
903 =================== =================
904 1 Private (Scratch)
905 2 Local (group/LDS)
906 *omitted* Global
907 *omitted* Constant
908 *omitted* Generic (Flat)
909 *not supported* Region (GDS)
910 =================== =================
912 See :ref:`amdgpu-address-spaces` for information on the memory space terminology
913 used in the table.
915 An ``address_class`` attribute is generated on pointer type DIEs to specify the
916 DWARF address space of the value of the pointer when it is in the *private* or
917 *local* address space. Otherwise the attribute is omitted.
919 An ``XDEREF`` operation is generated in location list expressions for variables
920 that are allocated in the *private* and *local* address space. Otherwise no
921 ``XDREF`` is omitted.
923 Register Mapping
924 ~~~~~~~~~~~~~~~~
926 *This section is WIP.*
928 .. TODO
929 Define DWARF register enumeration.
931 If want to present a wavefront state then should expose vector registers as
932 64 wide (rather than per work-item view that LLVM uses). Either as separate
933 registers, or a 64x4 byte single register. In either case use a new LANE op
934 (akin to XDREF) to select the current lane usage in a location
935 expression. This would also allow scalar register spilling to vector register
936 lanes to be expressed (currently no debug information is being generated for
937 spilling). If choose a wide single register approach then use LANE in
938 conjunction with PIECE operation to select the dword part of the register for
939 the current lane. If the separate register approach then use LANE to select
940 the register.
942 Source Text
943 ~~~~~~~~~~~
945 Source text for online-compiled programs (e.g. those compiled by the OpenCL
946 runtime) may be embedded into the DWARF v5 line table using the ``clang
947 -gembed-source`` option, described in table :ref:`amdgpu-debug-options`.
949 For example:
951 ``-gembed-source``
952 Enable the embedded source DWARF v5 extension.
953 ``-gno-embed-source``
954 Disable the embedded source DWARF v5 extension.
956 .. table:: AMDGPU Debug Options
957 :name: amdgpu-debug-options
959 ==================== ==================================================
960 Debug Flag Description
961 ==================== ==================================================
962 -g[no-]embed-source Enable/disable embedding source text in DWARF
963 debug sections. Useful for environments where
964 source cannot be written to disk, such as
965 when performing online compilation.
966 ==================== ==================================================
968 This option enables one extended content types in the DWARF v5 Line Number
969 Program Header, which is used to encode embedded source.
971 .. table:: AMDGPU DWARF Line Number Program Header Extended Content Types
972 :name: amdgpu-dwarf-extended-content-types
974 ============================ ======================
975 Content Type Form
976 ============================ ======================
977 ``DW_LNCT_LLVM_source`` ``DW_FORM_line_strp``
978 ============================ ======================
980 The source field will contain the UTF-8 encoded, null-terminated source text
981 with ``'\n'`` line endings. When the source field is present, consumers can use
982 the embedded source instead of attempting to discover the source on disk. When
983 the source field is absent, consumers can access the file to get the source
984 text.
986 The above content type appears in the ``file_name_entry_format`` field of the
987 line table prologue, and its corresponding value appear in the ``file_names``
988 field. The current encoding of the content type is documented in table
989 :ref:`amdgpu-dwarf-extended-content-types-encoding`
991 .. table:: AMDGPU DWARF Line Number Program Header Extended Content Types Encoding
992 :name: amdgpu-dwarf-extended-content-types-encoding
994 ============================ ====================
995 Content Type Value
996 ============================ ====================
997 ``DW_LNCT_LLVM_source`` 0x2001
998 ============================ ====================
1000 .. _amdgpu-code-conventions:
1002 Code Conventions
1003 ================
1005 This section provides code conventions used for each supported target triple OS
1006 (see :ref:`amdgpu-target-triples`).
1008 AMDHSA
1009 ------
1011 This section provides code conventions used when the target triple OS is
1012 ``amdhsa`` (see :ref:`amdgpu-target-triples`).
1014 .. _amdgpu-amdhsa-code-object-target-identification:
1016 Code Object Target Identification
1017 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1019 The AMDHSA OS uses the following syntax to specify the code object
1020 target as a single string:
1022 ``<Architecture>-<Vendor>-<OS>-<Environment>-<Processor><Target Features>``
1024 Where:
1026 - ``<Architecture>``, ``<Vendor>``, ``<OS>`` and ``<Environment>``
1027 are the same as the *Target Triple* (see
1028 :ref:`amdgpu-target-triples`).
1030 - ``<Processor>`` is the same as the *Processor* (see
1031 :ref:`amdgpu-processors`).
1033 - ``<Target Features>`` is a list of the enabled *Target Features*
1034 (see :ref:`amdgpu-target-features`), each prefixed by a plus, that
1035 apply to *Processor*. The list must be in the same order as listed
1036 in the table :ref:`amdgpu-target-feature-table`. Note that *Target
1037 Features* must be included in the list if they are enabled even if
1038 that is the default for *Processor*.
1040 For example:
1042 ``"amdgcn-amd-amdhsa--gfx902+xnack"``
1044 .. _amdgpu-amdhsa-code-object-metadata:
1046 Code Object Metadata
1047 ~~~~~~~~~~~~~~~~~~~~
1049 The code object metadata specifies extensible metadata associated with the code
1050 objects executed on HSA [HSA]_ compatible runtimes such as AMD's ROCm
1051 [AMD-ROCm]_. It is specified in a note record (see :ref:`amdgpu-note-records`)
1052 and is required when the target triple OS is ``amdhsa`` (see
1053 :ref:`amdgpu-target-triples`). It must contain the minimum information
1054 necessary to support the ROCM kernel queries. For example, the segment sizes
1055 needed in a dispatch packet. In addition, a high level language runtime may
1056 require other information to be included. For example, the AMD OpenCL runtime
1057 records kernel argument information.
1059 .. _amdgpu-amdhsa-code-object-metadata-v2:
1061 Code Object V2 Metadata (-mattr=-code-object-v3)
1062 ++++++++++++++++++++++++++++++++++++++++++++++++
1064 Code object V2 metadata is specified by the ``NT_AMD_AMDGPU_METADATA`` note
1065 record (see :ref:`amdgpu-note-records-v2`).
1067 The metadata is specified as a YAML formatted string (see [YAML]_ and
1068 :doc:`YamlIO`).
1070 .. TODO
1071 Is the string null terminated? It probably should not if YAML allows it to
1072 contain null characters, otherwise it should be.
1074 The metadata is represented as a single YAML document comprised of the mapping
1075 defined in table :ref:`amdgpu-amdhsa-code-object-metadata-map-table-v2` and
1076 referenced tables.
1078 For boolean values, the string values of ``false`` and ``true`` are used for
1079 false and true respectively.
1081 Additional information can be added to the mappings. To avoid conflicts, any
1082 non-AMD key names should be prefixed by "*vendor-name*.".
1084 .. table:: AMDHSA Code Object V2 Metadata Map
1085 :name: amdgpu-amdhsa-code-object-metadata-map-table-v2
1087 ========== ============== ========= =======================================
1088 String Key Value Type Required? Description
1089 ========== ============== ========= =======================================
1090 "Version" sequence of Required - The first integer is the major
1091 2 integers version. Currently 1.
1092 - The second integer is the minor
1093 version. Currently 0.
1094 "Printf" sequence of Each string is encoded information
1095 strings about a printf function call. The
1096 encoded information is organized as
1097 fields separated by colon (':'):
1099 ``ID:N:S[0]:S[1]:...:S[N-1]:FormatString``
1101 where:
1103 ``ID``
1104 A 32 bit integer as a unique id for
1105 each printf function call
1107 ``N``
1108 A 32 bit integer equal to the number
1109 of arguments of printf function call
1110 minus 1
1112 ``S[i]`` (where i = 0, 1, ... , N-1)
1113 32 bit integers for the size in bytes
1114 of the i-th FormatString argument of
1115 the printf function call
1117 FormatString
1118 The format string passed to the
1119 printf function call.
1120 "Kernels" sequence of Required Sequence of the mappings for each
1121 mapping kernel in the code object. See
1122 :ref:`amdgpu-amdhsa-code-object-kernel-metadata-map-table-v2`
1123 for the definition of the mapping.
1124 ========== ============== ========= =======================================
1126 ..
1128 .. table:: AMDHSA Code Object V2 Kernel Metadata Map
1129 :name: amdgpu-amdhsa-code-object-kernel-metadata-map-table-v2
1131 ================= ============== ========= ================================
1132 String Key Value Type Required? Description
1133 ================= ============== ========= ================================
1134 "Name" string Required Source name of the kernel.
1135 "SymbolName" string Required Name of the kernel
1136 descriptor ELF symbol.
1137 "Language" string Source language of the kernel.
1138 Values include:
1140 - "OpenCL C"
1141 - "OpenCL C++"
1142 - "HCC"
1143 - "OpenMP"
1145 "LanguageVersion" sequence of - The first integer is the major
1146 2 integers version.
1147 - The second integer is the
1148 minor version.
1149 "Attrs" mapping Mapping of kernel attributes.
1150 See
1151 :ref:`amdgpu-amdhsa-code-object-kernel-attribute-metadata-map-table-v2`
1152 for the mapping definition.
1153 "Args" sequence of Sequence of mappings of the
1154 mapping kernel arguments. See
1155 :ref:`amdgpu-amdhsa-code-object-kernel-argument-metadata-map-table-v2`
1156 for the definition of the mapping.
1157 "CodeProps" mapping Mapping of properties related to
1158 the kernel code. See
1159 :ref:`amdgpu-amdhsa-code-object-kernel-code-properties-metadata-map-table-v2`
1160 for the mapping definition.
1161 ================= ============== ========= ================================
1163 ..
1165 .. table:: AMDHSA Code Object V2 Kernel Attribute Metadata Map
1166 :name: amdgpu-amdhsa-code-object-kernel-attribute-metadata-map-table-v2
1168 =================== ============== ========= ==============================
1169 String Key Value Type Required? Description
1170 =================== ============== ========= ==============================
1171 "ReqdWorkGroupSize" sequence of If not 0, 0, 0 then all values
1172 3 integers must be >=1 and the dispatch
1173 work-group size X, Y, Z must
1174 correspond to the specified
1175 values. Defaults to 0, 0, 0.
1177 Corresponds to the OpenCL
1178 ``reqd_work_group_size``
1179 attribute.
1180 "WorkGroupSizeHint" sequence of The dispatch work-group size
1181 3 integers X, Y, Z is likely to be the
1182 specified values.
1184 Corresponds to the OpenCL
1185 ``work_group_size_hint``
1186 attribute.
1187 "VecTypeHint" string The name of a scalar or vector
1188 type.
1190 Corresponds to the OpenCL
1191 ``vec_type_hint`` attribute.
1193 "RuntimeHandle" string The external symbol name
1194 associated with a kernel.
1195 OpenCL runtime allocates a
1196 global buffer for the symbol
1197 and saves the kernel's address
1198 to it, which is used for
1199 device side enqueueing. Only
1200 available for device side
1201 enqueued kernels.
1202 =================== ============== ========= ==============================
1204 ..
1206 .. table:: AMDHSA Code Object V2 Kernel Argument Metadata Map
1207 :name: amdgpu-amdhsa-code-object-kernel-argument-metadata-map-table-v2
1209 ================= ============== ========= ================================
1210 String Key Value Type Required? Description
1211 ================= ============== ========= ================================
1212 "Name" string Kernel argument name.
1213 "TypeName" string Kernel argument type name.
1214 "Size" integer Required Kernel argument size in bytes.
1215 "Align" integer Required Kernel argument alignment in
1216 bytes. Must be a power of two.
1217 "ValueKind" string Required Kernel argument kind that
1218 specifies how to set up the
1219 corresponding argument.
1220 Values include:
1222 "ByValue"
1223 The argument is copied
1224 directly into the kernarg.
1226 "GlobalBuffer"
1227 A global address space pointer
1228 to the buffer data is passed
1229 in the kernarg.
1231 "DynamicSharedPointer"
1232 A group address space pointer
1233 to dynamically allocated LDS
1234 is passed in the kernarg.
1236 "Sampler"
1237 A global address space
1238 pointer to a S# is passed in
1239 the kernarg.
1241 "Image"
1242 A global address space
1243 pointer to a T# is passed in
1244 the kernarg.
1246 "Pipe"
1247 A global address space pointer
1248 to an OpenCL pipe is passed in
1249 the kernarg.
1251 "Queue"
1252 A global address space pointer
1253 to an OpenCL device enqueue
1254 queue is passed in the
1255 kernarg.
1257 "HiddenGlobalOffsetX"
1258 The OpenCL grid dispatch
1259 global offset for the X
1260 dimension is passed in the
1261 kernarg.
1263 "HiddenGlobalOffsetY"
1264 The OpenCL grid dispatch
1265 global offset for the Y
1266 dimension is passed in the
1267 kernarg.
1269 "HiddenGlobalOffsetZ"
1270 The OpenCL grid dispatch
1271 global offset for the Z
1272 dimension is passed in the
1273 kernarg.
1275 "HiddenNone"
1276 An argument that is not used
1277 by the kernel. Space needs to
1278 be left for it, but it does
1279 not need to be set up.
1281 "HiddenPrintfBuffer"
1282 A global address space pointer
1283 to the runtime printf buffer
1284 is passed in kernarg.
1286 "HiddenDefaultQueue"
1287 A global address space pointer
1288 to the OpenCL device enqueue
1289 queue that should be used by
1290 the kernel by default is
1291 passed in the kernarg.
1293 "HiddenCompletionAction"
1294 A global address space pointer
1295 to help link enqueued kernels into
1296 the ancestor tree for determining
1297 when the parent kernel has finished.
1299 "ValueType" string Required Kernel argument value type. Only
1300 present if "ValueKind" is
1301 "ByValue". For vector data
1302 types, the value is for the
1303 element type. Values include:
1305 - "Struct"
1306 - "I8"
1307 - "U8"
1308 - "I16"
1309 - "U16"
1310 - "F16"
1311 - "I32"
1312 - "U32"
1313 - "F32"
1314 - "I64"
1315 - "U64"
1316 - "F64"
1318 .. TODO
1319 How can it be determined if a
1320 vector type, and what size
1321 vector?
1322 "PointeeAlign" integer Alignment in bytes of pointee
1323 type for pointer type kernel
1324 argument. Must be a power
1325 of 2. Only present if
1326 "ValueKind" is
1327 "DynamicSharedPointer".
1328 "AddrSpaceQual" string Kernel argument address space
1329 qualifier. Only present if
1330 "ValueKind" is "GlobalBuffer" or
1331 "DynamicSharedPointer". Values
1332 are:
1334 - "Private"
1335 - "Global"
1336 - "Constant"
1337 - "Local"
1338 - "Generic"
1339 - "Region"
1341 .. TODO
1342 Is GlobalBuffer only Global
1343 or Constant? Is
1344 DynamicSharedPointer always
1345 Local? Can HCC allow Generic?
1346 How can Private or Region
1347 ever happen?
1348 "AccQual" string Kernel argument access
1349 qualifier. Only present if
1350 "ValueKind" is "Image" or
1351 "Pipe". Values
1352 are:
1354 - "ReadOnly"
1355 - "WriteOnly"
1356 - "ReadWrite"
1358 .. TODO
1359 Does this apply to
1360 GlobalBuffer?
1361 "ActualAccQual" string The actual memory accesses
1362 performed by the kernel on the
1363 kernel argument. Only present if
1364 "ValueKind" is "GlobalBuffer",
1365 "Image", or "Pipe". This may be
1366 more restrictive than indicated
1367 by "AccQual" to reflect what the
1368 kernel actual does. If not
1369 present then the runtime must
1370 assume what is implied by
1371 "AccQual" and "IsConst". Values
1372 are:
1374 - "ReadOnly"
1375 - "WriteOnly"
1376 - "ReadWrite"
1378 "IsConst" boolean Indicates if the kernel argument
1379 is const qualified. Only present
1380 if "ValueKind" is
1381 "GlobalBuffer".
1383 "IsRestrict" boolean Indicates if the kernel argument
1384 is restrict qualified. Only
1385 present if "ValueKind" is
1386 "GlobalBuffer".
1388 "IsVolatile" boolean Indicates if the kernel argument
1389 is volatile qualified. Only
1390 present if "ValueKind" is
1391 "GlobalBuffer".
1393 "IsPipe" boolean Indicates if the kernel argument
1394 is pipe qualified. Only present
1395 if "ValueKind" is "Pipe".
1397 .. TODO
1398 Can GlobalBuffer be pipe
1399 qualified?
1400 ================= ============== ========= ================================
1402 ..
1404 .. table:: AMDHSA Code Object V2 Kernel Code Properties Metadata Map
1405 :name: amdgpu-amdhsa-code-object-kernel-code-properties-metadata-map-table-v2
1407 ============================ ============== ========= =====================
1408 String Key Value Type Required? Description
1409 ============================ ============== ========= =====================
1410 "KernargSegmentSize" integer Required The size in bytes of
1411 the kernarg segment
1412 that holds the values
1413 of the arguments to
1414 the kernel.
1415 "GroupSegmentFixedSize" integer Required The amount of group
1416 segment memory
1417 required by a
1418 work-group in
1419 bytes. This does not
1420 include any
1421 dynamically allocated
1422 group segment memory
1423 that may be added
1424 when the kernel is
1425 dispatched.
1426 "PrivateSegmentFixedSize" integer Required The amount of fixed
1427 private address space
1428 memory required for a
1429 work-item in
1430 bytes. If the kernel
1431 uses a dynamic call
1432 stack then additional
1433 space must be added
1434 to this value for the
1435 call stack.
1436 "KernargSegmentAlign" integer Required The maximum byte
1437 alignment of
1438 arguments in the
1439 kernarg segment. Must
1440 be a power of 2.
1441 "WavefrontSize" integer Required Wavefront size. Must
1442 be a power of 2.
1443 "NumSGPRs" integer Required Number of scalar
1444 registers used by a
1445 wavefront for
1446 GFX6-GFX9. This
1447 includes the special
1448 SGPRs for VCC, Flat
1449 Scratch (GFX7-GFX9)
1450 and XNACK (for
1451 GFX8-GFX9). It does
1452 not include the 16
1453 SGPR added if a trap
1454 handler is
1455 enabled. It is not
1456 rounded up to the
1457 allocation
1458 granularity.
1459 "NumVGPRs" integer Required Number of vector
1460 registers used by
1461 each work-item for
1462 GFX6-GFX9
1463 "MaxFlatWorkGroupSize" integer Required Maximum flat
1464 work-group size
1465 supported by the
1466 kernel in work-items.
1467 Must be >=1 and
1468 consistent with
1469 ReqdWorkGroupSize if
1470 not 0, 0, 0.
1471 "NumSpilledSGPRs" integer Number of stores from
1472 a scalar register to
1473 a register allocator
1474 created spill
1475 location.
1476 "NumSpilledVGPRs" integer Number of stores from
1477 a vector register to
1478 a register allocator
1479 created spill
1480 location.
1481 ============================ ============== ========= =====================
1483 .. _amdgpu-amdhsa-code-object-metadata-v3:
1485 Code Object V3 Metadata (-mattr=+code-object-v3)
1486 ++++++++++++++++++++++++++++++++++++++++++++++++
1488 Code object V3 metadata is specified by the ``NT_AMDGPU_METADATA`` note record
1489 (see :ref:`amdgpu-note-records-v3`).
1491 The metadata is represented as Message Pack formatted binary data (see
1492 [MsgPack]_). The top level is a Message Pack map that includes the
1493 keys defined in table
1494 :ref:`amdgpu-amdhsa-code-object-metadata-map-table-v3` and referenced
1495 tables.
