1 ======================================
2 Syntax of AMDGPU Instruction Modifiers
3 ======================================
11 The following notation is used throughout this document:
13 =================== =============================================================
15 =================== =============================================================
16 {0..N} Any integer value in the range from 0 to N (inclusive).
17 <x> Syntax and meaning of *x* is explained elsewhere.
18 =================== =============================================================
20 .. _amdgpu_syn_modifiers:
28 .. _amdgpu_synid_ds_offset8:
33 Specifies an immediate unsigned 8-bit offset, in bytes. The default value is 0.
35 Used with DS instructions which have 2 addresses.
37 =================== ====================================================================
39 =================== ====================================================================
40 offset:{0..0xFF} Specifies an unsigned 8-bit offset as a positive
41 :ref:`integer number <amdgpu_synid_integer_number>`
42 or an :ref:`absolute expression<amdgpu_synid_absolute_expression>`.
43 =================== ====================================================================
53 .. _amdgpu_synid_ds_offset16:
58 Specifies an immediate unsigned 16-bit offset, in bytes. The default value is 0.
60 Used with DS instructions which have 1 address.
62 ==================== ====================================================================
64 ==================== ====================================================================
65 offset:{0..0xFFFF} Specifies an unsigned 16-bit offset as a positive
66 :ref:`integer number <amdgpu_synid_integer_number>`
67 or an :ref:`absolute expression<amdgpu_synid_absolute_expression>`.
68 ==================== ====================================================================
78 .. _amdgpu_synid_sw_offset16:
83 This is a special modifier which may be used with *ds_swizzle_b32* instruction only.
84 It specifies a swizzle pattern in numeric or symbolic form. The default value is 0.
86 See AMD documentation for more information.
88 ======================================================= ===========================================================
90 ======================================================= ===========================================================
91 offset:{0..0xFFFF} Specifies a 16-bit swizzle pattern.
92 offset:swizzle(QUAD_PERM,{0..3},{0..3},{0..3},{0..3}) Specifies a quad permute mode pattern
94 Each number is a lane *id*.
95 offset:swizzle(BITMASK_PERM, "<mask>") Specifies a bitmask permute mode pattern.
97 The pattern converts a 5-bit lane *id* to another
98 lane *id* with which the lane interacts.
100 *mask* is a 5 character sequence which
101 specifies how to transform the bits of the
104 The following characters are allowed:
106 * "0" - set bit to 0.
108 * "1" - set bit to 1.
110 * "p" - preserve bit.
114 offset:swizzle(BROADCAST,{2..32},{0..N}) Specifies a broadcast mode.
116 Broadcasts the value of any particular lane to
117 all lanes in its group.
119 The first numeric parameter is a group
120 size and must be equal to 2, 4, 8, 16 or 32.
122 The second numeric parameter is an index of the
123 lane being broadcasted.
125 The index must not exceed group size.
126 offset:swizzle(SWAP,{1..16}) Specifies a swap mode.
128 Swaps the neighboring groups of
129 1, 2, 4, 8 or 16 lanes.
130 offset:swizzle(REVERSE,{2..32}) Specifies a reverse mode.
132 Reverses the lanes for groups of 2, 4, 8, 16 or 32 lanes.
133 ======================================================= ===========================================================
135 Note: numeric values may be specified as either :ref:`integer numbers<amdgpu_synid_integer_number>` or
136 :ref:`absolute expressions<amdgpu_synid_absolute_expression>`.
144 offset:swizzle(QUAD_PERM, 0, 1, 2, 3)
145 offset:swizzle(BITMASK_PERM, "01pi0")
146 offset:swizzle(BROADCAST, 2, 0)
147 offset:swizzle(SWAP, 8)
148 offset:swizzle(REVERSE, 30 + 2)
150 .. _amdgpu_synid_gds:
155 Specifies whether to use GDS or LDS memory (LDS is the default).
157 ======================================== ================================================
159 ======================================== ================================================
161 ======================================== ================================================
167 .. _amdgpu_synid_done:
172 Specifies if this is the last export from the shader to the target. By default,
173 *exp* instruction does not finish an export sequence.
175 ======================================== ================================================
177 ======================================== ================================================
178 done Indicates the last export operation.
179 ======================================== ================================================
181 .. _amdgpu_synid_compr:
186 Indicates if the data are compressed (data are not compressed by default).
188 ======================================== ================================================
190 ======================================== ================================================
191 compr Data are compressed.
192 ======================================== ================================================
199 Specifies valid mask flag state (off by default).
201 ======================================== ================================================
203 ======================================== ================================================
204 vm Set valid mask flag.
205 ======================================== ================================================
210 .. _amdgpu_synid_flat_offset12:
215 Specifies an immediate unsigned 12-bit offset, in bytes. The default value is 0.
217 Cannot be used with *global/scratch* opcodes. GFX9 only.
219 ================= ====================================================================
221 ================= ====================================================================
222 offset:{0..4095} Specifies a 12-bit unsigned offset as a positive
223 :ref:`integer number <amdgpu_synid_integer_number>`
224 or an :ref:`absolute expression<amdgpu_synid_absolute_expression>`.
