6aebb057619c89abe01401dd215ffa22ccf17f27
2 * Copyright (c) 2013 Rob Clark <robdclark@gmail.com>
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
27 #define PACKED __attribute__((__packed__))
32 /* size of largest OPC field of all the instruction categories: */
35 #define _OPC(cat, opc) (((cat) << NOPC_BITS) | opc)
41 OPC_JUMP
= _OPC(0, 2),
42 OPC_CALL
= _OPC(0, 3),
44 OPC_KILL
= _OPC(0, 5),
46 OPC_EMIT
= _OPC(0, 7),
48 OPC_CHMASK
= _OPC(0, 9),
49 OPC_CHSH
= _OPC(0, 10),
50 OPC_FLOW_REV
= _OPC(0, 11),
56 OPC_ADD_F
= _OPC(2, 0),
57 OPC_MIN_F
= _OPC(2, 1),
58 OPC_MAX_F
= _OPC(2, 2),
59 OPC_MUL_F
= _OPC(2, 3),
60 OPC_SIGN_F
= _OPC(2, 4),
61 OPC_CMPS_F
= _OPC(2, 5),
62 OPC_ABSNEG_F
= _OPC(2, 6),
63 OPC_CMPV_F
= _OPC(2, 7),
65 OPC_FLOOR_F
= _OPC(2, 9),
66 OPC_CEIL_F
= _OPC(2, 10),
67 OPC_RNDNE_F
= _OPC(2, 11),
68 OPC_RNDAZ_F
= _OPC(2, 12),
69 OPC_TRUNC_F
= _OPC(2, 13),
71 OPC_ADD_U
= _OPC(2, 16),
72 OPC_ADD_S
= _OPC(2, 17),
73 OPC_SUB_U
= _OPC(2, 18),
74 OPC_SUB_S
= _OPC(2, 19),
75 OPC_CMPS_U
= _OPC(2, 20),
76 OPC_CMPS_S
= _OPC(2, 21),
77 OPC_MIN_U
= _OPC(2, 22),
78 OPC_MIN_S
= _OPC(2, 23),
79 OPC_MAX_U
= _OPC(2, 24),
80 OPC_MAX_S
= _OPC(2, 25),
81 OPC_ABSNEG_S
= _OPC(2, 26),
83 OPC_AND_B
= _OPC(2, 28),
84 OPC_OR_B
= _OPC(2, 29),
85 OPC_NOT_B
= _OPC(2, 30),
86 OPC_XOR_B
= _OPC(2, 31),
88 OPC_CMPV_U
= _OPC(2, 33),
89 OPC_CMPV_S
= _OPC(2, 34),
91 OPC_MUL_U
= _OPC(2, 48),
92 OPC_MUL_S
= _OPC(2, 49),
93 OPC_MULL_U
= _OPC(2, 50),
94 OPC_BFREV_B
= _OPC(2, 51),
95 OPC_CLZ_S
= _OPC(2, 52),
96 OPC_CLZ_B
= _OPC(2, 53),
97 OPC_SHL_B
= _OPC(2, 54),
98 OPC_SHR_B
= _OPC(2, 55),
99 OPC_ASHR_B
= _OPC(2, 56),
100 OPC_BARY_F
= _OPC(2, 57),
101 OPC_MGEN_B
= _OPC(2, 58),
102 OPC_GETBIT_B
= _OPC(2, 59),
103 OPC_SETRM
= _OPC(2, 60),
104 OPC_CBITS_B
= _OPC(2, 61),
105 OPC_SHB
= _OPC(2, 62),
106 OPC_MSAD
= _OPC(2, 63),
109 OPC_MAD_U16
= _OPC(3, 0),
110 OPC_MADSH_U16
= _OPC(3, 1),
111 OPC_MAD_S16
= _OPC(3, 2),
112 OPC_MADSH_M16
= _OPC(3, 3), /* should this be .s16? */
113 OPC_MAD_U24
= _OPC(3, 4),
114 OPC_MAD_S24
= _OPC(3, 5),
115 OPC_MAD_F16
= _OPC(3, 6),
116 OPC_MAD_F32
= _OPC(3, 7),
117 OPC_SEL_B16
= _OPC(3, 8),
118 OPC_SEL_B32
= _OPC(3, 9),
119 OPC_SEL_S16
= _OPC(3, 10),
120 OPC_SEL_S32
= _OPC(3, 11),
121 OPC_SEL_F16
= _OPC(3, 12),
122 OPC_SEL_F32
= _OPC(3, 13),
123 OPC_SAD_S16
= _OPC(3, 14),
124 OPC_SAD_S32
= _OPC(3, 15),
127 OPC_RCP
= _OPC(4, 0),
128 OPC_RSQ
= _OPC(4, 1),
129 OPC_LOG2
= _OPC(4, 2),
130 OPC_EXP2
= _OPC(4, 3),
131 OPC_SIN
= _OPC(4, 4),
132 OPC_COS
= _OPC(4, 5),
133 OPC_SQRT
= _OPC(4, 6),
137 OPC_ISAM
= _OPC(5, 0),
138 OPC_ISAML
= _OPC(5, 1),
139 OPC_ISAMM
= _OPC(5, 2),
140 OPC_SAM
= _OPC(5, 3),
141 OPC_SAMB
= _OPC(5, 4),
142 OPC_SAML
= _OPC(5, 5),
143 OPC_SAMGQ
= _OPC(5, 6),
144 OPC_GETLOD
= _OPC(5, 7),
145 OPC_CONV
= _OPC(5, 8),
146 OPC_CONVM
= _OPC(5, 9),
147 OPC_GETSIZE
= _OPC(5, 10),
148 OPC_GETBUF
= _OPC(5, 11),
149 OPC_GETPOS
= _OPC(5, 12),
150 OPC_GETINFO
= _OPC(5, 13),
151 OPC_DSX
= _OPC(5, 14),
152 OPC_DSY