1497 Additional information can be added to the maps. To avoid conflicts,
1498 any key names should be prefixed by "*vendor-name*." where
1499 ``vendor-name`` can be the the name of the vendor and specific vendor
1500 tool that generates the information. The prefix is abbreviated to
1501 simply "." when it appears within a map that has been added by the
1502 same *vendor-name*.
1504 .. table:: AMDHSA Code Object V3 Metadata Map
1505 :name: amdgpu-amdhsa-code-object-metadata-map-table-v3
1507 ================= ============== ========= =======================================
1508 String Key Value Type Required? Description
1509 ================= ============== ========= =======================================
1510 "amdhsa.version" sequence of Required - The first integer is the major
1511 2 integers version. Currently 1.
1512 - The second integer is the minor
1513 version. Currently 0.
1514 "amdhsa.printf" sequence of Each string is encoded information
1515 strings about a printf function call. The
1516 encoded information is organized as
1517 fields separated by colon (':'):
1519 ``ID:N:S[0]:S[1]:...:S[N-1]:FormatString``
1521 where:
1523 ``ID``
1524 A 32 bit integer as a unique id for
1525 each printf function call
1527 ``N``
1528 A 32 bit integer equal to the number
1529 of arguments of printf function call
1530 minus 1
1532 ``S[i]`` (where i = 0, 1, ... , N-1)
1533 32 bit integers for the size in bytes
1534 of the i-th FormatString argument of
1535 the printf function call
1537 FormatString
1538 The format string passed to the
1539 printf function call.
1540 "amdhsa.kernels" sequence of Required Sequence of the maps for each
1541 map kernel in the code object. See
1542 :ref:`amdgpu-amdhsa-code-object-kernel-metadata-map-table-v3`
1543 for the definition of the keys included
1544 in that map.
1545 ================= ============== ========= =======================================
1547 ..
1549 .. table:: AMDHSA Code Object V3 Kernel Metadata Map
1550 :name: amdgpu-amdhsa-code-object-kernel-metadata-map-table-v3
1552 =================================== ============== ========= ================================
1553 String Key Value Type Required? Description
1554 =================================== ============== ========= ================================
1555 ".name" string Required Source name of the kernel.
1556 ".symbol" string Required Name of the kernel
1557 descriptor ELF symbol.
1558 ".language" string Source language of the kernel.
1559 Values include:
1561 - "OpenCL C"
1562 - "OpenCL C++"
1563 - "HCC"
1564 - "HIP"
1565 - "OpenMP"
1566 - "Assembler"
1568 ".language_version" sequence of - The first integer is the major
1569 2 integers version.
1570 - The second integer is the
1571 minor version.
1572 ".args" sequence of Sequence of maps of the
1573 map kernel arguments. See
1574 :ref:`amdgpu-amdhsa-code-object-kernel-argument-metadata-map-table-v3`
1575 for the definition of the keys
1576 included in that map.
1577 ".reqd_workgroup_size" sequence of If not 0, 0, 0 then all values
1578 3 integers must be >=1 and the dispatch
1579 work-group size X, Y, Z must
1580 correspond to the specified
1581 values. Defaults to 0, 0, 0.
1583 Corresponds to the OpenCL
1584 ``reqd_work_group_size``
1585 attribute.
1586 ".workgroup_size_hint" sequence of The dispatch work-group size
1587 3 integers X, Y, Z is likely to be the
1588 specified values.
1590 Corresponds to the OpenCL
1591 ``work_group_size_hint``
1592 attribute.
1593 ".vec_type_hint" string The name of a scalar or vector
1594 type.
1596 Corresponds to the OpenCL
1597 ``vec_type_hint`` attribute.
1599 ".device_enqueue_symbol" string The external symbol name
1600 associated with a kernel.
1601 OpenCL runtime allocates a
1602 global buffer for the symbol
1603 and saves the kernel's address
1604 to it, which is used for
1605 device side enqueueing. Only
1606 available for device side
1607 enqueued kernels.
1608 ".kernarg_segment_size" integer Required The size in bytes of
1609 the kernarg segment
1610 that holds the values
1611 of the arguments to
1612 the kernel.
1613 ".group_segment_fixed_size" integer Required The amount of group
1614 segment memory
1615 required by a
1616 work-group in
1617 bytes. This does not
1618 include any
1619 dynamically allocated
1620 group segment memory
1621 that may be added
1622 when the kernel is
1623 dispatched.
1624 ".private_segment_fixed_size" integer Required The amount of fixed
1625 private address space
1626 memory required for a
1627 work-item in
1628 bytes. If the kernel
1629 uses a dynamic call
1630 stack then additional
1631 space must be added
1632 to this value for the
1633 call stack.
1634 ".kernarg_segment_align" integer Required The maximum byte
1635 alignment of
1636 arguments in the
1637 kernarg segment. Must
1638 be a power of 2.
1639 ".wavefront_size" integer Required Wavefront size. Must
1640 be a power of 2.
1641 ".sgpr_count" integer Required Number of scalar
1642 registers required by a
1643 wavefront for
1644 GFX6-GFX9. A register
1645 is required if it is
1646 used explicitly, or
1647 if a higher numbered
1648 register is used
1649 explicitly. This
1650 includes the special
1651 SGPRs for VCC, Flat
1652 Scratch (GFX7-GFX9)
1653 and XNACK (for
1654 GFX8-GFX9). It does
1655 not include the 16
1656 SGPR added if a trap
1657 handler is
1658 enabled. It is not
1659 rounded up to the
1660 allocation
1661 granularity.
1662 ".vgpr_count" integer Required Number of vector
1663 registers required by
1664 each work-item for
1665 GFX6-GFX9. A register
1666 is required if it is
1667 used explicitly, or
1668 if a higher numbered
1669 register is used
1670 explicitly.
1671 ".max_flat_workgroup_size" integer Required Maximum flat
1672 work-group size
1673 supported by the
1674 kernel in work-items.
1675 Must be >=1 and
1676 consistent with
1677 ReqdWorkGroupSize if
1678 not 0, 0, 0.
1679 ".sgpr_spill_count" integer Number of stores from
1680 a scalar register to
1681 a register allocator
1682 created spill
1683 location.
1684 ".vgpr_spill_count" integer Number of stores from
1685 a vector register to
1686 a register allocator
1687 created spill
1688 location.
1689 =================================== ============== ========= ================================
1691 ..
1693 .. table:: AMDHSA Code Object V3 Kernel Argument Metadata Map
1694 :name: amdgpu-amdhsa-code-object-kernel-argument-metadata-map-table-v3
1696 ====================== ============== ========= ================================
1697 String Key Value Type Required? Description
1698 ====================== ============== ========= ================================
1699 ".name" string Kernel argument name.
1700 ".type_name" string Kernel argument type name.
1701 ".size" integer Required Kernel argument size in bytes.
1702 ".offset" integer Required Kernel argument offset in
1703 bytes. The offset must be a
1704 multiple of the alignment
1705 required by the argument.
1706 ".value_kind" string Required Kernel argument kind that
1707 specifies how to set up the
1708 corresponding argument.
1709 Values include:
1711 "by_value"
1712 The argument is copied
1713 directly into the kernarg.
1715 "global_buffer"
1716 A global address space pointer
1717 to the buffer data is passed
1718 in the kernarg.
1720 "dynamic_shared_pointer"
1721 A group address space pointer
1722 to dynamically allocated LDS
1723 is passed in the kernarg.
1725 "sampler"
1726 A global address space
1727 pointer to a S# is passed in
1728 the kernarg.
1730 "image"
1731 A global address space
1732 pointer to a T# is passed in
1733 the kernarg.
1735 "pipe"
1736 A global address space pointer
1737 to an OpenCL pipe is passed in
1738 the kernarg.
1740 "queue"
1741 A global address space pointer
1742 to an OpenCL device enqueue
1743 queue is passed in the
1744 kernarg.
1746 "hidden_global_offset_x"
1747 The OpenCL grid dispatch
1748 global offset for the X
1749 dimension is passed in the
1750 kernarg.
1752 "hidden_global_offset_y"
1753 The OpenCL grid dispatch
1754 global offset for the Y
1755 dimension is passed in the
1756 kernarg.
1758 "hidden_global_offset_z"
1759 The OpenCL grid dispatch
1760 global offset for the Z
1761 dimension is passed in the
1762 kernarg.
1764 "hidden_none"
1765 An argument that is not used
1766 by the kernel. Space needs to
1767 be left for it, but it does
1768 not need to be set up.
1770 "hidden_printf_buffer"
1771 A global address space pointer
1772 to the runtime printf buffer
1773 is passed in kernarg.
1775 "hidden_default_queue"
1776 A global address space pointer
1777 to the OpenCL device enqueue
1778 queue that should be used by
1779 the kernel by default is
1780 passed in the kernarg.
1782 "hidden_completion_action"
1783 A global address space pointer
1784 to help link enqueued kernels into
1785 the ancestor tree for determining
1786 when the parent kernel has finished.
1788 ".value_type" string Required Kernel argument value type. Only
1789 present if ".value_kind" is
1790 "by_value". For vector data
1791 types, the value is for the
1792 element type. Values include:
1794 - "struct"
1795 - "i8"
1796 - "u8"
1797 - "i16"
1798 - "u16"
1799 - "f16"
1800 - "i32"
1801 - "u32"
1802 - "f32"
1803 - "i64"
1804 - "u64"
1805 - "f64"
1807 .. TODO
1808 How can it be determined if a
1809 vector type, and what size
1810 vector?
1811 ".pointee_align" integer Alignment in bytes of pointee
1812 type for pointer type kernel
1813 argument. Must be a power
1814 of 2. Only present if
1815 ".value_kind" is
1816 "dynamic_shared_pointer".
1817 ".address_space" string Kernel argument address space
1818 qualifier. Only present if
1819 ".value_kind" is "global_buffer" or
1820 "dynamic_shared_pointer". Values
1821 are:
1823 - "private"
1824 - "global"
1825 - "constant"
1826 - "local"
1827 - "generic"
1828 - "region"
1830 .. TODO
1831 Is "global_buffer" only "global"
1832 or "constant"? Is
1833 "dynamic_shared_pointer" always
1834 "local"? Can HCC allow "generic"?
1835 How can "private" or "region"
1836 ever happen?
1837 ".access" string Kernel argument access
1838 qualifier. Only present if
1839 ".value_kind" is "image" or
1840 "pipe". Values
1841 are:
1843 - "read_only"
1844 - "write_only"
1845 - "read_write"
1847 .. TODO
1848 Does this apply to
1849 "global_buffer"?
1850 ".actual_access" string The actual memory accesses
1851 performed by the kernel on the
1852 kernel argument. Only present if
1853 ".value_kind" is "global_buffer",
1854 "image", or "pipe". This may be
1855 more restrictive than indicated
1856 by ".access" to reflect what the
1857 kernel actual does. If not
1858 present then the runtime must
1859 assume what is implied by
1860 ".access" and ".is_const" . Values
1861 are:
1863 - "read_only"
1864 - "write_only"
1865 - "read_write"
1867 ".is_const" boolean Indicates if the kernel argument
1868 is const qualified. Only present
1869 if ".value_kind" is
1870 "global_buffer".
1872 ".is_restrict" boolean Indicates if the kernel argument
1873 is restrict qualified. Only
1874 present if ".value_kind" is
1875 "global_buffer".
1877 ".is_volatile" boolean Indicates if the kernel argument
1878 is volatile qualified. Only
1879 present if ".value_kind" is
1880 "global_buffer".
1882 ".is_pipe" boolean Indicates if the kernel argument
1883 is pipe qualified. Only present
1884 if ".value_kind" is "pipe".
1886 .. TODO
1887 Can "global_buffer" be pipe
1888 qualified?
1889 ====================== ============== ========= ================================
1891 ..
1893 Kernel Dispatch
1894 ~~~~~~~~~~~~~~~
1896 The HSA architected queuing language (AQL) defines a user space memory interface
1897 that can be used to control the dispatch of kernels, in an agent independent
1898 way. An agent can have zero or more AQL queues created for it using the ROCm
1899 runtime, in which AQL packets (all of which are 64 bytes) can be placed. See the
1900 *HSA Platform System Architecture Specification* [HSA]_ for the AQL queue
1901 mechanics and packet layouts.
1903 The packet processor of a kernel agent is responsible for detecting and
1904 dispatching HSA kernels from the AQL queues associated with it. For AMD GPUs the
1905 packet processor is implemented by the hardware command processor (CP),
1906 asynchronous dispatch controller (ADC) and shader processor input controller
1907 (SPI).
1909 The ROCm runtime can be used to allocate an AQL queue object. It uses the kernel
1910 mode driver to initialize and register the AQL queue with CP.
1912 To dispatch a kernel the following actions are performed. This can occur in the
1913 CPU host program, or from an HSA kernel executing on a GPU.
1915 1. A pointer to an AQL queue for the kernel agent on which the kernel is to be
1916 executed is obtained.
1917 2. A pointer to the kernel descriptor (see
1918 :ref:`amdgpu-amdhsa-kernel-descriptor`) of the kernel to execute is
1919 obtained. It must be for a kernel that is contained in a code object that that
1920 was loaded by the ROCm runtime on the kernel agent with which the AQL queue is
1921 associated.
1922 3. Space is allocated for the kernel arguments using the ROCm runtime allocator
1923 for a memory region with the kernarg property for the kernel agent that will
1924 execute the kernel. It must be at least 16 byte aligned.
1925 4. Kernel argument values are assigned to the kernel argument memory
1926 allocation. The layout is defined in the *HSA Programmer's Language Reference*
1927 [HSA]_. For AMDGPU the kernel execution directly accesses the kernel argument
1928 memory in the same way constant memory is accessed. (Note that the HSA
1929 specification allows an implementation to copy the kernel argument contents to
1930 another location that is accessed by the kernel.)
1931 5. An AQL kernel dispatch packet is created on the AQL queue. The ROCm runtime
1932 api uses 64 bit atomic operations to reserve space in the AQL queue for the
1933 packet. The packet must be set up, and the final write must use an atomic
1934 store release to set the packet kind to ensure the packet contents are
1935 visible to the kernel agent. AQL defines a doorbell signal mechanism to
1936 notify the kernel agent that the AQL queue has been updated. These rules, and
1937 the layout of the AQL queue and kernel dispatch packet is defined in the *HSA
1938 System Architecture Specification* [HSA]_.
1939 6. A kernel dispatch packet includes information about the actual dispatch,
1940 such as grid and work-group size, together with information from the code
1941 object about the kernel, such as segment sizes. The ROCm runtime queries on
1942 the kernel symbol can be used to obtain the code object values which are
1943 recorded in the :ref:`amdgpu-amdhsa-code-object-metadata`.
1944 7. CP executes micro-code and is responsible for detecting and setting up the
1945 GPU to execute the wavefronts of a kernel dispatch.
1946 8. CP ensures that when the a wavefront starts executing the kernel machine
1947 code, the scalar general purpose registers (SGPR) and vector general purpose
1948 registers (VGPR) are set up as required by the machine code. The required
1949 setup is defined in the :ref:`amdgpu-amdhsa-kernel-descriptor`. The initial
1950 register state is defined in
1951 :ref:`amdgpu-amdhsa-initial-kernel-execution-state`.
1952 9. The prolog of the kernel machine code (see
1953 :ref:`amdgpu-amdhsa-kernel-prolog`) sets up the machine state as necessary
1954 before continuing executing the machine code that corresponds to the kernel.
1955 10. When the kernel dispatch has completed execution, CP signals the completion
1956 signal specified in the kernel dispatch packet if not 0.
1958 .. _amdgpu-amdhsa-memory-spaces:
1960 Memory Spaces
1961 ~~~~~~~~~~~~~
1963 The memory space properties are:
1965 .. table:: AMDHSA Memory Spaces
1966 :name: amdgpu-amdhsa-memory-spaces-table
1968 ================= =========== ======== ======= ==================
1969 Memory Space Name HSA Segment Hardware Address NULL Value
1970 Name Name Size
1971 ================= =========== ======== ======= ==================
1972 Private private scratch 32 0x00000000
1973 Local group LDS 32 0xFFFFFFFF
1974 Global global global 64 0x0000000000000000
1975 Constant constant *same as 64 0x0000000000000000
1976 global*
1977 Generic flat flat 64 0x0000000000000000
1978 Region N/A GDS 32 *not implemented
1979 for AMDHSA*
1980 ================= =========== ======== ======= ==================
1982 The global and constant memory spaces both use global virtual addresses, which
1983 are the same virtual address space used by the CPU. However, some virtual
1984 addresses may only be accessible to the CPU, some only accessible by the GPU,
1985 and some by both.
1987 Using the constant memory space indicates that the data will not change during
1988 the execution of the kernel. This allows scalar read instructions to be
1989 used. The vector and scalar L1 caches are invalidated of volatile data before
1990 each kernel dispatch execution to allow constant memory to change values between
1991 kernel dispatches.
1993 The local memory space uses the hardware Local Data Store (LDS) which is
1994 automatically allocated when the hardware creates work-groups of wavefronts, and
1995 freed when all the wavefronts of a work-group have terminated. The data store
1996 (DS) instructions can be used to access it.
1998 The private memory space uses the hardware scratch memory support. If the kernel
1999 uses scratch, then the hardware allocates memory that is accessed using
2000 wavefront lane dword (4 byte) interleaving. The mapping used from private
2001 address to physical address is:
2003 ``wavefront-scratch-base +
2004 (private-address * wavefront-size * 4) +
2005 (wavefront-lane-id * 4)``
2007 There are different ways that the wavefront scratch base address is determined
2008 by a wavefront (see :ref:`amdgpu-amdhsa-initial-kernel-execution-state`). This
2009 memory can be accessed in an interleaved manner using buffer instruction with
2010 the scratch buffer descriptor and per wavefront scratch offset, by the scratch
2011 instructions, or by flat instructions. If each lane of a wavefront accesses the
2012 same private address, the interleaving results in adjacent dwords being accessed
2013 and hence requires fewer cache lines to be fetched. Multi-dword access is not
2014 supported except by flat and scratch instructions in GFX9.
2016 The generic address space uses the hardware flat address support available in
2017 GFX7-GFX9. This uses two fixed ranges of virtual addresses (the private and
2018 local appertures), that are outside the range of addressible global memory, to
2019 map from a flat address to a private or local address.
2021 FLAT instructions can take a flat address and access global, private (scratch)
2022 and group (LDS) memory depending in if the address is within one of the
2023 apperture ranges. Flat access to scratch requires hardware aperture setup and
2024 setup in the kernel prologue (see :ref:`amdgpu-amdhsa-flat-scratch`). Flat
2025 access to LDS requires hardware aperture setup and M0 (GFX7-GFX8) register setup
2026 (see :ref:`amdgpu-amdhsa-m0`).
2028 To convert between a segment address and a flat address the base address of the
2029 appertures address can be used. For GFX7-GFX8 these are available in the
2030 :ref:`amdgpu-amdhsa-hsa-aql-queue` the address of which can be obtained with
2031 Queue Ptr SGPR (see :ref:`amdgpu-amdhsa-initial-kernel-execution-state`). For
2032 GFX9 the appature base addresses are directly available as inline constant
2033 registers ``SRC_SHARED_BASE/LIMIT`` and ``SRC_PRIVATE_BASE/LIMIT``. In 64 bit
2034 address mode the apperture sizes are 2^32 bytes and the base is aligned to 2^32
2035 which makes it easier to convert from flat to segment or segment to flat.