225 ================= ====================================================================
234 .. _amdgpu_synid_flat_offset13s:
239 Specifies an immediate signed 13-bit offset, in bytes. The default value is 0.
241 Can be used with *global/scratch* opcodes only. GFX9 only.
243 ===================== ====================================================================
245 ===================== ====================================================================
246 offset:{-4096..4095} Specifies a 13-bit signed offset as an
247 :ref:`integer number <amdgpu_synid_integer_number>`
248 or an :ref:`absolute expression<amdgpu_synid_absolute_expression>`.
249 ===================== ====================================================================
259 .. _amdgpu_synid_flat_offset12s:
264 Specifies an immediate signed 12-bit offset, in bytes. The default value is 0.
266 Can be used with *global/scratch* opcodes only.
270 ===================== ====================================================================
272 ===================== ====================================================================
273 offset:{-2048..2047} Specifies a 12-bit signed offset as an
274 :ref:`integer number <amdgpu_synid_integer_number>`
275 or an :ref:`absolute expression<amdgpu_synid_absolute_expression>`.
276 ===================== ====================================================================
286 .. _amdgpu_synid_flat_offset11:
291 Specifies an immediate unsigned 11-bit offset, in bytes. The default value is 0.
293 Cannot be used with *global/scratch* opcodes.
297 ================= ====================================================================
299 ================= ====================================================================
300 offset:{0..2047} Specifies an 11-bit unsigned offset as a positive
301 :ref:`integer number <amdgpu_synid_integer_number>`
302 or an :ref:`absolute expression<amdgpu_synid_absolute_expression>`.
303 ================= ====================================================================
315 See a description :ref:`here<amdgpu_synid_dlc>`. GFX10 only.
320 See a description :ref:`here<amdgpu_synid_glc>`.
325 See a description :ref:`here<amdgpu_synid_lds>`. GFX10 only.
330 See a description :ref:`here<amdgpu_synid_slc>`.
335 See a description :ref:`here<amdgpu_synid_tfe>`.
340 See a description :ref:`here<amdgpu_synid_nv>`.
345 .. _amdgpu_synid_dmask:
350 Specifies which channels (image components) are used by the operation. By default, no channels
353 =============== ====================================================================
355 =============== ====================================================================
356 dmask:{0..15} Specifies image channels as a positive
357 :ref:`integer number <amdgpu_synid_integer_number>`
358 or an :ref:`absolute expression<amdgpu_synid_absolute_expression>`.
360 Each bit corresponds to one of 4 image components (RGBA).
362 If the specified bit value is 0, the component is not used,
363 value 1 means that the component is used.
364 =============== ====================================================================
366 This modifier has some limitations depending on instruction kind:
368 =================================================== ========================
369 Instruction Kind Valid dmask Values
370 =================================================== ========================
371 32-bit atomic *cmpswap* 0x3
372 32-bit atomic instructions except for *cmpswap* 0x1
373 64-bit atomic *cmpswap* 0xF
374 64-bit atomic instructions except for *cmpswap* 0x3
375 *gather4* 0x1, 0x2, 0x4, 0x8
376 Other instructions any value
377 =================================================== ========================
387 .. _amdgpu_synid_unorm:
392 Specifies whether the address is normalized or not (the address is normalized by default).
394 ======================== ========================================
396 ======================== ========================================
397 unorm Force the address to be unnormalized.
398 ======================== ========================================
403 See a description :ref:`here<amdgpu_synid_glc>`.
408 See a description :ref:`here<amdgpu_synid_slc>`.
410 .. _amdgpu_synid_r128:
415 Specifies texture resource size. The default size is 256 bits.
417 GFX7, GFX8 and GFX10 only.
419 =================== ================================================
421 =================== ================================================
422 r128 Specifies 128 bits texture resource size.
423 =================== ================================================
425 .. WARNING:: Using this modifier should descrease *rsrc* operand size from 8 to 4 dwords, but assembler does not currently support this feature.
430 See a description :ref:`here<amdgpu_synid_tfe>`.
432 .. _amdgpu_synid_lwe:
437 Specifies LOD warning status (LOD warning is disabled by default).
439 ======================================== ================================================
441 ======================================== ================================================
442 lwe Enables LOD warning.
443 ======================================== ================================================
450 Specifies if an array index must be sent to TA. By default, array index is not sent.
452 ======================================== ================================================
454 ======================================== ================================================
455 da Send an array-index to TA.
456 ======================================== ================================================
458 .. _amdgpu_synid_d16:
463 Specifies data size: 16 or 32 bits (32 bits by default). Not supported by GFX7.
465 ======================================== ================================================
467 ======================================== ================================================
468 d16 Enables 16-bits data mode.
470 On loads, convert data in memory to 16-bit
471 format before storing it in VGPRs.
473 For stores, convert 16-bit data in VGPRs to
474 32 bits before going to memory.
476 Note that GFX8.0 does not support data packing.
477 Each 16-bit data element occupies 1 VGPR.
479 GFX8.1, GFX9 and GFX10 support data packing.
480 Each pair of 16-bit data elements
482 ======================================== ================================================
484 .. _amdgpu_synid_a16:
489 Specifies size of image address components: 16 or 32 bits (32 bits by default).
492 ======================================== ================================================
494 ======================================== ================================================
495 a16 Enables 16-bits image address components.
496 ======================================== ================================================
498 .. _amdgpu_synid_dim:
503 Specifies surface dimension. This is a mandatory modifier. There is no default value.
507 =============================== =========================================================
509 =============================== =========================================================
510 dim:1D One-dimensional image.
511 dim:2D Two-dimensional image.
512 dim:3D Three-dimensional image.
513 dim:CUBE Cubemap array.
514 dim:1D_ARRAY One-dimensional image array.
515 dim:2D_ARRAY Two-dimensional image array.
516 dim:2D_MSAA Two-dimensional multi-sample auto-aliasing image.
517 dim:2D_MSAA_ARRAY Two-dimensional multi-sample auto-aliasing image array.
518 =============================== =========================================================
520 The following table defines an alternative syntax which is supported
521 for compatibility with SP3 assembler:
523 =============================== =========================================================
525 =============================== =========================================================
526 dim:SQ_RSRC_IMG_1D One-dimensional image.
527 dim:SQ_RSRC_IMG_2D Two-dimensional image.
528 dim:SQ_RSRC_IMG_3D Three-dimensional image.
529 dim:SQ_RSRC_IMG_CUBE Cubemap array.
530 dim:SQ_RSRC_IMG_1D_ARRAY One-dimensional image array.
531 dim:SQ_RSRC_IMG_2D_ARRAY Two-dimensional image array.
532 dim:SQ_RSRC_IMG_2D_MSAA Two-dimensional multi-sample auto-aliasing image.
533 dim:SQ_RSRC_IMG_2D_MSAA_ARRAY Two-dimensional multi-sample auto-aliasing image array.
534 =============================== =========================================================
539 See a description :ref:`here<amdgpu_synid_dlc>`. GFX10 only.
541 Miscellaneous Modifiers
542 -----------------------
544 .. _amdgpu_synid_dlc:
549 Controls device level cache policy for memory operations. Used for synchronization.
550 When specified, forces operation to bypass device level cache making the operation device
551 level coherent. By default, instructions use device level cache.
555 ======================================== ================================================
557 ======================================== ================================================
558 dlc Bypass device level cache.
559 ======================================== ================================================
561 .. _amdgpu_synid_glc:
566 This modifier has different meaning for loads, stores, and atomic operations.
567 The default value is off (0).
569 See AMD documentation for details.
571 ======================================== ================================================
573 ======================================== ================================================
574 glc Set glc bit to 1.
575 ======================================== ================================================
577 .. _amdgpu_synid_lds:
582 Specifies where to store the result: VGPRs or LDS (VGPRs by default).
584 ======================================== ===========================
586 ======================================== ===========================
587 lds Store result in LDS.
588 ======================================== ===========================
595 Specifies if instruction is operating on non-volatile memory. By default, memory is volatile.
599 ======================================== ================================================
601 ======================================== ================================================
602 nv Indicates that instruction operates on
604 ======================================== ================================================
606 .. _amdgpu_synid_slc:
611 Specifies cache policy. The default value is off (0).
613 See AMD documentation for details.
615 ======================================== ================================================
617 ======================================== ================================================
618 slc Set slc bit to 1.
619 ======================================== ================================================
621 .. _amdgpu_synid_tfe:
626 Controls access to partially resident textures. The default value is off (0).
628 See AMD documentation for details.
630 ======================================== ================================================
632 ======================================== ================================================
633 tfe Set tfe bit to 1.
634 ======================================== ================================================
636 MUBUF/MTBUF Modifiers
637 ---------------------
639 .. _amdgpu_synid_idxen:
644 Specifies whether address components include an index. By default, no components are used.
646 Can be used together with :ref:`offen<amdgpu_synid_offen>`.
648 Cannot be used with :ref:`addr64<amdgpu_synid_addr64>`.
650 ======================================== ================================================
652 ======================================== ================================================
653 idxen Address components include an index.
654 ======================================== ================================================
656 .. _amdgpu_synid_offen:
661 Specifies whether address components include an offset. By default, no components are used.
663 Can be used together with :ref:`idxen<amdgpu_synid_idxen>`.
665 Cannot be used with :ref:`addr64<amdgpu_synid_addr64>`.
667 ======================================== ================================================
669 ======================================== ================================================
670 offen Address components include an offset.
671 ======================================== ================================================
673 .. _amdgpu_synid_addr64:
678 Specifies whether a 64-bit address is used. By default, no address is used.
680 GFX7 only. Cannot be used with :ref:`offen<amdgpu_synid_offen>` and
681 :ref:`idxen<amdgpu_synid_idxen>` modifiers.