2037 Image and Samplers
2038 ~~~~~~~~~~~~~~~~~~
2040 Image and sample handles created by the ROCm runtime are 64 bit addresses of a
2041 hardware 32 byte V# and 48 byte S# object respectively. In order to support the
2042 HSA ``query_sampler`` operations two extra dwords are used to store the HSA BRIG
2043 enumeration values for the queries that are not trivially deducible from the S#
2044 representation.
2046 HSA Signals
2047 ~~~~~~~~~~~
2049 HSA signal handles created by the ROCm runtime are 64 bit addresses of a
2050 structure allocated in memory accessible from both the CPU and GPU. The
2051 structure is defined by the ROCm runtime and subject to change between releases
2052 (see [AMD-ROCm-github]_).
2054 .. _amdgpu-amdhsa-hsa-aql-queue:
2056 HSA AQL Queue
2057 ~~~~~~~~~~~~~
2059 The HSA AQL queue structure is defined by the ROCm runtime and subject to change
2060 between releases (see [AMD-ROCm-github]_). For some processors it contains
2061 fields needed to implement certain language features such as the flat address
2062 aperture bases. It also contains fields used by CP such as managing the
2063 allocation of scratch memory.
2065 .. _amdgpu-amdhsa-kernel-descriptor:
2067 Kernel Descriptor
2068 ~~~~~~~~~~~~~~~~~
2070 A kernel descriptor consists of the information needed by CP to initiate the
2071 execution of a kernel, including the entry point address of the machine code
2072 that implements the kernel.
2074 Kernel Descriptor for GFX6-GFX9
2075 +++++++++++++++++++++++++++++++
2077 CP microcode requires the Kernel descriptor to be allocated on 64 byte
2078 alignment.
2080 .. table:: Kernel Descriptor for GFX6-GFX9
2081 :name: amdgpu-amdhsa-kernel-descriptor-gfx6-gfx9-table
2083 ======= ======= =============================== ============================
2084 Bits Size Field Name Description
2085 ======= ======= =============================== ============================
2086 31:0 4 bytes GROUP_SEGMENT_FIXED_SIZE The amount of fixed local
2087 address space memory
2088 required for a work-group
2089 in bytes. This does not
2090 include any dynamically
2091 allocated local address
2092 space memory that may be
2093 added when the kernel is
2094 dispatched.
2095 63:32 4 bytes PRIVATE_SEGMENT_FIXED_SIZE The amount of fixed
2096 private address space
2097 memory required for a
2098 work-item in bytes. If
2099 is_dynamic_callstack is 1
2100 then additional space must
2101 be added to this value for
2102 the call stack.
2103 127:64 8 bytes Reserved, must be 0.
2104 191:128 8 bytes KERNEL_CODE_ENTRY_BYTE_OFFSET Byte offset (possibly
2105 negative) from base
2106 address of kernel
2107 descriptor to kernel's
2108 entry point instruction
2109 which must be 256 byte
2110 aligned.
2111 383:192 24 Reserved, must be 0.
2112 bytes
2113 415:384 4 bytes COMPUTE_PGM_RSRC1 Compute Shader (CS)
2114 program settings used by
2115 CP to set up
2116 ``COMPUTE_PGM_RSRC1``
2117 configuration
2118 register. See
2119 :ref:`amdgpu-amdhsa-compute_pgm_rsrc1-gfx6-gfx9-table`.
2120 447:416 4 bytes COMPUTE_PGM_RSRC2 Compute Shader (CS)
2121 program settings used by
2122 CP to set up
2123 ``COMPUTE_PGM_RSRC2``
2124 configuration
2125 register. See
2126 :ref:`amdgpu-amdhsa-compute_pgm_rsrc2-gfx6-gfx9-table`.
2127 448 1 bit ENABLE_SGPR_PRIVATE_SEGMENT Enable the setup of the
2128 _BUFFER SGPR user data registers
2129 (see
2130 :ref:`amdgpu-amdhsa-initial-kernel-execution-state`).
2132 The total number of SGPR
2133 user data registers
2134 requested must not exceed
2135 16 and match value in
2136 ``compute_pgm_rsrc2.user_sgpr.user_sgpr_count``.
2137 Any requests beyond 16
2138 will be ignored.
2139 449 1 bit ENABLE_SGPR_DISPATCH_PTR *see above*
2140 450 1 bit ENABLE_SGPR_QUEUE_PTR *see above*
2141 451 1 bit ENABLE_SGPR_KERNARG_SEGMENT_PTR *see above*
2142 452 1 bit ENABLE_SGPR_DISPATCH_ID *see above*
2143 453 1 bit ENABLE_SGPR_FLAT_SCRATCH_INIT *see above*
2144 454 1 bit ENABLE_SGPR_PRIVATE_SEGMENT *see above*
2145 _SIZE
2146 455 1 bit Reserved, must be 0.
2147 511:456 8 bytes Reserved, must be 0.
2148 512 **Total size 64 bytes.**
2149 ======= ====================================================================
2151 ..
2153 .. table:: compute_pgm_rsrc1 for GFX6-GFX9
2154 :name: amdgpu-amdhsa-compute_pgm_rsrc1-gfx6-gfx9-table
2156 ======= ======= =============================== ===========================================================================
2157 Bits Size Field Name Description
2158 ======= ======= =============================== ===========================================================================
2159 5:0 6 bits GRANULATED_WORKITEM_VGPR_COUNT Number of vector register
2160 blocks used by each work-item;
2161 granularity is device
2162 specific:
2164 GFX6-GFX9
2165 - vgprs_used 0..256
2166 - max(0, ceil(vgprs_used / 4) - 1)
2168 Where vgprs_used is defined
2169 as the highest VGPR number
2170 explicitly referenced plus
2171 one.
2173 Used by CP to set up
2174 ``COMPUTE_PGM_RSRC1.VGPRS``.
2176 The
2177 :ref:`amdgpu-assembler`
2178 calculates this
2179 automatically for the
2180 selected processor from
2181 values provided to the
2182 `.amdhsa_kernel` directive
2183 by the
2184 `.amdhsa_next_free_vgpr`
2185 nested directive (see
2186 :ref:`amdhsa-kernel-directives-table`).
2187 9:6 4 bits GRANULATED_WAVEFRONT_SGPR_COUNT Number of scalar register
2188 blocks used by a wavefront;
2189 granularity is device
2190 specific:
2192 GFX6-GFX8
2193 - sgprs_used 0..112
2194 - max(0, ceil(sgprs_used / 8) - 1)
2195 GFX9
2196 - sgprs_used 0..112
2197 - 2 * max(0, ceil(sgprs_used / 16) - 1)
2199 Where sgprs_used is
2200 defined as the highest
2201 SGPR number explicitly
2202 referenced plus one, plus
2203 a target-specific number
2204 of additional special
2205 SGPRs for VCC,
2206 FLAT_SCRATCH (GFX7+) and
2207 XNACK_MASK (GFX8+), and
2208 any additional
2209 target-specific
2210 limitations. It does not
2211 include the 16 SGPRs added
2212 if a trap handler is
2213 enabled.
2215 The target-specific
2216 limitations and special
2217 SGPR layout are defined in
2218 the hardware
2219 documentation, which can
2220 be found in the
2221 :ref:`amdgpu-processors`
2222 table.
2224 Used by CP to set up
2225 ``COMPUTE_PGM_RSRC1.SGPRS``.
2227 The
2228 :ref:`amdgpu-assembler`
2229 calculates this
2230 automatically for the
2231 selected processor from
2232 values provided to the
2233 `.amdhsa_kernel` directive
2234 by the
2235 `.amdhsa_next_free_sgpr`
2236 and `.amdhsa_reserve_*`
2237 nested directives (see
2238 :ref:`amdhsa-kernel-directives-table`).
2239 11:10 2 bits PRIORITY Must be 0.
2241 Start executing wavefront
2242 at the specified priority.
2244 CP is responsible for
2245 filling in
2246 ``COMPUTE_PGM_RSRC1.PRIORITY``.
2247 13:12 2 bits FLOAT_ROUND_MODE_32 Wavefront starts execution
2248 with specified rounding
2249 mode for single (32
2250 bit) floating point
2251 precision floating point
2252 operations.
2254 Floating point rounding
2255 mode values are defined in
2256 :ref:`amdgpu-amdhsa-floating-point-rounding-mode-enumeration-values-table`.
2258 Used by CP to set up
2259 ``COMPUTE_PGM_RSRC1.FLOAT_MODE``.
2260 15:14 2 bits FLOAT_ROUND_MODE_16_64 Wavefront starts execution
2261 with specified rounding
2262 denorm mode for half/double (16
2263 and 64 bit) floating point
2264 precision floating point
2265 operations.
2267 Floating point rounding
2268 mode values are defined in
2269 :ref:`amdgpu-amdhsa-floating-point-rounding-mode-enumeration-values-table`.
2271 Used by CP to set up
2272 ``COMPUTE_PGM_RSRC1.FLOAT_MODE``.
2273 17:16 2 bits FLOAT_DENORM_MODE_32 Wavefront starts execution
2274 with specified denorm mode
2275 for single (32
2276 bit) floating point
2277 precision floating point
2278 operations.
2280 Floating point denorm mode
2281 values are defined in
2282 :ref:`amdgpu-amdhsa-floating-point-denorm-mode-enumeration-values-table`.
2284 Used by CP to set up
2285 ``COMPUTE_PGM_RSRC1.FLOAT_MODE``.
2286 19:18 2 bits FLOAT_DENORM_MODE_16_64 Wavefront starts execution
2287 with specified denorm mode
2288 for half/double (16
2289 and 64 bit) floating point
2290 precision floating point
2291 operations.
2293 Floating point denorm mode
2294 values are defined in
2295 :ref:`amdgpu-amdhsa-floating-point-denorm-mode-enumeration-values-table`.
2297 Used by CP to set up
2298 ``COMPUTE_PGM_RSRC1.FLOAT_MODE``.
2299 20 1 bit PRIV Must be 0.
2301 Start executing wavefront
2302 in privilege trap handler
2303 mode.
2305 CP is responsible for
2306 filling in
2307 ``COMPUTE_PGM_RSRC1.PRIV``.
2308 21 1 bit ENABLE_DX10_CLAMP Wavefront starts execution
2309 with DX10 clamp mode
2310 enabled. Used by the vector
2311 ALU to force DX10 style
2312 treatment of NaN's (when
2313 set, clamp NaN to zero,
2314 otherwise pass NaN
2315 through).
2317 Used by CP to set up
2318 ``COMPUTE_PGM_RSRC1.DX10_CLAMP``.
2319 22 1 bit DEBUG_MODE Must be 0.
2321 Start executing wavefront
2322 in single step mode.
2324 CP is responsible for
2325 filling in
2326 ``COMPUTE_PGM_RSRC1.DEBUG_MODE``.
2327 23 1 bit ENABLE_IEEE_MODE Wavefront starts execution
2328 with IEEE mode
2329 enabled. Floating point
2330 opcodes that support
2331 exception flag gathering
2332 will quiet and propagate
2333 signaling-NaN inputs per
2334 IEEE 754-2008. Min_dx10 and
2335 max_dx10 become IEEE
2336 754-2008 compliant due to
2337 signaling-NaN propagation
2338 and quieting.
2340 Used by CP to set up
2341 ``COMPUTE_PGM_RSRC1.IEEE_MODE``.
2342 24 1 bit BULKY Must be 0.
2344 Only one work-group allowed
2345 to execute on a compute
2346 unit.
2348 CP is responsible for
2349 filling in
2350 ``COMPUTE_PGM_RSRC1.BULKY``.
2351 25 1 bit CDBG_USER Must be 0.
2353 Flag that can be used to
2354 control debugging code.
2356 CP is responsible for
2357 filling in
2358 ``COMPUTE_PGM_RSRC1.CDBG_USER``.
2359 26 1 bit FP16_OVFL GFX6-GFX8
2360 Reserved, must be 0.
2361 GFX9
2362 Wavefront starts execution
2363 with specified fp16 overflow
2364 mode.
2366 - If 0, fp16 overflow generates
2367 +/-INF values.
2368 - If 1, fp16 overflow that is the
2369 result of an +/-INF input value
2370 or divide by 0 produces a +/-INF,
2371 otherwise clamps computed
2372 overflow to +/-MAX_FP16 as
2373 appropriate.
2375 Used by CP to set up
2376 ``COMPUTE_PGM_RSRC1.FP16_OVFL``.
2377 31:27 5 bits Reserved, must be 0.
2378 32 **Total size 4 bytes**
2379 ======= ===================================================================================================================
2381 ..
2383 .. table:: compute_pgm_rsrc2 for GFX6-GFX9
2384 :name: amdgpu-amdhsa-compute_pgm_rsrc2-gfx6-gfx9-table
2386 ======= ======= =============================== ===========================================================================
2387 Bits Size Field Name Description
2388 ======= ======= =============================== ===========================================================================
2389 0 1 bit ENABLE_SGPR_PRIVATE_SEGMENT Enable the setup of the
2390 _WAVEFRONT_OFFSET SGPR wavefront scratch offset
2391 system register (see
2392 :ref:`amdgpu-amdhsa-initial-kernel-execution-state`).
2394 Used by CP to set up
2395 ``COMPUTE_PGM_RSRC2.SCRATCH_EN``.
2396 5:1 5 bits USER_SGPR_COUNT The total number of SGPR
2397 user data registers
2398 requested. This number must
2399 match the number of user
2400 data registers enabled.
2402 Used by CP to set up
2403 ``COMPUTE_PGM_RSRC2.USER_SGPR``.
2404 6 1 bit ENABLE_TRAP_HANDLER Must be 0.
2406 This bit represents
2407 ``COMPUTE_PGM_RSRC2.TRAP_PRESENT``,
2408 which is set by the CP if
2409 the runtime has installed a
2410 trap handler.
2411 7 1 bit ENABLE_SGPR_WORKGROUP_ID_X Enable the setup of the
2412 system SGPR register for
2413 the work-group id in the X
2414 dimension (see
2415 :ref:`amdgpu-amdhsa-initial-kernel-execution-state`).
2417 Used by CP to set up
2418 ``COMPUTE_PGM_RSRC2.TGID_X_EN``.
2419 8 1 bit ENABLE_SGPR_WORKGROUP_ID_Y Enable the setup of the
2420 system SGPR register for
2421 the work-group id in the Y
2422 dimension (see
2423 :ref:`amdgpu-amdhsa-initial-kernel-execution-state`).
2425 Used by CP to set up
2426 ``COMPUTE_PGM_RSRC2.TGID_Y_EN``.
2427 9 1 bit ENABLE_SGPR_WORKGROUP_ID_Z Enable the setup of the
2428 system SGPR register for
2429 the work-group id in the Z
2430 dimension (see
2431 :ref:`amdgpu-amdhsa-initial-kernel-execution-state`).
2433 Used by CP to set up
2434 ``COMPUTE_PGM_RSRC2.TGID_Z_EN``.
2435 10 1 bit ENABLE_SGPR_WORKGROUP_INFO Enable the setup of the
2436 system SGPR register for
2437 work-group information (see
2438 :ref:`amdgpu-amdhsa-initial-kernel-execution-state`).
2440 Used by CP to set up
2441 ``COMPUTE_PGM_RSRC2.TGID_SIZE_EN``.
2442 12:11 2 bits ENABLE_VGPR_WORKITEM_ID Enable the setup of the
2443 VGPR system registers used
2444 for the work-item ID.
2445 :ref:`amdgpu-amdhsa-system-vgpr-work-item-id-enumeration-values-table`
2446 defines the values.
2448 Used by CP to set up
2449 ``COMPUTE_PGM_RSRC2.TIDIG_CMP_CNT``.
2450 13 1 bit ENABLE_EXCEPTION_ADDRESS_WATCH Must be 0.
2452 Wavefront starts execution
2453 with address watch
2454 exceptions enabled which
2455 are generated when L1 has
2456 witnessed a thread access
2457 an *address of
2458 interest*.
2460 CP is responsible for
2461 filling in the address
2462 watch bit in
2463 ``COMPUTE_PGM_RSRC2.EXCP_EN_MSB``
2464 according to what the
2465 runtime requests.
2466 14 1 bit ENABLE_EXCEPTION_MEMORY Must be 0.
2468 Wavefront starts execution
2469 with memory violation
2470 exceptions exceptions
2471 enabled which are generated
2472 when a memory violation has
2473 occurred for this wavefront from
2474 L1 or LDS
2475 (write-to-read-only-memory,
2476 mis-aligned atomic, LDS
2477 address out of range,
2478 illegal address, etc.).
2480 CP sets the memory
2481 violation bit in
2482 ``COMPUTE_PGM_RSRC2.EXCP_EN_MSB``
2483 according to what the
2484 runtime requests.
2485 23:15 9 bits GRANULATED_LDS_SIZE Must be 0.
2487 CP uses the rounded value
2488 from the dispatch packet,
2489 not this value, as the
2490 dispatch may contain
2491 dynamically allocated group
2492 segment memory. CP writes
2493 directly to
2494 ``COMPUTE_PGM_RSRC2.LDS_SIZE``.
2496 Amount of group segment
2497 (LDS) to allocate for each
2498 work-group. Granularity is
2499 device specific:
2501 GFX6:
2502 roundup(lds-size / (64 * 4))
2503 GFX7-GFX9:
2504 roundup(lds-size / (128 * 4))
2506 24 1 bit ENABLE_EXCEPTION_IEEE_754_FP Wavefront starts execution
2507 _INVALID_OPERATION with specified exceptions
2508 enabled.
2510 Used by CP to set up
2511 ``COMPUTE_PGM_RSRC2.EXCP_EN``
2512 (set from bits 0..6).
2514 IEEE 754 FP Invalid
2515 Operation
2516 25 1 bit ENABLE_EXCEPTION_FP_DENORMAL FP Denormal one or more
2517 _SOURCE input operands is a
2518 denormal number
2519 26 1 bit ENABLE_EXCEPTION_IEEE_754_FP IEEE 754 FP Division by
2520 _DIVISION_BY_ZERO Zero
2521 27 1 bit ENABLE_EXCEPTION_IEEE_754_FP IEEE 754 FP FP Overflow
2522 _OVERFLOW
2523 28 1 bit ENABLE_EXCEPTION_IEEE_754_FP IEEE 754 FP Underflow
2524 _UNDERFLOW
2525 29 1 bit ENABLE_EXCEPTION_IEEE_754_FP IEEE 754 FP Inexact
2526 _INEXACT
2527 30 1 bit ENABLE_EXCEPTION_INT_DIVIDE_BY Integer Division by Zero
2528 _ZERO (rcp_iflag_f32 instruction
2529 only)
2530 31 1 bit Reserved, must be 0.
2531 32 **Total size 4 bytes.**
2532 ======= ===================================================================================================================
2534 ..
2536 .. table:: Floating Point Rounding Mode Enumeration Values
2537 :name: amdgpu-amdhsa-floating-point-rounding-mode-enumeration-values-table
2539 ====================================== ===== ==============================
2540 Enumeration Name Value Description
2541 ====================================== ===== ==============================
2542 FLOAT_ROUND_MODE_NEAR_EVEN 0 Round Ties To Even
2543 FLOAT_ROUND_MODE_PLUS_INFINITY 1 Round Toward +infinity
2544 FLOAT_ROUND_MODE_MINUS_INFINITY 2 Round Toward -infinity
2545 FLOAT_ROUND_MODE_ZERO 3 Round Toward 0
2546 ====================================== ===== ==============================
2548 ..