683 ======================================== ================================================
685 ======================================== ================================================
686 addr64 A 64-bit address is used.
687 ======================================== ================================================
689 .. _amdgpu_synid_buf_offset12:
694 Specifies an immediate unsigned 12-bit offset, in bytes. The default value is 0.
696 ================== ====================================================================
698 ================== ====================================================================
699 offset:{0..0xFFF} Specifies a 12-bit unsigned offset as a positive
700 :ref:`integer number <amdgpu_synid_integer_number>`
701 or an :ref:`absolute expression<amdgpu_synid_absolute_expression>`.
702 ================== ====================================================================
714 See a description :ref:`here<amdgpu_synid_glc>`.
719 See a description :ref:`here<amdgpu_synid_slc>`.
724 See a description :ref:`here<amdgpu_synid_lds>`.
729 See a description :ref:`here<amdgpu_synid_dlc>`. GFX10 only.
734 See a description :ref:`here<amdgpu_synid_tfe>`.
736 .. _amdgpu_synid_dfmt:
743 .. _amdgpu_synid_nfmt:
756 See a description :ref:`here<amdgpu_synid_glc>`.
761 See a description :ref:`here<amdgpu_synid_nv>`. GFX9 only.
766 See a description :ref:`here<amdgpu_synid_dlc>`. GFX10 only.
771 .. _amdgpu_synid_high:
776 Specifies which half of the LDS word to use. Low half of LDS word is used by default.
779 ======================================== ================================
781 ======================================== ================================
782 high Use high half of LDS word.
783 ======================================== ================================
790 .. _amdgpu_synid_dpp8_sel:
795 Selects which lanes to pull data from, within a group of 8 lanes. This is a mandatory modifier.
796 There is no default value.
800 The *dpp8_sel* modifier must specify exactly 8 values.
801 First value selects which lane to read from to supply data into lane 0.
802 Second value controls lane 1 and so on.
804 Each value may be specified as either
805 an :ref:`integer number<amdgpu_synid_integer_number>` or
806 an :ref:`absolute expression<amdgpu_synid_absolute_expression>`.
808 =============================================================== ===========================
810 =============================================================== ===========================
811 dpp8:[{0..7},{0..7},{0..7},{0..7},{0..7},{0..7},{0..7},{0..7}] Select lanes to read from.
812 =============================================================== ===========================
818 dpp8:[7,6,5,4,3,2,1,0]
819 dpp8:[0,1,0,1,0,1,0,1]
821 .. _amdgpu_synid_fi8:
826 Controls interaction with inactive lanes for *dpp8* instructions. The default value is zero.
828 Note: *inactive* lanes are those whose :ref:`exec<amdgpu_synid_exec>` mask bit is zero.
832 ==================================== =====================================================
834 ==================================== =====================================================
835 fi:0 Fetch zero when accessing data from inactive lanes.
836 fi:1 Fetch pre-exist values from inactive lanes.
837 ==================================== =====================================================
839 Note: numeric values may be specified as either :ref:`integer numbers<amdgpu_synid_integer_number>` or
840 :ref:`absolute expressions<amdgpu_synid_absolute_expression>`.
845 GFX8, GFX9 and GFX10 only.
847 .. _amdgpu_synid_dpp_ctrl:
852 Specifies how data are shared between threads. This is a mandatory modifier.
853 There is no default value.
855 GFX8 and GFX9 only. Use :ref:`dpp16_ctrl<amdgpu_synid_dpp16_ctrl>` for GFX10.
857 Note: the lanes of a wavefront are organized in four *rows* and four *banks*.
859 ======================================== ================================================
861 ======================================== ================================================
862 quad_perm:[{0..3},{0..3},{0..3},{0..3}] Full permute of 4 threads.
863 row_mirror Mirror threads within row.
864 row_half_mirror Mirror threads within 1/2 row (8 threads).
865 row_bcast:15 Broadcast 15th thread of each row to next row.
866 row_bcast:31 Broadcast thread 31 to rows 2 and 3.
867 wave_shl:1 Wavefront left shift by 1 thread.
868 wave_rol:1 Wavefront left rotate by 1 thread.
869 wave_shr:1 Wavefront right shift by 1 thread.
870 wave_ror:1 Wavefront right rotate by 1 thread.
871 row_shl:{1..15} Row shift left by 1-15 threads.
872 row_shr:{1..15} Row shift right by 1-15 threads.
873 row_ror:{1..15} Row rotate right by 1-15 threads.
874 ======================================== ================================================
876 Note: numeric values may be specified as either
877 :ref:`integer numbers<amdgpu_synid_integer_number>` or
878 :ref:`absolute expressions<amdgpu_synid_absolute_expression>`.