2550 .. table:: Floating Point Denorm Mode Enumeration Values
2551 :name: amdgpu-amdhsa-floating-point-denorm-mode-enumeration-values-table
2553 ====================================== ===== ==============================
2554 Enumeration Name Value Description
2555 ====================================== ===== ==============================
2556 FLOAT_DENORM_MODE_FLUSH_SRC_DST 0 Flush Source and Destination
2557 Denorms
2558 FLOAT_DENORM_MODE_FLUSH_DST 1 Flush Output Denorms
2559 FLOAT_DENORM_MODE_FLUSH_SRC 2 Flush Source Denorms
2560 FLOAT_DENORM_MODE_FLUSH_NONE 3 No Flush
2561 ====================================== ===== ==============================
2563 ..
2565 .. table:: System VGPR Work-Item ID Enumeration Values
2566 :name: amdgpu-amdhsa-system-vgpr-work-item-id-enumeration-values-table
2568 ======================================== ===== ============================
2569 Enumeration Name Value Description
2570 ======================================== ===== ============================
2571 SYSTEM_VGPR_WORKITEM_ID_X 0 Set work-item X dimension
2572 ID.
2573 SYSTEM_VGPR_WORKITEM_ID_X_Y 1 Set work-item X and Y
2574 dimensions ID.
2575 SYSTEM_VGPR_WORKITEM_ID_X_Y_Z 2 Set work-item X, Y and Z
2576 dimensions ID.
2577 SYSTEM_VGPR_WORKITEM_ID_UNDEFINED 3 Undefined.
2578 ======================================== ===== ============================
2580 .. _amdgpu-amdhsa-initial-kernel-execution-state:
2582 Initial Kernel Execution State
2583 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2585 This section defines the register state that will be set up by the packet
2586 processor prior to the start of execution of every wavefront. This is limited by
2587 the constraints of the hardware controllers of CP/ADC/SPI.
2589 The order of the SGPR registers is defined, but the compiler can specify which
2590 ones are actually setup in the kernel descriptor using the ``enable_sgpr_*`` bit
2591 fields (see :ref:`amdgpu-amdhsa-kernel-descriptor`). The register numbers used
2592 for enabled registers are dense starting at SGPR0: the first enabled register is
2593 SGPR0, the next enabled register is SGPR1 etc.; disabled registers do not have
2594 an SGPR number.
2596 The initial SGPRs comprise up to 16 User SRGPs that are set by CP and apply to
2597 all wavefronts of the grid. It is possible to specify more than 16 User SGPRs using
2598 the ``enable_sgpr_*`` bit fields, in which case only the first 16 are actually
2599 initialized. These are then immediately followed by the System SGPRs that are
2600 set up by ADC/SPI and can have different values for each wavefront of the grid
2601 dispatch.
2603 SGPR register initial state is defined in
2604 :ref:`amdgpu-amdhsa-sgpr-register-set-up-order-table`.
2606 .. table:: SGPR Register Set Up Order
2607 :name: amdgpu-amdhsa-sgpr-register-set-up-order-table
2609 ========== ========================== ====== ==============================
2610 SGPR Order Name Number Description
2611 (kernel descriptor enable of
2612 field) SGPRs
2613 ========== ========================== ====== ==============================
2614 First Private Segment Buffer 4 V# that can be used, together
2615 (enable_sgpr_private with Scratch Wavefront Offset
2616 _segment_buffer) as an offset, to access the
2617 private memory space using a
2618 segment address.
2620 CP uses the value provided by
2621 the runtime.
2622 then Dispatch Ptr 2 64 bit address of AQL dispatch
2623 (enable_sgpr_dispatch_ptr) packet for kernel dispatch
2624 actually executing.
2625 then Queue Ptr 2 64 bit address of amd_queue_t
2626 (enable_sgpr_queue_ptr) object for AQL queue on which
2627 the dispatch packet was
2628 queued.
2629 then Kernarg Segment Ptr 2 64 bit address of Kernarg
2630 (enable_sgpr_kernarg segment. This is directly
2631 _segment_ptr) copied from the
2632 kernarg_address in the kernel
2633 dispatch packet.
2635 Having CP load it once avoids
2636 loading it at the beginning of
2637 every wavefront.
2638 then Dispatch Id 2 64 bit Dispatch ID of the
2639 (enable_sgpr_dispatch_id) dispatch packet being
2640 executed.
2641 then Flat Scratch Init 2 This is 2 SGPRs:
2642 (enable_sgpr_flat_scratch
2643 _init) GFX6
2644 Not supported.
2645 GFX7-GFX8
2646 The first SGPR is a 32 bit
2647 byte offset from
2648 ``SH_HIDDEN_PRIVATE_BASE_VIMID``
2649 to per SPI base of memory
2650 for scratch for the queue
2651 executing the kernel
2652 dispatch. CP obtains this
2653 from the runtime. (The
2654 Scratch Segment Buffer base
2655 address is
2656 ``SH_HIDDEN_PRIVATE_BASE_VIMID``
2657 plus this offset.) The value
2658 of Scratch Wavefront Offset must
2659 be added to this offset by
2660 the kernel machine code,
2661 right shifted by 8, and
2662 moved to the FLAT_SCRATCH_HI
2663 SGPR register.
2664 FLAT_SCRATCH_HI corresponds
2665 to SGPRn-4 on GFX7, and
2666 SGPRn-6 on GFX8 (where SGPRn
2667 is the highest numbered SGPR
2668 allocated to the wavefront).
2669 FLAT_SCRATCH_HI is
2670 multiplied by 256 (as it is
2671 in units of 256 bytes) and
2672 added to
2673 ``SH_HIDDEN_PRIVATE_BASE_VIMID``
2674 to calculate the per wavefront
2675 FLAT SCRATCH BASE in flat
2676 memory instructions that
2677 access the scratch
2678 apperture.
2680 The second SGPR is 32 bit
2681 byte size of a single
2682 work-item's scratch memory
2683 usage. CP obtains this from
2684 the runtime, and it is
2685 always a multiple of DWORD.
2686 CP checks that the value in
2687 the kernel dispatch packet
2688 Private Segment Byte Size is
2689 not larger, and requests the
2690 runtime to increase the
2691 queue's scratch size if
2692 necessary. The kernel code
2693 must move it to
2694 FLAT_SCRATCH_LO which is
2695 SGPRn-3 on GFX7 and SGPRn-5
2696 on GFX8. FLAT_SCRATCH_LO is
2697 used as the FLAT SCRATCH
2698 SIZE in flat memory
2699 instructions. Having CP load
2700 it once avoids loading it at
2701 the beginning of every
2702 wavefront.
2703 GFX9
2704 This is the
2705 64 bit base address of the
2706 per SPI scratch backing
2707 memory managed by SPI for
2708 the queue executing the
2709 kernel dispatch. CP obtains
2710 this from the runtime (and
2711 divides it if there are
2712 multiple Shader Arrays each
2713 with its own SPI). The value
2714 of Scratch Wavefront Offset must
2715 be added by the kernel
2716 machine code and the result
2717 moved to the FLAT_SCRATCH
2718 SGPR which is SGPRn-6 and
2719 SGPRn-5. It is used as the
2720 FLAT SCRATCH BASE in flat
2721 memory instructions.
2722 then Private Segment Size 1 The 32 bit byte size of a
2723 (enable_sgpr_private single
2724 work-item's
2725 scratch_segment_size) memory
2726 allocation. This is the
2727 value from the kernel
2728 dispatch packet Private
2729 Segment Byte Size rounded up
2730 by CP to a multiple of
2731 DWORD.
2733 Having CP load it once avoids
2734 loading it at the beginning of
2735 every wavefront.
2737 This is not used for
2738 GFX7-GFX8 since it is the same
2739 value as the second SGPR of
2740 Flat Scratch Init. However, it
2741 may be needed for GFX9 which
2742 changes the meaning of the
2743 Flat Scratch Init value.
2744 then Grid Work-Group Count X 1 32 bit count of the number of
2745 (enable_sgpr_grid work-groups in the X dimension
2746 _workgroup_count_X) for the grid being
2747 executed. Computed from the
2748 fields in the kernel dispatch
2749 packet as ((grid_size.x +
2750 workgroup_size.x - 1) /
2751 workgroup_size.x).
2752 then Grid Work-Group Count Y 1 32 bit count of the number of
2753 (enable_sgpr_grid work-groups in the Y dimension
2754 _workgroup_count_Y && for the grid being
2755 less than 16 previous executed. Computed from the
2756 SGPRs) fields in the kernel dispatch
2757 packet as ((grid_size.y +
2758 workgroup_size.y - 1) /
2759 workgroupSize.y).
2761 Only initialized if <16
2762 previous SGPRs initialized.
2763 then Grid Work-Group Count Z 1 32 bit count of the number of
2764 (enable_sgpr_grid work-groups in the Z dimension
2765 _workgroup_count_Z && for the grid being
2766 less than 16 previous executed. Computed from the
2767 SGPRs) fields in the kernel dispatch
2768 packet as ((grid_size.z +
2769 workgroup_size.z - 1) /
2770 workgroupSize.z).
2772 Only initialized if <16
2773 previous SGPRs initialized.
2774 then Work-Group Id X 1 32 bit work-group id in X
2775 (enable_sgpr_workgroup_id dimension of grid for
2776 _X) wavefront.
2777 then Work-Group Id Y 1 32 bit work-group id in Y
2778 (enable_sgpr_workgroup_id dimension of grid for
2779 _Y) wavefront.
2780 then Work-Group Id Z 1 32 bit work-group id in Z
2781 (enable_sgpr_workgroup_id dimension of grid for
2782 _Z) wavefront.
2783 then Work-Group Info 1 {first_wavefront, 14'b0000,
2784 (enable_sgpr_workgroup ordered_append_term[10:0],
2785 _info) threadgroup_size_in_wavefronts[5:0]}
2786 then Scratch Wavefront Offset 1 32 bit byte offset from base
2787 (enable_sgpr_private of scratch base of queue
2788 _segment_wavefront_offset) executing the kernel
2789 dispatch. Must be used as an
2790 offset with Private
2791 segment address when using
2792 Scratch Segment Buffer. It
2793 must be used to set up FLAT
2794 SCRATCH for flat addressing
2795 (see
2796 :ref:`amdgpu-amdhsa-flat-scratch`).
2797 ========== ========================== ====== ==============================
2799 The order of the VGPR registers is defined, but the compiler can specify which
2800 ones are actually setup in the kernel descriptor using the ``enable_vgpr*`` bit
2801 fields (see :ref:`amdgpu-amdhsa-kernel-descriptor`). The register numbers used
2802 for enabled registers are dense starting at VGPR0: the first enabled register is
2803 VGPR0, the next enabled register is VGPR1 etc.; disabled registers do not have a
2804 VGPR number.
2806 VGPR register initial state is defined in
2807 :ref:`amdgpu-amdhsa-vgpr-register-set-up-order-table`.
2809 .. table:: VGPR Register Set Up Order
2810 :name: amdgpu-amdhsa-vgpr-register-set-up-order-table
2812 ========== ========================== ====== ==============================
2813 VGPR Order Name Number Description
2814 (kernel descriptor enable of
2815 field) VGPRs
2816 ========== ========================== ====== ==============================
2817 First Work-Item Id X 1 32 bit work item id in X
2818 (Always initialized) dimension of work-group for
2819 wavefront lane.
2820 then Work-Item Id Y 1 32 bit work item id in Y
2821 (enable_vgpr_workitem_id dimension of work-group for
2822 > 0) wavefront lane.
2823 then Work-Item Id Z 1 32 bit work item id in Z
2824 (enable_vgpr_workitem_id dimension of work-group for
2825 > 1) wavefront lane.
2826 ========== ========================== ====== ==============================
2828 The setting of registers is done by GPU CP/ADC/SPI hardware as follows:
2830 1. SGPRs before the Work-Group Ids are set by CP using the 16 User Data
2831 registers.
2832 2. Work-group Id registers X, Y, Z are set by ADC which supports any
2833 combination including none.
2834 3. Scratch Wavefront Offset is set by SPI in a per wavefront basis which is why
2835 its value cannot included with the flat scratch init value which is per queue.
2836 4. The VGPRs are set by SPI which only supports specifying either (X), (X, Y)
2837 or (X, Y, Z).
2839 Flat Scratch register pair are adjacent SGRRs so they can be moved as a 64 bit
2840 value to the hardware required SGPRn-3 and SGPRn-4 respectively.
2842 The global segment can be accessed either using buffer instructions (GFX6 which
2843 has V# 64 bit address support), flat instructions (GFX7-GFX9), or global
2844 instructions (GFX9).
2846 If buffer operations are used then the compiler can generate a V# with the
2847 following properties:
2849 * base address of 0
2850 * no swizzle
2851 * ATC: 1 if IOMMU present (such as APU)
2852 * ptr64: 1
2853 * MTYPE set to support memory coherence that matches the runtime (such as CC for
2854 APU and NC for dGPU).
2856 .. _amdgpu-amdhsa-kernel-prolog:
2858 Kernel Prolog
2859 ~~~~~~~~~~~~~
2861 .. _amdgpu-amdhsa-m0:
2863 M0
2864 ++
2866 GFX6-GFX8
2867 The M0 register must be initialized with a value at least the total LDS size
2868 if the kernel may access LDS via DS or flat operations. Total LDS size is
2869 available in dispatch packet. For M0, it is also possible to use maximum
2870 possible value of LDS for given target (0x7FFF for GFX6 and 0xFFFF for
2871 GFX7-GFX8).
2872 GFX9
2873 The M0 register is not used for range checking LDS accesses and so does not
2874 need to be initialized in the prolog.
2876 .. _amdgpu-amdhsa-flat-scratch:
2878 Flat Scratch
2879 ++++++++++++
2881 If the kernel may use flat operations to access scratch memory, the prolog code
2882 must set up FLAT_SCRATCH register pair (FLAT_SCRATCH_LO/FLAT_SCRATCH_HI which
2883 are in SGPRn-4/SGPRn-3). Initialization uses Flat Scratch Init and Scratch Wavefront
2884 Offset SGPR registers (see :ref:`amdgpu-amdhsa-initial-kernel-execution-state`):
2886 GFX6
2887 Flat scratch is not supported.
2889 GFX7-GFX8
2890 1. The low word of Flat Scratch Init is 32 bit byte offset from
2891 ``SH_HIDDEN_PRIVATE_BASE_VIMID`` to the base of scratch backing memory
2892 being managed by SPI for the queue executing the kernel dispatch. This is
2893 the same value used in the Scratch Segment Buffer V# base address. The
2894 prolog must add the value of Scratch Wavefront Offset to get the wavefront's byte
2895 scratch backing memory offset from ``SH_HIDDEN_PRIVATE_BASE_VIMID``. Since
2896 FLAT_SCRATCH_LO is in units of 256 bytes, the offset must be right shifted
2897 by 8 before moving into FLAT_SCRATCH_LO.
2898 2. The second word of Flat Scratch Init is 32 bit byte size of a single
2899 work-items scratch memory usage. This is directly loaded from the kernel
2900 dispatch packet Private Segment Byte Size and rounded up to a multiple of
2901 DWORD. Having CP load it once avoids loading it at the beginning of every
2902 wavefront. The prolog must move it to FLAT_SCRATCH_LO for use as FLAT SCRATCH
2903 SIZE.
2905 GFX9
2906 The Flat Scratch Init is the 64 bit address of the base of scratch backing
2907 memory being managed by SPI for the queue executing the kernel dispatch. The
2908 prolog must add the value of Scratch Wavefront Offset and moved to the FLAT_SCRATCH
2909 pair for use as the flat scratch base in flat memory instructions.
2911 .. _amdgpu-amdhsa-memory-model:
2913 Memory Model
2914 ~~~~~~~~~~~~
2916 This section describes the mapping of LLVM memory model onto AMDGPU machine code
2917 (see :ref:`memmodel`). *The implementation is WIP.*
2919 .. TODO
2920 Update when implementation complete.
2922 The AMDGPU backend supports the memory synchronization scopes specified in
2923 :ref:`amdgpu-memory-scopes`.
2925 The code sequences used to implement the memory model are defined in table
2926 :ref:`amdgpu-amdhsa-memory-model-code-sequences-gfx6-gfx9-table`.
2928 The sequences specify the order of instructions that a single thread must
2929 execute. The ``s_waitcnt`` and ``buffer_wbinvl1_vol`` are defined with respect
2930 to other memory instructions executed by the same thread. This allows them to be
2931 moved earlier or later which can allow them to be combined with other instances
2932 of the same instruction, or hoisted/sunk out of loops to improve
2933 performance. Only the instructions related to the memory model are given;
2934 additional ``s_waitcnt`` instructions are required to ensure registers are
2935 defined before being used. These may be able to be combined with the memory
2936 model ``s_waitcnt`` instructions as described above.
2938 The AMDGPU backend supports the following memory models:
2940 HSA Memory Model [HSA]_
2941 The HSA memory model uses a single happens-before relation for all address
2942 spaces (see :ref:`amdgpu-address-spaces`).
2943 OpenCL Memory Model [OpenCL]_
2944 The OpenCL memory model which has separate happens-before relations for the
2945 global and local address spaces. Only a fence specifying both global and
2946 local address space, and seq_cst instructions join the relationships. Since
2947 the LLVM ``memfence`` instruction does not allow an address space to be
2948 specified the OpenCL fence has to convervatively assume both local and
2949 global address space was specified. However, optimizations can often be
2950 done to eliminate the additional ``s_waitcnt`` instructions when there are
2951 no intervening memory instructions which access the corresponding address
2952 space. The code sequences in the table indicate what can be omitted for the
2953 OpenCL memory. The target triple environment is used to determine if the
2954 source language is OpenCL (see :ref:`amdgpu-opencl`).
2956 ``ds/flat_load/store/atomic`` instructions to local memory are termed LDS
2957 operations.
2959 ``buffer/global/flat_load/store/atomic`` instructions to global memory are
2960 termed vector memory operations.
2962 For GFX6-GFX9:
2964 * Each agent has multiple compute units (CU).
2965 * Each CU has multiple SIMDs that execute wavefronts.
2966 * The wavefronts for a single work-group are executed in the same CU but may be
2967 executed by different SIMDs.
2968 * Each CU has a single LDS memory shared by the wavefronts of the work-groups
2969 executing on it.
2970 * All LDS operations of a CU are performed as wavefront wide operations in a
2971 global order and involve no caching. Completion is reported to a wavefront in
2972 execution order.
2973 * The LDS memory has multiple request queues shared by the SIMDs of a
2974 CU. Therefore, the LDS operations performed by different wavefronts of a work-group
2975 can be reordered relative to each other, which can result in reordering the
2976 visibility of vector memory operations with respect to LDS operations of other
2977 wavefronts in the same work-group. A ``s_waitcnt lgkmcnt(0)`` is required to
2978 ensure synchronization between LDS operations and vector memory operations
2979 between wavefronts of a work-group, but not between operations performed by the
2980 same wavefront.
2981 * The vector memory operations are performed as wavefront wide operations and
2982 completion is reported to a wavefront in execution order. The exception is
2983 that for GFX7-GFX9 ``flat_load/store/atomic`` instructions can report out of
2984 vector memory order if they access LDS memory, and out of LDS operation order
2985 if they access global memory.
2986 * The vector memory operations access a single vector L1 cache shared by all
2987 SIMDs a CU. Therefore, no special action is required for coherence between the
2988 lanes of a single wavefront, or for coherence between wavefronts in the same
2989 work-group. A ``buffer_wbinvl1_vol`` is required for coherence between wavefronts
2990 executing in different work-groups as they may be executing on different CUs.
2991 * The scalar memory operations access a scalar L1 cache shared by all wavefronts
2992 on a group of CUs. The scalar and vector L1 caches are not coherent. However,
2993 scalar operations are used in a restricted way so do not impact the memory
2994 model. See :ref:`amdgpu-amdhsa-memory-spaces`.
2995 * The vector and scalar memory operations use an L2 cache shared by all CUs on
2996 the same agent.
2997 * The L2 cache has independent channels to service disjoint ranges of virtual
2998 addresses.
2999 * Each CU has a separate request queue per channel. Therefore, the vector and
3000 scalar memory operations performed by wavefronts executing in different work-groups
3001 (which may be executing on different CUs) of an agent can be reordered
3002 relative to each other. A ``s_waitcnt vmcnt(0)`` is required to ensure
3003 synchronization between vector memory operations of different CUs. It ensures a
3004 previous vector memory operation has completed before executing a subsequent
3005 vector memory or LDS operation and so can be used to meet the requirements of
3006 acquire and release.
3007 * The L2 cache can be kept coherent with other agents on some targets, or ranges
3008 of virtual addresses can be set up to bypass it to ensure system coherence.
3010 Private address space uses ``buffer_load/store`` using the scratch V# (GFX6-GFX8),
3011 or ``scratch_load/store`` (GFX9). Since only a single thread is accessing the
3012 memory, atomic memory orderings are not meaningful and all accesses are treated
3013 as non-atomic.
3015 Constant address space uses ``buffer/global_load`` instructions (or equivalent
3016 scalar memory instructions). Since the constant address space contents do not
3017 change during the execution of a kernel dispatch it is not legal to perform
3018 stores, and atomic memory orderings are not meaningful and all access are
3019 treated as non-atomic.
3021 A memory synchronization scope wider than work-group is not meaningful for the
3022 group (LDS) address space and is treated as work-group.
3024 The memory model does not support the region address space which is treated as
3025 non-atomic.
3027 Acquire memory ordering is not meaningful on store atomic instructions and is
3028 treated as non-atomic.
3030 Release memory ordering is not meaningful on load atomic instructions and is
3031 treated a non-atomic.
3033 Acquire-release memory ordering is not meaningful on load or store atomic
3034 instructions and is treated as acquire and release respectively.
3036 AMDGPU backend only uses scalar memory operations to access memory that is
3037 proven to not change during the execution of the kernel dispatch. This includes
3038 constant address space and global address space for program scope const
3039 variables. Therefore the kernel machine code does not have to maintain the
3040 scalar L1 cache to ensure it is coherent with the vector L1 cache. The scalar
3041 and vector L1 caches are invalidated between kernel dispatches by CP since
3042 constant address space data may change between kernel dispatch executions. See
3043 :ref:`amdgpu-amdhsa-memory-spaces`.
3045 The one execption is if scalar writes are used to spill SGPR registers. In this
3046 case the AMDGPU backend ensures the memory location used to spill is never
3047 accessed by vector memory operations at the same time. If scalar writes are used
3048 then a ``s_dcache_wb`` is inserted before the ``s_endpgm`` and before a function
3049 return since the locations may be used for vector memory instructions by a
3050 future wavefront that uses the same scratch area, or a function call that creates a
3051 frame at the same address, respectively. There is no need for a ``s_dcache_inv``
3052 as all scalar writes are write-before-read in the same thread.
3054 Scratch backing memory (which is used for the private address space)
3055 is accessed with MTYPE NC_NV (non-coherenent non-volatile). Since the private
3056 address space is only accessed by a single thread, and is always
3057 write-before-read, there is never a need to invalidate these entries from the L1
3058 cache. Hence all cache invalidates are done as ``*_vol`` to only invalidate the
3059 volatile cache lines.
3061 On dGPU the kernarg backing memory is accessed as UC (uncached) to avoid needing
3062 to invalidate the L2 cache. This also causes it to be treated as
3063 non-volatile and so is not invalidated by ``*_vol``. On APU it is accessed as CC
3064 (cache coherent) and so the L2 cache will coherent with the CPU and other
3065 agents.
3067 .. table:: AMDHSA Memory Model Code Sequences GFX6-GFX9
3068 :name: amdgpu-amdhsa-memory-model-code-sequences-gfx6-gfx9-table
3070 ============ ============ ============== ========== ===============================
3071 LLVM Instr LLVM Memory LLVM Memory AMDGPU AMDGPU Machine Code
3072 Ordering Sync Scope Address
3073 Space
3074 ============ ============ ============== ========== ===============================
3075 **Non-Atomic**
3076 -----------------------------------------------------------------------------------
3077 load *none* *none* - global - !volatile & !nontemporal
3078 - generic
3079 - private 1. buffer/global/flat_load
3080 - constant
3081 - volatile & !nontemporal
3083 1. buffer/global/flat_load
3084 glc=1
3086 - nontemporal
3088 1. buffer/global/flat_load
3089 glc=1 slc=1
3091 load *none* *none* - local 1. ds_load
3092 store *none* *none* - global - !nontemporal
3093 - generic
3094 - private 1. buffer/global/flat_store
3095 - constant
3096 - nontemporal
3098 1. buffer/global/flat_stote
3099 glc=1 slc=1
3101 store *none* *none* - local 1. ds_store
3102 **Unordered Atomic**
3103 -----------------------------------------------------------------------------------
3104 load atomic unordered *any* *any* *Same as non-atomic*.
3105 store atomic unordered *any* *any* *Same as non-atomic*.
3106 atomicrmw unordered *any* *any* *Same as monotonic
3107 atomic*.
3108 **Monotonic Atomic**
3109 -----------------------------------------------------------------------------------
3110 load atomic monotonic - singlethread - global 1. buffer/global/flat_load
3111 - wavefront - generic
3112 - workgroup
3113 load atomic monotonic - singlethread - local 1. ds_load
3114 - wavefront
3115 - workgroup
3116 load atomic monotonic - agent - global 1. buffer/global/flat_load
3117 - system - generic glc=1
3118 store atomic monotonic - singlethread - global 1. buffer/global/flat_store
3119 - wavefront - generic
3120 - workgroup
3121 - agent
3122 - system
3123 store atomic monotonic - singlethread - local 1. ds_store
3124 - wavefront
3125 - workgroup
3126 atomicrmw monotonic - singlethread - global 1. buffer/global/flat_atomic
3127 - wavefront - generic
3128 - workgroup
3129 - agent
3130 - system
3131 atomicrmw monotonic - singlethread - local 1. ds_atomic
3132 - wavefront
3133 - workgroup
3134 **Acquire Atomic**
3135 -----------------------------------------------------------------------------------
3136 load atomic acquire - singlethread - global 1. buffer/global/ds/flat_load
3137 - wavefront - local
3138 - generic
3139 load atomic acquire - workgroup - global 1. buffer/global/flat_load
3140 load atomic acquire - workgroup - local 1. ds_load
3141 2. s_waitcnt lgkmcnt(0)
3143 - If OpenCL, omit.
3144 - Must happen before
3145 any following
3146 global/generic
3147 load/load
3148 atomic/store/store
3149 atomic/atomicrmw.
3150 - Ensures any
3151 following global
3152 data read is no
3153 older than the load
3154 atomic value being
3155 acquired.
3156 load atomic acquire - workgroup - generic 1. flat_load
3157 2. s_waitcnt lgkmcnt(0)
3159 - If OpenCL, omit.
3160 - Must happen before
3161 any following
3162 global/generic
3163 load/load
3164 atomic/store/store
3165 atomic/atomicrmw.
3166 - Ensures any
3167 following global
3168 data read is no
3169 older than the load
3170 atomic value being
3171 acquired.
3172 load atomic acquire - agent - global 1. buffer/global/flat_load
3173 - system glc=1
3174 2. s_waitcnt vmcnt(0)
3176 - Must happen before
3177 following
3178 buffer_wbinvl1_vol.
3179 - Ensures the load
3180 has completed
3181 before invalidating
3182 the cache.
3184 3. buffer_wbinvl1_vol
3186 - Must happen before
3187 any following
3188 global/generic
3189 load/load
3190 atomic/atomicrmw.
3191 - Ensures that
3192 following
3193 loads will not see
3194 stale global data.
3196 load atomic acquire - agent - generic 1. flat_load glc=1
3197 - system 2. s_waitcnt vmcnt(0) &
3198 lgkmcnt(0)
3200 - If OpenCL omit
3201 lgkmcnt(0).
3202 - Must happen before
3203 following
3204 buffer_wbinvl1_vol.
3205 - Ensures the flat_load
3206 has completed
3207 before invalidating
3208 the cache.
3210 3. buffer_wbinvl1_vol
3212 - Must happen before
3213 any following
3214 global/generic
3215 load/load
3216 atomic/atomicrmw.
3217 - Ensures that
3218 following loads
3219 will not see stale
3220 global data.
3222 atomicrmw acquire - singlethread - global 1. buffer/global/ds/flat_atomic
3223 - wavefront - local
3224 - generic
3225 atomicrmw acquire - workgroup - global 1. buffer/global/flat_atomic
3226 atomicrmw acquire - workgroup - local 1. ds_atomic
3227 2. waitcnt lgkmcnt(0)
3229 - If OpenCL, omit.
3230 - Must happen before
3231 any following
3232 global/generic
3233 load/load
3234 atomic/store/store
3235 atomic/atomicrmw.
3236 - Ensures any
3237 following global
3238 data read is no
3239 older than the
3240 atomicrmw value
3241 being acquired.
3243 atomicrmw acquire - workgroup - generic 1. flat_atomic
3244 2. waitcnt lgkmcnt(0)
3246 - If OpenCL, omit.
3247 - Must happen before
3248 any following
3249 global/generic
3250 load/load
3251 atomic/store/store
3252 atomic/atomicrmw.
3253 - Ensures any
3254 following global
3255 data read is no
3256 older than the
3257 atomicrmw value
3258 being acquired.
3260 atomicrmw acquire - agent - global 1. buffer/global/flat_atomic
3261 - system 2. s_waitcnt vmcnt(0)
3263 - Must happen before
3264 following
3265 buffer_wbinvl1_vol.
3266 - Ensures the
3267 atomicrmw has
3268 completed before
3269 invalidating the
3270 cache.
3272 3. buffer_wbinvl1_vol
3274 - Must happen before
3275 any following
3276 global/generic
3277 load/load
3278 atomic/atomicrmw.
3279 - Ensures that
3280 following loads
3281 will not see stale
3282 global data.
3284 atomicrmw acquire - agent - generic 1. flat_atomic
3285 - system 2. s_waitcnt vmcnt(0) &
3286 lgkmcnt(0)
3288 - If OpenCL, omit
3289 lgkmcnt(0).
3290 - Must happen before
3291 following
3292 buffer_wbinvl1_vol.
3293 - Ensures the
3294 atomicrmw has
3295 completed before
3296 invalidating the
3297 cache.
3299 3. buffer_wbinvl1_vol
3301 - Must happen before
3302 any following
3303 global/generic
3304 load/load
3305 atomic/atomicrmw.
3306 - Ensures that
3307 following loads
3308 will not see stale
3309 global data.
3311 fence acquire - singlethread *none* *none*
3312 - wavefront
3313 fence acquire - workgroup *none* 1. s_waitcnt lgkmcnt(0)
3315 - If OpenCL and
3316 address space is
3317 not generic, omit.
3318 - However, since LLVM
3319 currently has no
3320 address space on
3321 the fence need to
3322 conservatively
3323 always generate. If
3324 fence had an
3325 address space then
3326 set to address
3327 space of OpenCL
3328 fence flag, or to
3329 generic if both
3330 local and global
3331 flags are
3332 specified.
3333 - Must happen after
3334 any preceding
3335 local/generic load
3336 atomic/atomicrmw
3337 with an equal or
3338 wider sync scope
3339 and memory ordering
3340 stronger than
3341 unordered (this is
3342 termed the
3343 fence-paired-atomic).
3344 - Must happen before
3345 any following
3346 global/generic
3347 load/load
3348 atomic/store/store
3349 atomic/atomicrmw.
3350 - Ensures any
3351 following global
3352 data read is no
3353 older than the
3354 value read by the
3355 fence-paired-atomic.
3357 fence acquire - agent *none* 1. s_waitcnt lgkmcnt(0) &
3358 - system vmcnt(0)
3360 - If OpenCL and
3361 address space is
3362 not generic, omit
3363 lgkmcnt(0).
3364 - However, since LLVM
3365 currently has no
3366 address space on
3367 the fence need to
3368 conservatively
3369 always generate
3370 (see comment for
3371 previous fence).
3372 - Could be split into
3373 separate s_waitcnt
3374 vmcnt(0) and
3375 s_waitcnt
3376 lgkmcnt(0) to allow
3377 them to be
3378 independently moved
3379 according to the
3380 following rules.
3381 - s_waitcnt vmcnt(0)
3382 must happen after
3383 any preceding
3384 global/generic load
3385 atomic/atomicrmw
3386 with an equal or
3387 wider sync scope
3388 and memory ordering
3389 stronger than
3390 unordered (this is
3391 termed the
3392 fence-paired-atomic).
3393 - s_waitcnt lgkmcnt(0)
3394 must happen after
3395 any preceding
3396 local/generic load
3397 atomic/atomicrmw
3398 with an equal or
3399 wider sync scope
3400 and memory ordering
3401 stronger than
3402 unordered (this is
3403 termed the
3404 fence-paired-atomic).
3405 - Must happen before
3406 the following
3407 buffer_wbinvl1_vol.
3408 - Ensures that the
3409 fence-paired atomic
3410 has completed
3411 before invalidating
3412 the
3413 cache. Therefore
3414 any following
3415 locations read must
3416 be no older than
3417 the value read by
3418 the
3419 fence-paired-atomic.
3421 2. buffer_wbinvl1_vol
3423 - Must happen before any
3424 following global/generic
3425 load/load
3426 atomic/store/store
3427 atomic/atomicrmw.
3428 - Ensures that
3429 following loads
3430 will not see stale
3431 global data.
3433 **Release Atomic**
3434 -----------------------------------------------------------------------------------
3435 store atomic release - singlethread - global 1. buffer/global/ds/flat_store
3436 - wavefront - local
3437 - generic
3438 store atomic release - workgroup - global 1. s_waitcnt lgkmcnt(0)
3440 - If OpenCL, omit.
3441 - Must happen after
3442 any preceding
3443 local/generic
3444 load/store/load
3445 atomic/store
3446 atomic/atomicrmw.
3447 - Must happen before
3448 the following
3449 store.
3450 - Ensures that all
3451 memory operations
3452 to local have
3453 completed before
3454 performing the
3455 store that is being
3456 released.
3458 2. buffer/global/flat_store
3459 store atomic release - workgroup - local 1. ds_store
3460 store atomic release - workgroup - generic 1. s_waitcnt lgkmcnt(0)
3462 - If OpenCL, omit.
3463 - Must happen after
3464 any preceding
3465 local/generic
3466 load/store/load
3467 atomic/store
3468 atomic/atomicrmw.
3469 - Must happen before
3470 the following
3471 store.
3472 - Ensures that all
3473 memory operations
3474 to local have
3475 completed before
3476 performing the
3477 store that is being
3478 released.
3480 2. flat_store
3481 store atomic release - agent - global 1. s_waitcnt lgkmcnt(0) &
3482 - system - generic vmcnt(0)
3484 - If OpenCL, omit
3485 lgkmcnt(0).
3486 - Could be split into
3487 separate s_waitcnt
3488 vmcnt(0) and
3489 s_waitcnt
3490 lgkmcnt(0) to allow
3491 them to be
3492 independently moved
3493 according to the
3494 following rules.
3495 - s_waitcnt vmcnt(0)
3496 must happen after
3497 any preceding
3498 global/generic
3499 load/store/load
3500 atomic/store
3501 atomic/atomicrmw.
3502 - s_waitcnt lgkmcnt(0)
3503 must happen after
3504 any preceding
3505 local/generic
3506 load/store/load
3507 atomic/store
3508 atomic/atomicrmw.
3509 - Must happen before
3510 the following
3511 store.
3512 - Ensures that all
3513 memory operations
3514 to memory have
3515 completed before
3516 performing the
3517 store that is being
3518 released.
3520 2. buffer/global/ds/flat_store
3521 atomicrmw release - singlethread - global 1. buffer/global/ds/flat_atomic
3522 - wavefront - local
3523 - generic
3524 atomicrmw release - workgroup - global 1. s_waitcnt lgkmcnt(0)
3526 - If OpenCL, omit.
3527 - Must happen after
3528 any preceding
3529 local/generic
3530 load/store/load
3531 atomic/store
3532 atomic/atomicrmw.
3533 - Must happen before
3534 the following
3535 atomicrmw.
3536 - Ensures that all
3537 memory operations
3538 to local have
3539 completed before
3540 performing the
3541 atomicrmw that is
3542 being released.
3544 2. buffer/global/flat_atomic
3545 atomicrmw release - workgroup - local 1. ds_atomic
3546 atomicrmw release - workgroup - generic 1. s_waitcnt lgkmcnt(0)
3548 - If OpenCL, omit.
3549 - Must happen after
3550 any preceding
3551 local/generic
3552 load/store/load
3553 atomic/store
3554 atomic/atomicrmw.
3555 - Must happen before
3556 the following
3557 atomicrmw.
3558 - Ensures that all
3559 memory operations
3560 to local have
3561 completed before
3562 performing the
3563 atomicrmw that is
3564 being released.
3566 2. flat_atomic
3567 atomicrmw release - agent - global 1. s_waitcnt lgkmcnt(0) &
3568 - system - generic vmcnt(0)
3570 - If OpenCL, omit
3571 lgkmcnt(0).
3572 - Could be split into
3573 separate s_waitcnt
3574 vmcnt(0) and
3575 s_waitcnt
3576 lgkmcnt(0) to allow
3577 them to be
3578 independently moved
3579 according to the
3580 following rules.
3581 - s_waitcnt vmcnt(0)
3582 must happen after
3583 any preceding
3584 global/generic
3585 load/store/load
3586 atomic/store
3587 atomic/atomicrmw.
3588 - s_waitcnt lgkmcnt(0)
3589 must happen after
3590 any preceding
3591 local/generic
3592 load/store/load
3593 atomic/store
3594 atomic/atomicrmw.
3595 - Must happen before
3596 the following
3597 atomicrmw.
3598 - Ensures that all
3599 memory operations
3600 to global and local
3601 have completed
3602 before performing
3603 the atomicrmw that
3604 is being released.
3606 2. buffer/global/ds/flat_atomic
3607 fence release - singlethread *none* *none*
3608 - wavefront
3609 fence release - workgroup *none* 1. s_waitcnt lgkmcnt(0)
3611 - If OpenCL and
3612 address space is
3613 not generic, omit.
3614 - However, since LLVM
3615 currently has no
3616 address space on
3617 the fence need to
3618 conservatively
3619 always generate. If
3620 fence had an
3621 address space then
3622 set to address
3623 space of OpenCL
3624 fence flag, or to
3625 generic if both
3626 local and global
3627 flags are
3628 specified.
3629 - Must happen after
3630 any preceding
3631 local/generic
3632 load/load
3633 atomic/store/store
3634 atomic/atomicrmw.
3635 - Must happen before
3636 any following store
3637 atomic/atomicrmw
3638 with an equal or
3639 wider sync scope
3640 and memory ordering
3641 stronger than
3642 unordered (this is
3643 termed the
3644 fence-paired-atomic).
3645 - Ensures that all
3646 memory operations
3647 to local have
3648 completed before
3649 performing the
3650 following
3651 fence-paired-atomic.
3653 fence release - agent *none* 1. s_waitcnt lgkmcnt(0) &
3654 - system vmcnt(0)
3656 - If OpenCL and
3657 address space is
3658 not generic, omit
3659 lgkmcnt(0).
3660 - If OpenCL and
3661 address space is
3662 local, omit
3663 vmcnt(0).