884 quad_perm:[0, 1, 2, 3]
887 .. _amdgpu_synid_dpp16_ctrl:
892 Specifies how data are shared between threads. This is a mandatory modifier.
893 There is no default value.
895 GFX10 only. Use :ref:`dpp_ctrl<amdgpu_synid_dpp_ctrl>` for GFX8 and GFX9.
897 Note: the lanes of a wavefront are organized in four *rows* and four *banks*.
898 (There are only two rows in *wave32* mode.)
900 ======================================== ====================================================
902 ======================================== ====================================================
903 quad_perm:[{0..3},{0..3},{0..3},{0..3}] Full permute of 4 threads.
904 row_mirror Mirror threads within row.
905 row_half_mirror Mirror threads within 1/2 row (8 threads).
906 row_share:{0..15} Share the value from the specified lane with other
908 row_xmask:{0..15} Fetch from XOR(current lane id, specified lane id).
909 row_shl:{1..15} Row shift left by 1-15 threads.
910 row_shr:{1..15} Row shift right by 1-15 threads.
911 row_ror:{1..15} Row rotate right by 1-15 threads.
912 ======================================== ====================================================
914 Note: numeric values may be specified as either
915 :ref:`integer numbers<amdgpu_synid_integer_number>` or
916 :ref:`absolute expressions<amdgpu_synid_absolute_expression>`.
922 quad_perm:[0, 1, 2, 3]
925 .. _amdgpu_synid_row_mask:
930 Controls which rows are enabled for data sharing. By default, all rows are enabled.
932 Note: the lanes of a wavefront are organized in four *rows* and four *banks*.
933 (There are only two rows in *wave32* mode.)
935 ================= ====================================================================
937 ================= ====================================================================
938 row_mask:{0..15} Specifies a *row mask* as a positive
939 :ref:`integer number <amdgpu_synid_integer_number>`
940 or an :ref:`absolute expression<amdgpu_synid_absolute_expression>`.
942 Each of 4 bits in the mask controls one row
943 (0 - disabled, 1 - enabled).
945 In *wave32* mode the values should be limited to 0..7.
946 ================= ====================================================================
956 .. _amdgpu_synid_bank_mask:
961 Controls which banks are enabled for data sharing. By default, all banks are enabled.
963 Note: the lanes of a wavefront are organized in four *rows* and four *banks*.
964 (There are only two rows in *wave32* mode.)
966 ================== ====================================================================
968 ================== ====================================================================
969 bank_mask:{0..15} Specifies a *bank mask* as a positive
970 :ref:`integer number <amdgpu_synid_integer_number>`
971 or an :ref:`absolute expression<amdgpu_synid_absolute_expression>`.
973 Each of 4 bits in the mask controls one bank
974 (0 - disabled, 1 - enabled).
975 ================== ====================================================================
985 .. _amdgpu_synid_bound_ctrl:
990 Controls data sharing when accessing an invalid lane. By default, data sharing with
991 invalid lanes is disabled.
993 ======================================== ================================================
995 ======================================== ================================================
996 bound_ctrl:0 Enables data sharing with invalid lanes.
998 Accessing data from an invalid lane will
1000 ======================================== ================================================
1002 .. _amdgpu_synid_fi16:
1007 Controls interaction with *inactive* lanes for *dpp16* instructions. The default value is zero.
1009 Note: *inactive* lanes are those whose :ref:`exec<amdgpu_synid_exec>` mask bit is zero.
1013 ======================================== ==================================================
1015 ======================================== ==================================================
1016 fi:0 Interaction with inactive lanes is controlled by
1017 :ref:`bound_ctrl<amdgpu_synid_bound_ctrl>`.
1019 fi:1 Fetch pre-exist values from inactive lanes.
1020 ======================================== ==================================================
1022 Note: numeric values may be specified as either :ref:`integer numbers<amdgpu_synid_integer_number>` or
1023 :ref:`absolute expressions<amdgpu_synid_absolute_expression>`.
1028 GFX8, GFX9 and GFX10 only.
1033 See a description :ref:`here<amdgpu_synid_clamp>`.
1038 See a description :ref:`here<amdgpu_synid_omod>`.
1040 GFX9 and GFX10 only.
1042 .. _amdgpu_synid_dst_sel:
1047 Selects which bits in the destination are affected. By default, all bits are affected.
1049 ======================================== ================================================
1051 ======================================== ================================================
1052 dst_sel:DWORD Use bits 31:0.
1053 dst_sel:BYTE_0 Use bits 7:0.
1054 dst_sel:BYTE_1 Use bits 15:8.
1055 dst_sel:BYTE_2 Use bits 23:16.
1056 dst_sel:BYTE_3 Use bits 31:24.
1057 dst_sel:WORD_0 Use bits 15:0.
1058 dst_sel:WORD_1 Use bits 31:16.
1059 ======================================== ================================================
1061 .. _amdgpu_synid_dst_unused:
1066 Controls what to do with the bits in the destination which are not selected
1067 by :ref:`dst_sel<amdgpu_synid_dst_sel>`.