3664 - However, since LLVM
3665 currently has no
3666 address space on
3667 the fence need to
3668 conservatively
3669 always generate. If
3670 fence had an
3671 address space then
3672 set to address
3673 space of OpenCL
3674 fence flag, or to
3675 generic if both
3676 local and global
3677 flags are
3678 specified.
3679 - Could be split into
3680 separate s_waitcnt
3681 vmcnt(0) and
3682 s_waitcnt
3683 lgkmcnt(0) to allow
3684 them to be
3685 independently moved
3686 according to the
3687 following rules.
3688 - s_waitcnt vmcnt(0)
3689 must happen after
3690 any preceding
3691 global/generic
3692 load/store/load
3693 atomic/store
3694 atomic/atomicrmw.
3695 - s_waitcnt lgkmcnt(0)
3696 must happen after
3697 any preceding
3698 local/generic
3699 load/store/load
3700 atomic/store
3701 atomic/atomicrmw.
3702 - Must happen before
3703 any following store
3704 atomic/atomicrmw
3705 with an equal or
3706 wider sync scope
3707 and memory ordering
3708 stronger than
3709 unordered (this is
3710 termed the
3711 fence-paired-atomic).
3712 - Ensures that all
3713 memory operations
3714 have
3715 completed before
3716 performing the
3717 following
3718 fence-paired-atomic.
3720 **Acquire-Release Atomic**
3721 -----------------------------------------------------------------------------------
3722 atomicrmw acq_rel - singlethread - global 1. buffer/global/ds/flat_atomic
3723 - wavefront - local
3724 - generic
3725 atomicrmw acq_rel - workgroup - global 1. s_waitcnt lgkmcnt(0)
3727 - If OpenCL, omit.
3728 - Must happen after
3729 any preceding
3730 local/generic
3731 load/store/load
3732 atomic/store
3733 atomic/atomicrmw.
3734 - Must happen before
3735 the following
3736 atomicrmw.
3737 - Ensures that all
3738 memory operations
3739 to local have
3740 completed before
3741 performing the
3742 atomicrmw that is
3743 being released.
3745 2. buffer/global/flat_atomic
3746 atomicrmw acq_rel - workgroup - local 1. ds_atomic
3747 2. s_waitcnt lgkmcnt(0)
3749 - If OpenCL, omit.
3750 - Must happen before
3751 any following
3752 global/generic
3753 load/load
3754 atomic/store/store
3755 atomic/atomicrmw.
3756 - Ensures any
3757 following global
3758 data read is no
3759 older than the load
3760 atomic value being
3761 acquired.
3763 atomicrmw acq_rel - workgroup - generic 1. s_waitcnt lgkmcnt(0)
3765 - If OpenCL, omit.
3766 - Must happen after
3767 any preceding
3768 local/generic
3769 load/store/load
3770 atomic/store
3771 atomic/atomicrmw.
3772 - Must happen before
3773 the following
3774 atomicrmw.
3775 - Ensures that all
3776 memory operations
3777 to local have
3778 completed before
3779 performing the
3780 atomicrmw that is
3781 being released.
3783 2. flat_atomic
3784 3. s_waitcnt lgkmcnt(0)
3786 - If OpenCL, omit.
3787 - Must happen before
3788 any following
3789 global/generic
3790 load/load
3791 atomic/store/store
3792 atomic/atomicrmw.
3793 - Ensures any
3794 following global
3795 data read is no
3796 older than the load
3797 atomic value being
3798 acquired.
3800 atomicrmw acq_rel - agent - global 1. s_waitcnt lgkmcnt(0) &
3801 - system vmcnt(0)
3803 - If OpenCL, omit
3804 lgkmcnt(0).
3805 - Could be split into
3806 separate s_waitcnt
3807 vmcnt(0) and
3808 s_waitcnt
3809 lgkmcnt(0) to allow
3810 them to be
3811 independently moved
3812 according to the
3813 following rules.
3814 - s_waitcnt vmcnt(0)
3815 must happen after
3816 any preceding
3817 global/generic
3818 load/store/load
3819 atomic/store
3820 atomic/atomicrmw.
3821 - s_waitcnt lgkmcnt(0)
3822 must happen after
3823 any preceding
3824 local/generic
3825 load/store/load
3826 atomic/store
3827 atomic/atomicrmw.
3828 - Must happen before
3829 the following
3830 atomicrmw.
3831 - Ensures that all
3832 memory operations
3833 to global have
3834 completed before
3835 performing the
3836 atomicrmw that is
3837 being released.
3839 2. buffer/global/flat_atomic
3840 3. s_waitcnt vmcnt(0)
3842 - Must happen before
3843 following
3844 buffer_wbinvl1_vol.
3845 - Ensures the
3846 atomicrmw has
3847 completed before
3848 invalidating the
3849 cache.
3851 4. buffer_wbinvl1_vol
3853 - Must happen before
3854 any following
3855 global/generic
3856 load/load
3857 atomic/atomicrmw.
3858 - Ensures that
3859 following loads
3860 will not see stale
3861 global data.
3863 atomicrmw acq_rel - agent - generic 1. s_waitcnt lgkmcnt(0) &
3864 - system vmcnt(0)
3866 - If OpenCL, omit
3867 lgkmcnt(0).
3868 - Could be split into
3869 separate s_waitcnt
3870 vmcnt(0) and
3871 s_waitcnt
3872 lgkmcnt(0) to allow
3873 them to be
3874 independently moved
3875 according to the
3876 following rules.
3877 - s_waitcnt vmcnt(0)
3878 must happen after
3879 any preceding
3880 global/generic
3881 load/store/load
3882 atomic/store
3883 atomic/atomicrmw.
3884 - s_waitcnt lgkmcnt(0)
3885 must happen after
3886 any preceding
3887 local/generic
3888 load/store/load
3889 atomic/store
3890 atomic/atomicrmw.
3891 - Must happen before
3892 the following
3893 atomicrmw.
3894 - Ensures that all
3895 memory operations
3896 to global have
3897 completed before
3898 performing the
3899 atomicrmw that is
3900 being released.
3902 2. flat_atomic
3903 3. s_waitcnt vmcnt(0) &
3904 lgkmcnt(0)
3906 - If OpenCL, omit
3907 lgkmcnt(0).
3908 - Must happen before
3909 following
3910 buffer_wbinvl1_vol.
3911 - Ensures the
3912 atomicrmw has
3913 completed before
3914 invalidating the
3915 cache.
3917 4. buffer_wbinvl1_vol
3919 - Must happen before
3920 any following
3921 global/generic
3922 load/load
3923 atomic/atomicrmw.
3924 - Ensures that
3925 following loads
3926 will not see stale
3927 global data.
3929 fence acq_rel - singlethread *none* *none*
3930 - wavefront
3931 fence acq_rel - workgroup *none* 1. s_waitcnt lgkmcnt(0)
3933 - If OpenCL and
3934 address space is
3935 not generic, omit.
3936 - However,
3937 since LLVM
3938 currently has no
3939 address space on
3940 the fence need to
3941 conservatively
3942 always generate
3943 (see comment for
3944 previous fence).
3945 - Must happen after
3946 any preceding
3947 local/generic
3948 load/load
3949 atomic/store/store
3950 atomic/atomicrmw.
3951 - Must happen before
3952 any following
3953 global/generic
3954 load/load
3955 atomic/store/store
3956 atomic/atomicrmw.
3957 - Ensures that all
3958 memory operations
3959 to local have
3960 completed before
3961 performing any
3962 following global
3963 memory operations.
3964 - Ensures that the
3965 preceding
3966 local/generic load
3967 atomic/atomicrmw
3968 with an equal or
3969 wider sync scope
3970 and memory ordering
3971 stronger than
3972 unordered (this is
3973 termed the
3974 acquire-fence-paired-atomic
3975 ) has completed
3976 before following
3977 global memory
3978 operations. This
3979 satisfies the
3980 requirements of
3981 acquire.
3982 - Ensures that all
3983 previous memory
3984 operations have
3985 completed before a
3986 following
3987 local/generic store
3988 atomic/atomicrmw
3989 with an equal or
3990 wider sync scope
3991 and memory ordering
3992 stronger than
3993 unordered (this is
3994 termed the
3995 release-fence-paired-atomic
3996 ). This satisfies the
3997 requirements of
3998 release.
4000 fence acq_rel - agent *none* 1. s_waitcnt lgkmcnt(0) &
4001 - system vmcnt(0)
4003 - If OpenCL and
4004 address space is
4005 not generic, omit
4006 lgkmcnt(0).
4007 - However, since LLVM
4008 currently has no
4009 address space on
4010 the fence need to
4011 conservatively
4012 always generate
4013 (see comment for
4014 previous fence).
4015 - Could be split into
4016 separate s_waitcnt
4017 vmcnt(0) and
4018 s_waitcnt
4019 lgkmcnt(0) to allow
4020 them to be
4021 independently moved
4022 according to the
4023 following rules.
4024 - s_waitcnt vmcnt(0)
4025 must happen after
4026 any preceding
4027 global/generic
4028 load/store/load
4029 atomic/store
4030 atomic/atomicrmw.
4031 - s_waitcnt lgkmcnt(0)
4032 must happen after
4033 any preceding
4034 local/generic
4035 load/store/load
4036 atomic/store
4037 atomic/atomicrmw.
4038 - Must happen before
4039 the following
4040 buffer_wbinvl1_vol.
4041 - Ensures that the
4042 preceding
4043 global/local/generic
4044 load
4045 atomic/atomicrmw
4046 with an equal or
4047 wider sync scope
4048 and memory ordering
4049 stronger than
4050 unordered (this is
4051 termed the
4052 acquire-fence-paired-atomic
4053 ) has completed
4054 before invalidating
4055 the cache. This
4056 satisfies the
4057 requirements of
4058 acquire.
4059 - Ensures that all
4060 previous memory
4061 operations have
4062 completed before a
4063 following
4064 global/local/generic
4065 store
4066 atomic/atomicrmw
4067 with an equal or
4068 wider sync scope
4069 and memory ordering
4070 stronger than
4071 unordered (this is
4072 termed the
4073 release-fence-paired-atomic
4074 ). This satisfies the
4075 requirements of
4076 release.
4078 2. buffer_wbinvl1_vol
4080 - Must happen before
4081 any following
4082 global/generic
4083 load/load
4084 atomic/store/store
4085 atomic/atomicrmw.
4086 - Ensures that
4087 following loads
4088 will not see stale
4089 global data. This
4090 satisfies the
4091 requirements of
4092 acquire.
4094 **Sequential Consistent Atomic**
4095 -----------------------------------------------------------------------------------
4096 load atomic seq_cst - singlethread - global *Same as corresponding
4097 - wavefront - local load atomic acquire,
4098 - generic except must generated
4099 all instructions even
4100 for OpenCL.*
4101 load atomic seq_cst - workgroup - global 1. s_waitcnt lgkmcnt(0)
4102 - generic
4103 - Must
4104 happen after
4105 preceding
4106 global/generic load
4107 atomic/store
4108 atomic/atomicrmw
4109 with memory
4110 ordering of seq_cst
4111 and with equal or
4112 wider sync scope.
4113 (Note that seq_cst
4114 fences have their
4115 own s_waitcnt
4116 lgkmcnt(0) and so do
4117 not need to be
4118 considered.)
4119 - Ensures any
4120 preceding
4121 sequential
4122 consistent local
4123 memory instructions
4124 have completed
4125 before executing
4126 this sequentially
4127 consistent
4128 instruction. This
4129 prevents reordering
4130 a seq_cst store
4131 followed by a
4132 seq_cst load. (Note
4133 that seq_cst is
4134 stronger than
4135 acquire/release as
4136 the reordering of
4137 load acquire
4138 followed by a store
4139 release is
4140 prevented by the
4141 waitcnt of
4142 the release, but
4143 there is nothing
4144 preventing a store
4145 release followed by
4146 load acquire from
4147 competing out of
4148 order.)
4150 2. *Following
4151 instructions same as
4152 corresponding load
4153 atomic acquire,
4154 except must generated
4155 all instructions even
4156 for OpenCL.*
4157 load atomic seq_cst - workgroup - local *Same as corresponding
4158 load atomic acquire,
4159 except must generated
4160 all instructions even
4161 for OpenCL.*
4162 load atomic seq_cst - agent - global 1. s_waitcnt lgkmcnt(0) &
4163 - system - generic vmcnt(0)
4165 - Could be split into
4166 separate s_waitcnt
4167 vmcnt(0)
4168 and s_waitcnt
4169 lgkmcnt(0) to allow
4170 them to be
4171 independently moved
4172 according to the
4173 following rules.
4174 - waitcnt lgkmcnt(0)
4175 must happen after
4176 preceding
4177 global/generic load
4178 atomic/store
4179 atomic/atomicrmw
4180 with memory
4181 ordering of seq_cst
4182 and with equal or
4183 wider sync scope.
4184 (Note that seq_cst
4185 fences have their
4186 own s_waitcnt
4187 lgkmcnt(0) and so do
4188 not need to be
4189 considered.)
4190 - waitcnt vmcnt(0)
4191 must happen after
4192 preceding
4193 global/generic load
4194 atomic/store
4195 atomic/atomicrmw
4196 with memory
4197 ordering of seq_cst
4198 and with equal or
4199 wider sync scope.
4200 (Note that seq_cst
4201 fences have their
4202 own s_waitcnt
4203 vmcnt(0) and so do
4204 not need to be
4205 considered.)
4206 - Ensures any
4207 preceding
4208 sequential
4209 consistent global
4210 memory instructions
4211 have completed
4212 before executing
4213 this sequentially
4214 consistent
4215 instruction. This
4216 prevents reordering
4217 a seq_cst store
4218 followed by a
4219 seq_cst load. (Note
4220 that seq_cst is
4221 stronger than
4222 acquire/release as
4223 the reordering of
4224 load acquire
4225 followed by a store
4226 release is
4227 prevented by the
4228 waitcnt of
4229 the release, but
4230 there is nothing
4231 preventing a store
4232 release followed by
4233 load acquire from
4234 competing out of
4235 order.)
4237 2. *Following
4238 instructions same as
4239 corresponding load
4240 atomic acquire,
4241 except must generated
4242 all instructions even
4243 for OpenCL.*
4244 store atomic seq_cst - singlethread - global *Same as corresponding
4245 - wavefront - local store atomic release,
4246 - workgroup - generic except must generated
4247 all instructions even
4248 for OpenCL.*
4249 store atomic seq_cst - agent - global *Same as corresponding
4250 - system - generic store atomic release,
4251 except must generated
4252 all instructions even
4253 for OpenCL.*
4254 atomicrmw seq_cst - singlethread - global *Same as corresponding
4255 - wavefront - local atomicrmw acq_rel,
4256 - workgroup - generic except must generated
4257 all instructions even
4258 for OpenCL.*
4259 atomicrmw seq_cst - agent - global *Same as corresponding
4260 - system - generic atomicrmw acq_rel,
4261 except must generated
4262 all instructions even
4263 for OpenCL.*
4264 fence seq_cst - singlethread *none* *Same as corresponding
4265 - wavefront fence acq_rel,
4266 - workgroup except must generated
4267 - agent all instructions even
4268 - system for OpenCL.*
4269 ============ ============ ============== ========== ===============================
4271 The memory order also adds the single thread optimization constrains defined in
4272 table
4273 :ref:`amdgpu-amdhsa-memory-model-single-thread-optimization-constraints-gfx6-gfx9-table`.
4275 .. table:: AMDHSA Memory Model Single Thread Optimization Constraints GFX6-GFX9
4276 :name: amdgpu-amdhsa-memory-model-single-thread-optimization-constraints-gfx6-gfx9-table
4278 ============ ==============================================================
4279 LLVM Memory Optimization Constraints
4280 Ordering
4281 ============ ==============================================================
4282 unordered *none*
4283 monotonic *none*
4284 acquire - If a load atomic/atomicrmw then no following load/load
4285 atomic/store/ store atomic/atomicrmw/fence instruction can
4286 be moved before the acquire.
4287 - If a fence then same as load atomic, plus no preceding
4288 associated fence-paired-atomic can be moved after the fence.
4289 release - If a store atomic/atomicrmw then no preceding load/load
4290 atomic/store/ store atomic/atomicrmw/fence instruction can
4291 be moved after the release.
4292 - If a fence then same as store atomic, plus no following
4293 associated fence-paired-atomic can be moved before the
4294 fence.
4295 acq_rel Same constraints as both acquire and release.
4296 seq_cst - If a load atomic then same constraints as acquire, plus no
4297 preceding sequentially consistent load atomic/store
4298 atomic/atomicrmw/fence instruction can be moved after the
4299 seq_cst.
4300 - If a store atomic then the same constraints as release, plus
4301 no following sequentially consistent load atomic/store
4302 atomic/atomicrmw/fence instruction can be moved before the
4303 seq_cst.
4304 - If an atomicrmw/fence then same constraints as acq_rel.
4305 ============ ==============================================================
4307 Trap Handler ABI
4308 ~~~~~~~~~~~~~~~~
4310 For code objects generated by AMDGPU backend for HSA [HSA]_ compatible runtimes
4311 (such as ROCm [AMD-ROCm]_), the runtime installs a trap handler that supports
4312 the ``s_trap`` instruction with the following usage:
4314 .. table:: AMDGPU Trap Handler for AMDHSA OS
4315 :name: amdgpu-trap-handler-for-amdhsa-os-table
4317 =================== =============== =============== =======================
4318 Usage Code Sequence Trap Handler Description
4319 Inputs
4320 =================== =============== =============== =======================
4321 reserved ``s_trap 0x00`` Reserved by hardware.
4322 ``debugtrap(arg)`` ``s_trap 0x01`` ``SGPR0-1``: Reserved for HSA
4323 ``queue_ptr`` ``debugtrap``
4324 ``VGPR0``: intrinsic (not
4325 ``arg`` implemented).
4326 ``llvm.trap`` ``s_trap 0x02`` ``SGPR0-1``: Causes dispatch to be
4327 ``queue_ptr`` terminated and its
4328 associated queue put
4329 into the error state.
4330 ``llvm.debugtrap`` ``s_trap 0x03`` - If debugger not
4331 installed then
4332 behaves as a
4333 no-operation. The
4334 trap handler is
4335 entered and
4336 immediately returns
4337 to continue
4338 execution of the
4339 wavefront.
4340 - If the debugger is
4341 installed, causes
4342 the debug trap to be
4343 reported by the
4344 debugger and the
4345 wavefront is put in
4346 the halt state until
4347 resumed by the
4348 debugger.
4349 reserved ``s_trap 0x04`` Reserved.
4350 reserved ``s_trap 0x05`` Reserved.
4351 reserved ``s_trap 0x06`` Reserved.
4352 debugger breakpoint ``s_trap 0x07`` Reserved for debugger
4353 breakpoints.
4354 reserved ``s_trap 0x08`` Reserved.
4355 reserved ``s_trap 0xfe`` Reserved.
4356 reserved ``s_trap 0xff`` Reserved.
4357 =================== =============== =============== =======================
4359 AMDPAL
4360 ------
4362 This section provides code conventions used when the target triple OS is
4363 ``amdpal`` (see :ref:`amdgpu-target-triples`) for passing runtime parameters
4364 from the application/runtime to each invocation of a hardware shader. These
4365 parameters include both generic, application-controlled parameters called
4366 *user data* as well as system-generated parameters that are a product of the
4367 draw or dispatch execution.
4369 User Data
4370 ~~~~~~~~~
4372 Each hardware stage has a set of 32-bit *user data registers* which can be
4373 written from a command buffer and then loaded into SGPRs when waves are launched
4374 via a subsequent dispatch or draw operation. This is the way most arguments are
4375 passed from the application/runtime to a hardware shader.
4377 Compute User Data
4378 ~~~~~~~~~~~~~~~~~
4380 Compute shader user data mappings are simpler than graphics shaders, and have a
4381 fixed mapping.
4383 Note that there are always 10 available *user data entries* in registers -
4384 entries beyond that limit must be fetched from memory (via the spill table
4385 pointer) by the shader.
4387 .. table:: PAL Compute Shader User Data Registers
4388 :name: pal-compute-user-data-registers
4390 ============= ================================
4391 User Register Description
4392 ============= ================================
4393 0 Global Internal Table (32-bit pointer)
4394 1 Per-Shader Internal Table (32-bit pointer)
4395 2 - 11 Application-Controlled User Data (10 32-bit values)
4396 12 Spill Table (32-bit pointer)
4397 13 - 14 Thread Group Count (64-bit pointer)
4398 15 GDS Range
4399 ============= ================================
4401 Graphics User Data
4402 ~~~~~~~~~~~~~~~~~~
4404 Graphics pipelines support a much more flexible user data mapping:
4406 .. table:: PAL Graphics Shader User Data Registers
4407 :name: pal-graphics-user-data-registers
4409 ============= ================================
4410 User Register Description
4411 ============= ================================
4412 0 Global Internal Table (32-bit pointer)
4413 + Per-Shader Internal Table (32-bit pointer)
4414 + 1-15 Application Controlled User Data
4415 (1-15 Contiguous 32-bit Values in Registers)
4416 + Spill Table (32-bit pointer)
4417 + Draw Index (First Stage Only)
4418 + Vertex Offset (First Stage Only)
4419 + Instance Offset (First Stage Only)
4420 ============= ================================
4422 The placement of the global internal table remains fixed in the first *user
4423 data SGPR register*. Otherwise all parameters are optional, and can be mapped
4424 to any desired *user data SGPR register*, with the following regstrictions:
4426 * Draw Index, Vertex Offset, and Instance Offset can only be used by the first
4427 activehardware stage in a graphics pipeline (i.e. where the API vertex
4428 shader runs).
4430 * Application-controlled user data must be mapped into a contiguous range of
4431 user data registers.
4433 * The application-controlled user data range supports compaction remapping, so
4434 only *entries* that are actually consumed by the shader must be assigned to
4435 corresponding *registers*. Note that in order to support an efficient runtime
4436 implementation, the remapping must pack *registers* in the same order as
4437 *entries*, with unused *entries* removed.
4439 .. _pal_global_internal_table:
4441 Global Internal Table
4442 ~~~~~~~~~~~~~~~~~~~~~
4444 The global internal table is a table of *shader resource descriptors* (SRDs) that
4445 define how certain engine-wide, runtime-managed resources should be accessed
4446 from a shader. The majority of these resources have HW-defined formats, and it
4447 is up to the compiler to write/read data as required by the target hardware.
4449 The following table illustrates the required format:
4451 .. table:: PAL Global Internal Table
4452 :name: pal-git-table
4454 ============= ================================
4455 Offset Description
4456 ============= ================================
4457 0-3 Graphics Scratch SRD
4458 4-7 Compute Scratch SRD
4459 8-11 ES/GS Ring Output SRD
4460 12-15 ES/GS Ring Input SRD
4461 16-19 GS/VS Ring Output #0
4462 20-23 GS/VS Ring Output #1
4463 24-27 GS/VS Ring Output #2
4464 28-31 GS/VS Ring Output #3
4465 32-35 GS/VS Ring Input SRD
4466 36-39 Tessellation Factor Buffer SRD
4467 40-43 Off-Chip LDS Buffer SRD
4468 44-47 Off-Chip Param Cache Buffer SRD
4469 48-51 Sample Position Buffer SRD
4470 52 vaRange::ShadowDescriptorTable High Bits
4471 ============= ================================
4473 The pointer to the global internal table passed to the shader as user data
4474 is a 32-bit pointer. The top 32 bits should be assumed to be the same as
4475 the top 32 bits of the pipeline, so the shader may use the program
4476 counter's top 32 bits.
4478 Unspecified OS
4479 --------------
4481 This section provides code conventions used when the target triple OS is
4482 empty (see :ref:`amdgpu-target-triples`).
4484 Trap Handler ABI
4485 ~~~~~~~~~~~~~~~~
4487 For code objects generated by AMDGPU backend for non-amdhsa OS, the runtime does
4488 not install a trap handler. The ``llvm.trap`` and ``llvm.debugtrap``
4489 instructions are handled as follows:
4491 .. table:: AMDGPU Trap Handler for Non-AMDHSA OS
4492 :name: amdgpu-trap-handler-for-non-amdhsa-os-table
4494 =============== =============== ===========================================
4495 Usage Code Sequence Description
4496 =============== =============== ===========================================
4497 llvm.trap s_endpgm Causes wavefront to be terminated.
4498 llvm.debugtrap *none* Compiler warning given that there is no
4499 trap handler installed.
4500 =============== =============== ===========================================
4502 Source Languages
4503 ================
4505 .. _amdgpu-opencl:
4507 OpenCL
4508 ------
4510 When the language is OpenCL the following differences occur:
4512 1. The OpenCL memory model is used (see :ref:`amdgpu-amdhsa-memory-model`).
4513 2. The AMDGPU backend appends additional arguments to the kernel's explicit
4514 arguments for the AMDHSA OS (see
4515 :ref:`opencl-kernel-implicit-arguments-appended-for-amdhsa-os-table`).
4516 3. Additional metadata is generated
4517 (see :ref:`amdgpu-amdhsa-code-object-metadata`).
4519 .. table:: OpenCL kernel implicit arguments appended for AMDHSA OS
4520 :name: opencl-kernel-implicit-arguments-appended-for-amdhsa-os-table
4522 ======== ==== ========= ===========================================
4523 Position Byte Byte Description
4524 Size Alignment
4525 ======== ==== ========= ===========================================
4526 1 8 8 OpenCL Global Offset X
4527 2 8 8 OpenCL Global Offset Y
4528 3 8 8 OpenCL Global Offset Z
4529 4 8 8 OpenCL address of printf buffer
4530 5 8 8 OpenCL address of virtual queue used by
4531 enqueue_kernel.
4532 6 8 8 OpenCL address of AqlWrap struct used by
4533 enqueue_kernel.
4534 ======== ==== ========= ===========================================
4536 .. _amdgpu-hcc:
4538 HCC
4539 ---
4541 When the language is HCC the following differences occur:
4543 1. The HSA memory model is used (see :ref:`amdgpu-amdhsa-memory-model`).
4545 .. _amdgpu-assembler:
4547 Assembler
4548 ---------
4550 AMDGPU backend has LLVM-MC based assembler which is currently in development.
4551 It supports AMDGCN GFX6-GFX9.
4553 This section describes general syntax for instructions and operands.
4555 Instructions
4556 ~~~~~~~~~~~~
4558 .. toctree::
4559 :hidden:
4561 AMDGPU/AMDGPUAsmGFX7
4562 AMDGPU/AMDGPUAsmGFX8
4563 AMDGPU/AMDGPUAsmGFX9
4564 AMDGPUModifierSyntax
4565 AMDGPUOperandSyntax
4566 AMDGPUInstructionSyntax
4567 AMDGPUInstructionNotation
4569 An instruction has the following :doc:`syntax<AMDGPUInstructionSyntax>`:
4571 ``<``\ *opcode*\ ``> <``\ *operand0*\ ``>, <``\ *operand1*\ ``>,... <``\ *modifier0*\ ``> <``\ *modifier1*\ ``>...``
4573 :doc:`Operands<AMDGPUOperandSyntax>` are normally comma-separated while
4574 :doc:`modifiers<AMDGPUModifierSyntax>` are space-separated.
4576 The order of *operands* and *modifiers* is fixed.
4577 Most *modifiers* are optional and may be omitted.
4579 See detailed instruction syntax description for :doc:`GFX7<AMDGPU/AMDGPUAsmGFX7>`,
4580 :doc:`GFX8<AMDGPU/AMDGPUAsmGFX8>` and :doc:`GFX9<AMDGPU/AMDGPUAsmGFX9>`.
4582 Note that features under development are not included in this description.
4584 For more information about instructions, their semantics and supported combinations of
4585 operands, refer to one of instruction set architecture manuals
4586 [AMD-GCN-GFX6]_, [AMD-GCN-GFX7]_, [AMD-GCN-GFX8]_ and [AMD-GCN-GFX9]_.
4588 Operands
4589 ~~~~~~~~
4591 Detailed description of operands may be found :doc:`here<AMDGPUOperandSyntax>`.
4593 Modifiers
4594 ~~~~~~~~~
4596 Detailed description of modifiers may be found :doc:`here<AMDGPUModifierSyntax>`.
4598 Instruction Examples
4599 ~~~~~~~~~~~~~~~~~~~~
4601 DS
4602 ++
4604 .. code-block:: nasm
4606 ds_add_u32 v2, v4 offset:16
4607 ds_write_src2_b64 v2 offset0:4 offset1:8
4608 ds_cmpst_f32 v2, v4, v6
4609 ds_min_rtn_f64 v[8:9], v2, v[4:5]
4612 For full list of supported instructions, refer to "LDS/GDS instructions" in ISA Manual.
4614 FLAT
4615 ++++
4617 .. code-block:: nasm
4619 flat_load_dword v1, v[3:4]
4620 flat_store_dwordx3 v[3:4], v[5:7]
4621 flat_atomic_swap v1, v[3:4], v5 glc
4622 flat_atomic_cmpswap v1, v[3:4], v[5:6] glc slc
4623 flat_atomic_fmax_x2 v[1:2], v[3:4], v[5:6] glc
4625 For full list of supported instructions, refer to "FLAT instructions" in ISA Manual.
4627 MUBUF
4628 +++++
4630 .. code-block:: nasm
4632 buffer_load_dword v1, off, s[4:7], s1
4633 buffer_store_dwordx4 v[1:4], v2, ttmp[4:7], s1 offen offset:4 glc tfe
4634 buffer_store_format_xy v[1:2], off, s[4:7], s1
4635 buffer_wbinvl1
4636 buffer_atomic_inc v1, v2, s[8:11], s4 idxen offset:4 slc
4638 For full list of supported instructions, refer to "MUBUF Instructions" in ISA Manual.
4640 SMRD/SMEM
4641 +++++++++
4643 .. code-block:: nasm
4645 s_load_dword s1, s[2:3], 0xfc
4646 s_load_dwordx8 s[8:15], s[2:3], s4
4647 s_load_dwordx16 s[88:103], s[2:3], s4
4648 s_dcache_inv_vol
4649 s_memtime s[4:5]
4651 For full list of supported instructions, refer to "Scalar Memory Operations" in ISA Manual.
4653 SOP1
4654 ++++
4656 .. code-block:: nasm
4658 s_mov_b32 s1, s2
4659 s_mov_b64 s[0:1], 0x80000000
4660 s_cmov_b32 s1, 200
4661 s_wqm_b64 s[2:3], s[4:5]
4662 s_bcnt0_i32_b64 s1, s[2:3]
4663 s_swappc_b64 s[2:3], s[4:5]
4664 s_cbranch_join s[4:5]
4666 For full list of supported instructions, refer to "SOP1 Instructions" in ISA Manual.
4668 SOP2
4669 ++++
4671 .. code-block:: nasm
4673 s_add_u32 s1, s2, s3
4674 s_and_b64 s[2:3], s[4:5], s[6:7]
4675 s_cselect_b32 s1, s2, s3
4676 s_andn2_b32 s2, s4, s6
4677 s_lshr_b64 s[2:3], s[4:5], s6
4678 s_ashr_i32 s2, s4, s6
4679 s_bfm_b64 s[2:3], s4, s6
4680 s_bfe_i64 s[2:3], s[4:5], s6
4681 s_cbranch_g_fork s[4:5], s[6:7]
4683 For full list of supported instructions, refer to "SOP2 Instructions" in ISA Manual.
4685 SOPC
4686 ++++
4688 .. code-block:: nasm
4690 s_cmp_eq_i32 s1, s2
4691 s_bitcmp1_b32 s1, s2
4692 s_bitcmp0_b64 s[2:3], s4
4693 s_setvskip s3, s5
4695 For full list of supported instructions, refer to "SOPC Instructions" in ISA Manual.
4697 SOPP
4698 ++++
4700 .. code-block:: nasm
4702 s_barrier
4703 s_nop 2
4704 s_endpgm
4705 s_waitcnt 0 ; Wait for all counters to be 0
4706 s_waitcnt vmcnt(0) & expcnt(0) & lgkmcnt(0) ; Equivalent to above
4707 s_waitcnt vmcnt(1) ; Wait for vmcnt counter to be 1.
4708 s_sethalt 9
4709 s_sleep 10
4710 s_sendmsg 0x1
4711 s_sendmsg sendmsg(MSG_INTERRUPT)
4712 s_trap 1
4714 For full list of supported instructions, refer to "SOPP Instructions" in ISA Manual.
4716 Unless otherwise mentioned, little verification is performed on the operands
4717 of SOPP Instructions, so it is up to the programmer to be familiar with the
4718 range or acceptable values.
4720 VALU
4721 ++++
4723 For vector ALU instruction opcodes (VOP1, VOP2, VOP3, VOPC, VOP_DPP, VOP_SDWA),
4724 the assembler will automatically use optimal encoding based on its operands.
4725 To force specific encoding, one can add a suffix to the opcode of the instruction:
4727 * _e32 for 32-bit VOP1/VOP2/VOPC
4728 * _e64 for 64-bit VOP3
4729 * _dpp for VOP_DPP
4730 * _sdwa for VOP_SDWA
4732 VOP1/VOP2/VOP3/VOPC examples:
4734 .. code-block:: nasm
4736 v_mov_b32 v1, v2
4737 v_mov_b32_e32 v1, v2
4738 v_nop
4739 v_cvt_f64_i32_e32 v[1:2], v2
4740 v_floor_f32_e32 v1, v2
4741 v_bfrev_b32_e32 v1, v2
4742 v_add_f32_e32 v1, v2, v3
4743 v_mul_i32_i24_e64 v1, v2, 3
4744 v_mul_i32_i24_e32 v1, -3, v3
4745 v_mul_i32_i24_e32 v1, -100, v3
4746 v_addc_u32 v1, s[0:1], v2, v3, s[2:3]
4747 v_max_f16_e32 v1, v2, v3
4749 VOP_DPP examples:
4751 .. code-block:: nasm
4753 v_mov_b32 v0, v0 quad_perm:[0,2,1,1]
4754 v_sin_f32 v0, v0 row_shl:1 row_mask:0xa bank_mask:0x1 bound_ctrl:0
4755 v_mov_b32 v0, v0 wave_shl:1
4756 v_mov_b32 v0, v0 row_mirror
4757 v_mov_b32 v0, v0 row_bcast:31
4758 v_mov_b32 v0, v0 quad_perm:[1,3,0,1] row_mask:0xa bank_mask:0x1 bound_ctrl:0
4759 v_add_f32 v0, v0, |v0| row_shl:1 row_mask:0xa bank_mask:0x1 bound_ctrl:0
4760 v_max_f16 v1, v2, v3 row_shl:1 row_mask:0xa bank_mask:0x1 bound_ctrl:0
4762 VOP_SDWA examples:
4764 .. code-block:: nasm
4766 v_mov_b32 v1, v2 dst_sel:BYTE_0 dst_unused:UNUSED_PRESERVE src0_sel:DWORD
4767 v_min_u32 v200, v200, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_1 src1_sel:DWORD
4768 v_sin_f32 v0, v0 dst_unused:UNUSED_PAD src0_sel:WORD_1
4769 v_fract_f32 v0, |v0| dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
4770 v_cmpx_le_u32 vcc, v1, v2 src0_sel:BYTE_2 src1_sel:WORD_0
4772 For full list of supported instructions, refer to "Vector ALU instructions".
4774 .. TODO
4775 Remove once we switch to code object v3 by default.
4777 HSA Code Object Directives
4778 ~~~~~~~~~~~~~~~~~~~~~~~~~~
4780 AMDGPU ABI defines auxiliary data in output code object. In assembly source,
4781 one can specify them with assembler directives.
4783 .hsa_code_object_version major, minor
4784 +++++++++++++++++++++++++++++++++++++
4786 *major* and *minor* are integers that specify the version of the HSA code
4787 object that will be generated by the assembler.
4789 .hsa_code_object_isa [major, minor, stepping, vendor, arch]
4790 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
4793 *major*, *minor*, and *stepping* are all integers that describe the instruction
4794 set architecture (ISA) version of the assembly program.
4796 *vendor* and *arch* are quoted strings. *vendor* should always be equal to
4797 "AMD" and *arch* should always be equal to "AMDGPU".
4799 By default, the assembler will derive the ISA version, *vendor*, and *arch*
4800 from the value of the -mcpu option that is passed to the assembler.
4802 .amdgpu_hsa_kernel (name)
4803 +++++++++++++++++++++++++
4805 This directives specifies that the symbol with given name is a kernel entry point
4806 (label) and the object should contain corresponding symbol of type STT_AMDGPU_HSA_KERNEL.
4808 .amd_kernel_code_t
4809 ++++++++++++++++++
4811 This directive marks the beginning of a list of key / value pairs that are used
4812 to specify the amd_kernel_code_t object that will be emitted by the assembler.
4813 The list must be terminated by the *.end_amd_kernel_code_t* directive. For
4814 any amd_kernel_code_t values that are unspecified a default value will be
4815 used. The default value for all keys is 0, with the following exceptions:
4817 - *kernel_code_version_major* defaults to 1.
4818 - *machine_kind* defaults to 1.
4819 - *machine_version_major*, *machine_version_minor*, and
4820 *machine_version_stepping* are derived from the value of the -mcpu option
4821 that is passed to the assembler.
4822 - *kernel_code_entry_byte_offset* defaults to 256.
4823 - *wavefront_size* defaults to 6.
4824 - *kernarg_segment_alignment*, *group_segment_alignment*, and
4825 *private_segment_alignment* default to 4. Note that alignments are specified
4826 as a power of 2, so a value of **n** means an alignment of 2^ **n**.
4828 The *.amd_kernel_code_t* directive must be placed immediately after the
4829 function label and before any instructions.
4831 For a full list of amd_kernel_code_t keys, refer to AMDGPU ABI document,
4832 comments in lib/Target/AMDGPU/AmdKernelCodeT.h and test/CodeGen/AMDGPU/hsa.s.
4834 Here is an example of a minimal amd_kernel_code_t specification:
4836 .. code-block:: none
4838 .hsa_code_object_version 1,0
4839 .hsa_code_object_isa
4841 .hsatext
4842 .globl hello_world
4843 .p2align 8
4844 .amdgpu_hsa_kernel hello_world
4846 hello_world:
4848 .amd_kernel_code_t
4849 enable_sgpr_kernarg_segment_ptr = 1
4850 is_ptr64 = 1
4851 compute_pgm_rsrc1_vgprs = 0
4852 compute_pgm_rsrc1_sgprs = 0
4853 compute_pgm_rsrc2_user_sgpr = 2
4854 kernarg_segment_byte_size = 8
4855 wavefront_sgpr_count = 2
4856 workitem_vgpr_count = 3
4857 .end_amd_kernel_code_t
4859 s_load_dwordx2 s[0:1], s[0:1] 0x0
4860 v_mov_b32 v0, 3.14159
4861 s_waitcnt lgkmcnt(0)
4862 v_mov_b32 v1, s0
4863 v_mov_b32 v2, s1
4864 flat_store_dword v[1:2], v0
4865 s_endpgm
4866 .Lfunc_end0:
4867 .size hello_world, .Lfunc_end0-hello_world
4869 Predefined Symbols (-mattr=+code-object-v3)
4870 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
4872 The AMDGPU assembler defines and updates some symbols automatically. These
4873 symbols do not affect code generation.