1068 By default, unused bits are preserved.
1070 ======================================== ================================================
1072 ======================================== ================================================
1073 dst_unused:UNUSED_PAD Pad with zeros.
1074 dst_unused:UNUSED_SEXT Sign-extend upper bits, zero lower bits.
1075 dst_unused:UNUSED_PRESERVE Preserve bits.
1076 ======================================== ================================================
1078 .. _amdgpu_synid_src0_sel:
1083 Controls which bits in the src0 are used. By default, all bits are used.
1085 ======================================== ================================================
1087 ======================================== ================================================
1088 src0_sel:DWORD Use bits 31:0.
1089 src0_sel:BYTE_0 Use bits 7:0.
1090 src0_sel:BYTE_1 Use bits 15:8.
1091 src0_sel:BYTE_2 Use bits 23:16.
1092 src0_sel:BYTE_3 Use bits 31:24.
1093 src0_sel:WORD_0 Use bits 15:0.
1094 src0_sel:WORD_1 Use bits 31:16.
1095 ======================================== ================================================
1097 .. _amdgpu_synid_src1_sel:
1102 Controls which bits in the src1 are used. By default, all bits are used.
1104 ======================================== ================================================
1106 ======================================== ================================================
1107 src1_sel:DWORD Use bits 31:0.
1108 src1_sel:BYTE_0 Use bits 7:0.
1109 src1_sel:BYTE_1 Use bits 15:8.
1110 src1_sel:BYTE_2 Use bits 23:16.
1111 src1_sel:BYTE_3 Use bits 31:24.
1112 src1_sel:WORD_0 Use bits 15:0.
1113 src1_sel:WORD_1 Use bits 31:16.
1114 ======================================== ================================================
1116 .. _amdgpu_synid_sdwa_operand_modifiers:
1118 SDWA Operand Modifiers
1119 ----------------------
1121 Operand modifiers are not used separately. They are applied to source operands.
1123 GFX8, GFX9 and GFX10 only.
1128 See a description :ref:`here<amdgpu_synid_abs>`.
1133 See a description :ref:`here<amdgpu_synid_neg>`.
1135 .. _amdgpu_synid_sext:
1140 Sign-extends value of a (sub-dword) operand to fill all 32 bits.
1141 Has no effect for 32-bit operands.
1143 Valid for integer operands only.
1145 ======================================== ================================================
1147 ======================================== ================================================
1148 sext(<operand>) Sign-extend operand value.
1149 ======================================== ================================================
1161 .. _amdgpu_synid_vop3_op_sel:
1166 Selects the low [15:0] or high [31:16] operand bits for source and destination operands.
1167 By default, low bits are used for all operands.
1169 The number of values specified with the op_sel modifier must match the number of instruction
1170 operands (both source and destination). First value controls src0, second value controls src1
1171 and so on, except that the last value controls destination.
1172 The value 0 selects the low bits, while 1 selects the high bits.
1174 Note: op_sel modifier affects 16-bit operands only. For 32-bit operands the value specified
1175 by op_sel must be 0.
1177 GFX9 and GFX10 only.
1179 ======================================== ============================================================
1181 ======================================== ============================================================
1182 op_sel:[{0..1},{0..1}] Select operand bits for instructions with 1 source operand.
1183 op_sel:[{0..1},{0..1},{0..1}] Select operand bits for instructions with 2 source operands.
1184 op_sel:[{0..1},{0..1},{0..1},{0..1}] Select operand bits for instructions with 3 source operands.
1185 ======================================== ============================================================
1187 Note: numeric values may be specified as either
1188 :ref:`integer numbers<amdgpu_synid_integer_number>` or
1189 :ref:`absolute expressions<amdgpu_synid_absolute_expression>`.
1198 .. _amdgpu_synid_clamp:
1203 Clamp meaning depends on instruction.
1205 For *v_cmp* instructions, clamp modifier indicates that the compare signals
1206 if a floating point exception occurs. By default, signaling is disabled.
1207 Not supported by GFX7.
1209 For integer operations, clamp modifier indicates that the result must be clamped
1210 to the largest and smallest representable value. By default, there is no clamping.
1211 Integer clamping is not supported by GFX7.
1213 For floating point operations, clamp modifier indicates that the result must be clamped
1214 to the range [0.0, 1.0]. By default, there is no clamping.
1216 Note: clamp modifier is applied after :ref:`output modifiers<amdgpu_synid_omod>` (if any).
1218 ======================================== ================================================
1220 ======================================== ================================================
1221 clamp Enables clamping (or signaling).
1222 ======================================== ================================================
1224 .. _amdgpu_synid_omod:
1229 Specifies if an output modifier must be applied to the result.
1230 By default, no output modifiers are applied.
1232 Note: output modifiers are applied before :ref:`clamping<amdgpu_synid_clamp>` (if any).
1234 Output modifiers are valid for f32 and f64 floating point results only.
1235 They must not be used with f16.
1237 Note: *v_cvt_f16_f32* is an exception. This instruction produces f16 result
1238 but accepts output modifiers.