4875 .amdgcn.gfx_generation_number
4876 +++++++++++++++++++++++++++++
4878 Set to the GFX generation number of the target being assembled for. For
4879 example, when assembling for a "GFX9" target this will be set to the integer
4880 value "9". The possible GFX generation numbers are presented in
4881 :ref:`amdgpu-processors`.
4883 .amdgcn.next_free_vgpr
4884 ++++++++++++++++++++++
4886 Set to zero before assembly begins. At each instruction, if the current value
4887 of this symbol is less than or equal to the maximum VGPR number explicitly
4888 referenced within that instruction then the symbol value is updated to equal
4889 that VGPR number plus one.
4891 May be used to set the `.amdhsa_next_free_vpgr` directive in
4892 :ref:`amdhsa-kernel-directives-table`.
4894 May be set at any time, e.g. manually set to zero at the start of each kernel.
4896 .amdgcn.next_free_sgpr
4897 ++++++++++++++++++++++
4899 Set to zero before assembly begins. At each instruction, if the current value
4900 of this symbol is less than or equal the maximum SGPR number explicitly
4901 referenced within that instruction then the symbol value is updated to equal
4902 that SGPR number plus one.
4904 May be used to set the `.amdhsa_next_free_spgr` directive in
4905 :ref:`amdhsa-kernel-directives-table`.
4907 May be set at any time, e.g. manually set to zero at the start of each kernel.
4909 Code Object Directives (-mattr=+code-object-v3)
4910 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
4912 Directives which begin with ``.amdgcn`` are valid for all ``amdgcn``
4913 architecture processors, and are not OS-specific. Directives which begin with
4914 ``.amdhsa`` are specific to ``amdgcn`` architecture processors when the
4915 ``amdhsa`` OS is specified. See :ref:`amdgpu-target-triples` and
4916 :ref:`amdgpu-processors`.
4918 .amdgcn_target <target>
4919 +++++++++++++++++++++++
4921 Optional directive which declares the target supported by the containing
4922 assembler source file. Valid values are described in
4923 :ref:`amdgpu-amdhsa-code-object-target-identification`. Used by the assembler
4924 to validate command-line options such as ``-triple``, ``-mcpu``, and those
4925 which specify target features.
4927 .amdhsa_kernel <name>
4928 +++++++++++++++++++++
4930 Creates a correctly aligned AMDHSA kernel descriptor and a symbol,
4931 ``<name>.kd``, in the current location of the current section. Only valid when
4932 the OS is ``amdhsa``. ``<name>`` must be a symbol that labels the first
4933 instruction to execute, and does not need to be previously defined.
4935 Marks the beginning of a list of directives used to generate the bytes of a
4936 kernel descriptor, as described in :ref:`amdgpu-amdhsa-kernel-descriptor`.
4937 Directives which may appear in this list are described in
4938 :ref:`amdhsa-kernel-directives-table`. Directives may appear in any order, must
4939 be valid for the target being assembled for, and cannot be repeated. Directives
4940 support the range of values specified by the field they reference in
4941 :ref:`amdgpu-amdhsa-kernel-descriptor`. If a directive is not specified, it is
4942 assumed to have its default value, unless it is marked as "Required", in which
4943 case it is an error to omit the directive. This list of directives is
4944 terminated by an ``.end_amdhsa_kernel`` directive.
4946 .. table:: AMDHSA Kernel Assembler Directives
4947 :name: amdhsa-kernel-directives-table
4949 ======================================================== ================ ============ ===================
4950 Directive Default Supported On Description
4951 ======================================================== ================ ============ ===================
4952 ``.amdhsa_group_segment_fixed_size`` 0 GFX6-GFX9 Controls GROUP_SEGMENT_FIXED_SIZE in
4953 :ref:`amdgpu-amdhsa-kernel-descriptor-gfx6-gfx9-table`.
4954 ``.amdhsa_private_segment_fixed_size`` 0 GFX6-GFX9 Controls PRIVATE_SEGMENT_FIXED_SIZE in
4955 :ref:`amdgpu-amdhsa-kernel-descriptor-gfx6-gfx9-table`.
4956 ``.amdhsa_user_sgpr_private_segment_buffer`` 0 GFX6-GFX9 Controls ENABLE_SGPR_PRIVATE_SEGMENT_BUFFER in
4957 :ref:`amdgpu-amdhsa-kernel-descriptor-gfx6-gfx9-table`.
4958 ``.amdhsa_user_sgpr_dispatch_ptr`` 0 GFX6-GFX9 Controls ENABLE_SGPR_DISPATCH_PTR in
4959 :ref:`amdgpu-amdhsa-kernel-descriptor-gfx6-gfx9-table`.
4960 ``.amdhsa_user_sgpr_queue_ptr`` 0 GFX6-GFX9 Controls ENABLE_SGPR_QUEUE_PTR in
4961 :ref:`amdgpu-amdhsa-kernel-descriptor-gfx6-gfx9-table`.
4962 ``.amdhsa_user_sgpr_kernarg_segment_ptr`` 0 GFX6-GFX9 Controls ENABLE_SGPR_KERNARG_SEGMENT_PTR in
4963 :ref:`amdgpu-amdhsa-kernel-descriptor-gfx6-gfx9-table`.
4964 ``.amdhsa_user_sgpr_dispatch_id`` 0 GFX6-GFX9 Controls ENABLE_SGPR_DISPATCH_ID in
4965 :ref:`amdgpu-amdhsa-kernel-descriptor-gfx6-gfx9-table`.
4966 ``.amdhsa_user_sgpr_flat_scratch_init`` 0 GFX6-GFX9 Controls ENABLE_SGPR_FLAT_SCRATCH_INIT in
4967 :ref:`amdgpu-amdhsa-kernel-descriptor-gfx6-gfx9-table`.
4968 ``.amdhsa_user_sgpr_private_segment_size`` 0 GFX6-GFX9 Controls ENABLE_SGPR_PRIVATE_SEGMENT_SIZE in
4969 :ref:`amdgpu-amdhsa-kernel-descriptor-gfx6-gfx9-table`.
4970 ``.amdhsa_system_sgpr_private_segment_wavefront_offset`` 0 GFX6-GFX9 Controls ENABLE_SGPR_PRIVATE_SEGMENT_WAVEFRONT_OFFSET in
4971 :ref:`amdgpu-amdhsa-compute_pgm_rsrc2-gfx6-gfx9-table`.
4972 ``.amdhsa_system_sgpr_workgroup_id_x`` 1 GFX6-GFX9 Controls ENABLE_SGPR_WORKGROUP_ID_X in
4973 :ref:`amdgpu-amdhsa-compute_pgm_rsrc2-gfx6-gfx9-table`.
4974 ``.amdhsa_system_sgpr_workgroup_id_y`` 0 GFX6-GFX9 Controls ENABLE_SGPR_WORKGROUP_ID_Y in
4975 :ref:`amdgpu-amdhsa-compute_pgm_rsrc2-gfx6-gfx9-table`.
4976 ``.amdhsa_system_sgpr_workgroup_id_z`` 0 GFX6-GFX9 Controls ENABLE_SGPR_WORKGROUP_ID_Z in
4977 :ref:`amdgpu-amdhsa-compute_pgm_rsrc2-gfx6-gfx9-table`.
4978 ``.amdhsa_system_sgpr_workgroup_info`` 0 GFX6-GFX9 Controls ENABLE_SGPR_WORKGROUP_INFO in
4979 :ref:`amdgpu-amdhsa-compute_pgm_rsrc2-gfx6-gfx9-table`.
4980 ``.amdhsa_system_vgpr_workitem_id`` 0 GFX6-GFX9 Controls ENABLE_VGPR_WORKITEM_ID in
4981 :ref:`amdgpu-amdhsa-compute_pgm_rsrc2-gfx6-gfx9-table`.
4982 Possible values are defined in
4983 :ref:`amdgpu-amdhsa-system-vgpr-work-item-id-enumeration-values-table`.
4984 ``.amdhsa_next_free_vgpr`` Required GFX6-GFX9 Maximum VGPR number explicitly referenced, plus one.
4985 Used to calculate GRANULATED_WORKITEM_VGPR_COUNT in
4986 :ref:`amdgpu-amdhsa-compute_pgm_rsrc1-gfx6-gfx9-table`.
4987 ``.amdhsa_next_free_sgpr`` Required GFX6-GFX9 Maximum SGPR number explicitly referenced, plus one.
4988 Used to calculate GRANULATED_WAVEFRONT_SGPR_COUNT in
4989 :ref:`amdgpu-amdhsa-compute_pgm_rsrc1-gfx6-gfx9-table`.
4990 ``.amdhsa_reserve_vcc`` 1 GFX6-GFX9 Whether the kernel may use the special VCC SGPR.
4991 Used to calculate GRANULATED_WAVEFRONT_SGPR_COUNT in
4992 :ref:`amdgpu-amdhsa-compute_pgm_rsrc1-gfx6-gfx9-table`.
4993 ``.amdhsa_reserve_flat_scratch`` 1 GFX7-GFX9 Whether the kernel may use flat instructions to access
4994 scratch memory. Used to calculate
4995 GRANULATED_WAVEFRONT_SGPR_COUNT in
4996 :ref:`amdgpu-amdhsa-compute_pgm_rsrc1-gfx6-gfx9-table`.
4997 ``.amdhsa_reserve_xnack_mask`` Target GFX8-GFX9 Whether the kernel may trigger XNACK replay.
4998 Feature Used to calculate GRANULATED_WAVEFRONT_SGPR_COUNT in
4999 Specific :ref:`amdgpu-amdhsa-compute_pgm_rsrc1-gfx6-gfx9-table`.
5000 (+xnack)
5001 ``.amdhsa_float_round_mode_32`` 0 GFX6-GFX9 Controls FLOAT_ROUND_MODE_32 in
5002 :ref:`amdgpu-amdhsa-compute_pgm_rsrc1-gfx6-gfx9-table`.
5003 Possible values are defined in
5004 :ref:`amdgpu-amdhsa-floating-point-rounding-mode-enumeration-values-table`.
5005 ``.amdhsa_float_round_mode_16_64`` 0 GFX6-GFX9 Controls FLOAT_ROUND_MODE_16_64 in
5006 :ref:`amdgpu-amdhsa-compute_pgm_rsrc1-gfx6-gfx9-table`.
5007 Possible values are defined in
5008 :ref:`amdgpu-amdhsa-floating-point-rounding-mode-enumeration-values-table`.
5009 ``.amdhsa_float_denorm_mode_32`` 0 GFX6-GFX9 Controls FLOAT_DENORM_MODE_32 in
5010 :ref:`amdgpu-amdhsa-compute_pgm_rsrc1-gfx6-gfx9-table`.
5011 Possible values are defined in
5012 :ref:`amdgpu-amdhsa-floating-point-denorm-mode-enumeration-values-table`.
5013 ``.amdhsa_float_denorm_mode_16_64`` 3 GFX6-GFX9 Controls FLOAT_DENORM_MODE_16_64 in
5014 :ref:`amdgpu-amdhsa-compute_pgm_rsrc1-gfx6-gfx9-table`.
5015 Possible values are defined in
5016 :ref:`amdgpu-amdhsa-floating-point-denorm-mode-enumeration-values-table`.
5017 ``.amdhsa_dx10_clamp`` 1 GFX6-GFX9 Controls ENABLE_DX10_CLAMP in
5018 :ref:`amdgpu-amdhsa-compute_pgm_rsrc1-gfx6-gfx9-table`.
5019 ``.amdhsa_ieee_mode`` 1 GFX6-GFX9 Controls ENABLE_IEEE_MODE in
5020 :ref:`amdgpu-amdhsa-compute_pgm_rsrc1-gfx6-gfx9-table`.
5021 ``.amdhsa_fp16_overflow`` 0 GFX9 Controls FP16_OVFL in
5022 :ref:`amdgpu-amdhsa-compute_pgm_rsrc1-gfx6-gfx9-table`.
5023 ``.amdhsa_exception_fp_ieee_invalid_op`` 0 GFX6-GFX9 Controls ENABLE_EXCEPTION_IEEE_754_FP_INVALID_OPERATION in
5024 :ref:`amdgpu-amdhsa-compute_pgm_rsrc2-gfx6-gfx9-table`.
5025 ``.amdhsa_exception_fp_denorm_src`` 0 GFX6-GFX9 Controls ENABLE_EXCEPTION_FP_DENORMAL_SOURCE in
5026 :ref:`amdgpu-amdhsa-compute_pgm_rsrc2-gfx6-gfx9-table`.
5027 ``.amdhsa_exception_fp_ieee_div_zero`` 0 GFX6-GFX9 Controls ENABLE_EXCEPTION_IEEE_754_FP_DIVISION_BY_ZERO in
5028 :ref:`amdgpu-amdhsa-compute_pgm_rsrc2-gfx6-gfx9-table`.
5029 ``.amdhsa_exception_fp_ieee_overflow`` 0 GFX6-GFX9 Controls ENABLE_EXCEPTION_IEEE_754_FP_OVERFLOW in
5030 :ref:`amdgpu-amdhsa-compute_pgm_rsrc2-gfx6-gfx9-table`.
5031 ``.amdhsa_exception_fp_ieee_underflow`` 0 GFX6-GFX9 Controls ENABLE_EXCEPTION_IEEE_754_FP_UNDERFLOW in
5032 :ref:`amdgpu-amdhsa-compute_pgm_rsrc2-gfx6-gfx9-table`.
5033 ``.amdhsa_exception_fp_ieee_inexact`` 0 GFX6-GFX9 Controls ENABLE_EXCEPTION_IEEE_754_FP_INEXACT in
5034 :ref:`amdgpu-amdhsa-compute_pgm_rsrc2-gfx6-gfx9-table`.
5035 ``.amdhsa_exception_int_div_zero`` 0 GFX6-GFX9 Controls ENABLE_EXCEPTION_INT_DIVIDE_BY_ZERO in
5036 :ref:`amdgpu-amdhsa-compute_pgm_rsrc2-gfx6-gfx9-table`.
5037 ======================================================== ================ ============ ===================
5039 .amdgpu_metadata
5040 ++++++++++++++++
5042 Optional directive which declares the contents of the ``NT_AMDGPU_METADATA``
5043 note record (see :ref:`amdgpu-elf-note-records-table-v3`).
5045 The contents must be in the [YAML]_ markup format, with the same structure and
5046 semantics described in :ref:`amdgpu-amdhsa-code-object-metadata-v3`.
5048 This directive is terminated by an ``.end_amdgpu_metadata`` directive.
5050 Example HSA Source Code (-mattr=+code-object-v3)
5051 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
5053 Here is an example of a minimal assembly source file, defining one HSA kernel:
5055 .. code-block:: none
5057 .amdgcn_target "amdgcn-amd-amdhsa--gfx900+xnack" // optional
5059 .text
5060 .globl hello_world
5061 .p2align 8
5062 .type hello_world,@function
5063 hello_world:
5064 s_load_dwordx2 s[0:1], s[0:1] 0x0
5065 v_mov_b32 v0, 3.14159
5066 s_waitcnt lgkmcnt(0)
5067 v_mov_b32 v1, s0
5068 v_mov_b32 v2, s1
5069 flat_store_dword v[1:2], v0
5070 s_endpgm
5071 .Lfunc_end0:
5072 .size hello_world, .Lfunc_end0-hello_world
5074 .rodata
5075 .p2align 6
5076 .amdhsa_kernel hello_world
5077 .amdhsa_user_sgpr_kernarg_segment_ptr 1
5078 .amdhsa_next_free_vgpr .amdgcn.next_free_vgpr
5079 .amdhsa_next_free_sgpr .amdgcn.next_free_sgpr
5080 .end_amdhsa_kernel
5082 .amdgpu_metadata
5083 ---
5084 amdhsa.version:
5085 - 1
5086 - 0
5087 amdhsa.kernels:
5088 - .name: hello_world
5089 .symbol: hello_world.kd
5090 .kernarg_segment_size: 48
5091 .group_segment_fixed_size: 0
5092 .private_segment_fixed_size: 0
5093 .kernarg_segment_align: 4
5094 .wavefront_size: 64
5095 .sgpr_count: 2
5096 .vgpr_count: 3
5097 .max_flat_workgroup_size: 256
5098 ...
5099 .end_amdgpu_metadata
5101 Additional Documentation
5102 ========================
5104 .. [AMD-RADEON-HD-2000-3000] `AMD R6xx shader ISA <http://developer.amd.com/wordpress/media/2012/10/R600_Instruction_Set_Architecture.pdf>`__
5105 .. [AMD-RADEON-HD-4000] `AMD R7xx shader ISA <http://developer.amd.com/wordpress/media/2012/10/R700-Family_Instruction_Set_Architecture.pdf>`__
5106 .. [AMD-RADEON-HD-5000] `AMD Evergreen shader ISA <http://developer.amd.com/wordpress/media/2012/10/AMD_Evergreen-Family_Instruction_Set_Architecture.pdf>`__
5107 .. [AMD-RADEON-HD-6000] `AMD Cayman/Trinity shader ISA <http://developer.amd.com/wordpress/media/2012/10/AMD_HD_6900_Series_Instruction_Set_Architecture.pdf>`__
5108 .. [AMD-GCN-GFX6] `AMD Southern Islands Series ISA <http://developer.amd.com/wordpress/media/2012/12/AMD_Southern_Islands_Instruction_Set_Architecture.pdf>`__
5109 .. [AMD-GCN-GFX7] `AMD Sea Islands Series ISA <http://developer.amd.com/wordpress/media/2013/07/AMD_Sea_Islands_Instruction_Set_Architecture.pdf>`_
5110 .. [AMD-GCN-GFX8] `AMD GCN3 Instruction Set Architecture <http://amd-dev.wpengine.netdna-cdn.com/wordpress/media/2013/12/AMD_GCN3_Instruction_Set_Architecture_rev1.1.pdf>`__
5111 .. [AMD-GCN-GFX9] `AMD "Vega" Instruction Set Architecture <http://developer.amd.com/wordpress/media/2013/12/Vega_Shader_ISA_28July2017.pdf>`__
5112 .. [AMD-ROCm] `ROCm: Open Platform for Development, Discovery and Education Around GPU Computing <http://gpuopen.com/compute-product/rocm/>`__
5113 .. [AMD-ROCm-github] `ROCm github <http://github.com/RadeonOpenCompute>`__
5114 .. [HSA] `Heterogeneous System Architecture (HSA) Foundation <http://www.hsafoundation.com/>`__
5115 .. [ELF] `Executable and Linkable Format (ELF) <http://www.sco.com/developers/gabi/>`__
5116 .. [DWARF] `DWARF Debugging Information Format <http://dwarfstd.org/>`__
5117 .. [YAML] `YAML Ain't Markup Language (YAML™) Version 1.2 <http://www.yaml.org/spec/1.2/spec.html>`__
5118 .. [MsgPack] `Message Pack <http://www.msgpack.org/>`__
5119 .. [OpenCL] `The OpenCL Specification Version 2.0 <http://www.khronos.org/registry/cl/specs/opencl-2.0.pdf>`__
5120 .. [HRF] `Heterogeneous-race-free Memory Models <http://benedictgaster.org/wp-content/uploads/2014/01/asplos269-FINAL.pdf>`__
5121 .. [CLANG-ATTR] `Attributes in Clang <http://clang.llvm.org/docs/AttributeReference.html>`__