1240 ======================================== ================================================
1242 ======================================== ================================================
1243 mul:2 Multiply the result by 2.
1244 mul:4 Multiply the result by 4.
1245 div:2 Multiply the result by 0.5.
1246 ======================================== ================================================
1248 Note: numeric values may be specified as either :ref:`integer numbers<amdgpu_synid_integer_number>` or
1249 :ref:`absolute expressions<amdgpu_synid_absolute_expression>`.
1256 mul:x // x must be equal to 2 or 4
1258 .. _amdgpu_synid_vop3_operand_modifiers:
1260 VOP3 Operand Modifiers
1261 ----------------------
1263 Operand modifiers are not used separately. They are applied to source operands.
1265 .. _amdgpu_synid_abs:
1270 Computes the absolute value of its operand. Must be applied before :ref:`neg<amdgpu_synid_neg>`
1271 (if any). Valid for floating point operands only.
1273 ======================================== ====================================================
1275 ======================================== ====================================================
1276 abs(<operand>) Get the absolute value of a floating-point operand.
1277 \|<operand>| The same as above (an SP3 syntax).
1278 ======================================== ====================================================
1280 Note: avoid using SP3 syntax with operands specified as expressions because the trailing '|'
1281 may be misinterpreted. Such operands should be enclosed into additional parentheses as shown
1291 \|(x|y)| // additional parentheses are required
1293 .. _amdgpu_synid_neg:
1298 Computes the negative value of its operand. Must be applied after :ref:`abs<amdgpu_synid_abs>`
1299 (if any). Valid for floating point operands only.
1301 ================== ====================================================
1303 ================== ====================================================
1304 neg(<operand>) Get the negative value of a floating-point operand.
1305 The operand may include an optional
1306 :ref:`abs<amdgpu_synid_abs>` modifier.
1307 -<operand> The same as above (an SP3 syntax).
1308 ================== ====================================================
1310 Note: SP3 syntax is supported with limitations because of a potential ambiguity.
1311 Currently it is allowed in the following cases:
1313 * Before a register.
1314 * Before an :ref:`abs<amdgpu_synid_abs>` modifier.
1315 * Before an SP3 :ref:`abs<amdgpu_synid_abs>` modifier.
1317 In all other cases "-" is handled as a part of an expression that follows the sign.
1323 // Operands with negate modifiers
1331 // Operands without negate modifiers
1338 This section describes modifiers of *regular* VOP3P instructions.
1340 *v_mad_mix_f32*, *v_mad_mixhi_f16* and *v_mad_mixlo_f16*
1341 instructions use these modifiers :ref:`in a special manner<amdgpu_synid_mad_mix>`.
1343 GFX9 and GFX10 only.
1345 .. _amdgpu_synid_op_sel:
1350 Selects the low [15:0] or high [31:16] operand bits as input to the operation
1351 which results in the lower-half of the destination.
1352 By default, low bits are used for all operands.
1354 The number of values specified by the *op_sel* modifier must match the number of source
1355 operands. First value controls src0, second value controls src1 and so on.
1357 The value 0 selects the low bits, while 1 selects the high bits.
1359 ================================= =============================================================
1361 ================================= =============================================================
1362 op_sel:[{0..1}] Select operand bits for instructions with 1 source operand.
1363 op_sel:[{0..1},{0..1}] Select operand bits for instructions with 2 source operands.
1364 op_sel:[{0..1},{0..1},{0..1}] Select operand bits for instructions with 3 source operands.
1365 ================================= =============================================================
1367 Note: numeric values may be specified as either
1368 :ref:`integer numbers<amdgpu_synid_integer_number>` or
1369 :ref:`absolute expressions<amdgpu_synid_absolute_expression>`.
1378 .. _amdgpu_synid_op_sel_hi:
1383 Selects the low [15:0] or high [31:16] operand bits as input to the operation
1384 which results in the upper-half of the destination.
1385 By default, high bits are used for all operands.
1387 The number of values specified by the *op_sel_hi* modifier must match the number of source
1388 operands. First value controls src0, second value controls src1 and so on.
1390 The value 0 selects the low bits, while 1 selects the high bits.
1392 =================================== =============================================================
1394 =================================== =============================================================
1395 op_sel_hi:[{0..1}] Select operand bits for instructions with 1 source operand.
1396 op_sel_hi:[{0..1},{0..1}] Select operand bits for instructions with 2 source operands.
1397 op_sel_hi:[{0..1},{0..1},{0..1}] Select operand bits for instructions with 3 source operands.
1398 =================================== =============================================================
1400 Note: numeric values may be specified as either
1401 :ref:`integer numbers<amdgpu_synid_integer_number>` or
1402 :ref:`absolute expressions<amdgpu_synid_absolute_expression>`.
1411 .. _amdgpu_synid_neg_lo:
1416 Specifies whether to change sign of operand values selected by
1417 :ref:`op_sel<amdgpu_synid_op_sel>`. These values are then used
1418 as input to the operation which results in the upper-half of the destination.
1420 The number of values specified by this modifier must match the number of source
1421 operands. First value controls src0, second value controls src1 and so on.
1423 The value 0 indicates that the corresponding operand value is used unmodified,
1424 the value 1 indicates that negative value of the operand must be used.
1426 By default, operand values are used unmodified.
1428 This modifier is valid for floating point operands only.
1430 ================================ ==================================================================
1432 ================================ ==================================================================
1433 neg_lo:[{0..1}] Select affected operands for instructions with 1 source operand.
1434 neg_lo:[{0..1},{0..1}] Select affected operands for instructions with 2 source operands.
1435 neg_lo:[{0..1},{0..1},{0..1}] Select affected operands for instructions with 3 source operands.
1436 ================================ ==================================================================
1438 Note: numeric values may be specified as either
1439 :ref:`integer numbers<amdgpu_synid_integer_number>` or
1440 :ref:`absolute expressions<amdgpu_synid_absolute_expression>`.
1449 .. _amdgpu_synid_neg_hi:
1454 Specifies whether to change sign of operand values selected by
1455 :ref:`op_sel_hi<amdgpu_synid_op_sel_hi>`. These values are then used
1456 as input to the operation which results in the upper-half of the destination.
1458 The number of values specified by this modifier must match the number of source
1459 operands. First value controls src0, second value controls src1 and so on.
1461 The value 0 indicates that the corresponding operand value is used unmodified,
1462 the value 1 indicates that negative value of the operand must be used.
1464 By default, operand values are used unmodified.
1466 This modifier is valid for floating point operands only.
1468 =============================== ==================================================================
1470 =============================== ==================================================================
1471 neg_hi:[{0..1}] Select affected operands for instructions with 1 source operand.
1472 neg_hi:[{0..1},{0..1}] Select affected operands for instructions with 2 source operands.
1473 neg_hi:[{0..1},{0..1},{0..1}] Select affected operands for instructions with 3 source operands.
1474 =============================== ==================================================================
1476 Note: numeric values may be specified as either
1477 :ref:`integer numbers<amdgpu_synid_integer_number>` or
1478 :ref:`absolute expressions<amdgpu_synid_absolute_expression>`.
1490 See a description :ref:`here<amdgpu_synid_clamp>`.
1492 .. _amdgpu_synid_mad_mix:
1494 VOP3P V_MAD_MIX Modifiers
1495 -------------------------
1497 *v_mad_mix_f32*, *v_mad_mixhi_f16* and *v_mad_mixlo_f16* instructions
1498 use *op_sel* and *op_sel_hi* modifiers
1499 in a manner different from *regular* VOP3P instructions.
1501 See a description below.
1503 GFX9 and GFX10 only.
1505 .. _amdgpu_synid_mad_mix_op_sel:
1510 This operand has meaning only for 16-bit source operands as indicated by
1511 :ref:`m_op_sel_hi<amdgpu_synid_mad_mix_op_sel_hi>`.
1512 It specifies to select either the low [15:0] or high [31:16] operand bits
1513 as input to the operation.
1515 The number of values specified by the *op_sel* modifier must match the number of source
1516 operands. First value controls src0, second value controls src1 and so on.
1518 The value 0 indicates the low bits, the value 1 indicates the high 16 bits.
1520 By default, low bits are used for all operands.
1522 =============================== ================================================
1524 =============================== ================================================
1525 op_sel:[{0..1},{0..1},{0..1}] Select location of each 16-bit source operand.
1526 =============================== ================================================
1528 Note: numeric values may be specified as either
1529 :ref:`integer numbers<amdgpu_synid_integer_number>` or
1530 :ref:`absolute expressions<amdgpu_synid_absolute_expression>`.
1538 .. _amdgpu_synid_mad_mix_op_sel_hi:
1543 Selects the size of source operands: either 32 bits or 16 bits.
1544 By default, 32 bits are used for all source operands.
1546 The number of values specified by the *op_sel_hi* modifier must match the number of source
1547 operands. First value controls src0, second value controls src1 and so on.
1549 The value 0 indicates 32 bits, the value 1 indicates 16 bits.
1551 The location of 16 bits in the operand may be specified by
1552 :ref:`m_op_sel<amdgpu_synid_mad_mix_op_sel>`.
1554 ======================================== ====================================
1556 ======================================== ====================================
1557 op_sel_hi:[{0..1},{0..1},{0..1}] Select size of each source operand.
1558 ======================================== ====================================
1560 Note: numeric values may be specified as either
1561 :ref:`integer numbers<amdgpu_synid_integer_number>` or
1562 :ref:`absolute expressions<amdgpu_synid_absolute_expression>`.
1573 See a description :ref:`here<amdgpu_synid_abs>`.
1578 See a description :ref:`here<amdgpu_synid_neg>`.
1583 See a description :ref:`here<amdgpu_synid_clamp>`.