= _OPC(5, 15),
153 OPC_GATHER4R
= _OPC(5, 16),
154 OPC_GATHER4G
= _OPC(5, 17),
155 OPC_GATHER4B
= _OPC(5, 18),
156 OPC_GATHER4A
= _OPC(5, 19),
157 OPC_SAMGP0
= _OPC(5, 20),
158 OPC_SAMGP1
= _OPC(5, 21),
159 OPC_SAMGP2
= _OPC(5, 22),
160 OPC_SAMGP3
= _OPC(5, 23),
161 OPC_DSXPP_1
= _OPC(5, 24),
162 OPC_DSYPP_1
= _OPC(5, 25),
163 OPC_RGETPOS
= _OPC(5, 26),
164 OPC_RGETINFO
= _OPC(5, 27),
167 OPC_LDG
= _OPC(6, 0), /* load-global */
168 OPC_LDL
= _OPC(6, 1),
169 OPC_LDP
= _OPC(6, 2),
170 OPC_STG
= _OPC(6, 3), /* store-global */
171 OPC_STL
= _OPC(6, 4),
172 OPC_STP
= _OPC(6, 5),
173 OPC_STI
= _OPC(6, 6),
174 OPC_G2L
= _OPC(6, 7),
175 OPC_L2G
= _OPC(6, 8),
176 OPC_PREFETCH
= _OPC(6, 9),
177 OPC_LDLW
= _OPC(6, 10),
178 OPC_STLW
= _OPC(6, 11),
179 OPC_RESFMT
= _OPC(6, 14),
180 OPC_RESINFO
= _OPC(6, 15),
181 OPC_ATOMIC_ADD
= _OPC(6, 16),
182 OPC_ATOMIC_SUB
= _OPC(6, 17),
183 OPC_ATOMIC_XCHG
= _OPC(6, 18),
184 OPC_ATOMIC_INC
= _OPC(6, 19),
185 OPC_ATOMIC_DEC
= _OPC(6, 20),
186 OPC_ATOMIC_CMPXCHG
= _OPC(6, 21),
187 OPC_ATOMIC_MIN
= _OPC(6, 22),
188 OPC_ATOMIC_MAX
= _OPC(6, 23),
189 OPC_ATOMIC_AND
= _OPC(6, 24),
190 OPC_ATOMIC_OR
= _OPC(6, 25),
191 OPC_ATOMIC_XOR
= _OPC(6, 26),
192 OPC_LDGB
= _OPC(6, 27),
193 OPC_STGB
= _OPC(6, 28),
194 OPC_STIB
= _OPC(6, 29),
195 OPC_LDC
= _OPC(6, 30),
196 OPC_LDLV
= _OPC(6, 31),
199 OPC_BAR
= _OPC(7, 0),
200 OPC_FENCE
= _OPC(7, 1),
202 /* meta instructions (category -1): */
203 /* placeholder instr to mark shader inputs: */
204 OPC_META_INPUT
= _OPC(-1, 0),
205 OPC_META_PHI
= _OPC(-1, 1),
206 /* The "fan-in" and "fan-out" instructions are used for keeping
207 * track of instructions that write to multiple dst registers
208 * (fan-out) like texture sample instructions, or read multiple
209 * consecutive scalar registers (fan-in) (bary.f, texture samp)
211 OPC_META_FO
= _OPC(-1, 2),
212 OPC_META_FI
= _OPC(-1, 3),
216 #define opc_cat(opc) ((int)((opc) >> NOPC_BITS))
217 #define opc_op(opc) ((unsigned)((opc) & ((1 << NOPC_BITS) - 1)))
227 TYPE_S8
= 7, // XXX I assume?
230 static inline uint32_t type_size(type_t type
)
245 assert(0); /* invalid type */
250 static inline int type_float(type_t type
)
252 return (type
== TYPE_F32
) || (type
== TYPE_F16
);
255 static inline int type_uint(type_t type
)
257 return (type
== TYPE_U32
) || (type
== TYPE_U16
) || (type
== TYPE_U8
);
260 static inline int type_sint(type_t type
)
262 return (type
== TYPE_S32
) || (type
== TYPE_S16
) || (type
== TYPE_S8
);
265 typedef union PACKED
{
266 /* normal gpr or const src register: */
271 /* for immediate val: */
272 int32_t iim_val
: 11;
273 /* to make compiler happy: */
275 uint32_t dummy10
: 10;
276 int32_t idummy10
: 10;
277 uint32_t dummy11
: 11;
278 uint32_t dummy12
: 12;
279 uint32_t dummy13
: 13;
283 /* special registers: */
284 #define REG_A0 61 /* address register */
285 #define REG_P0 62 /* predicate register */
287 static inline int reg_special(reg_t reg
)
289 return (reg
.num
== REG_A0
) || (reg
.num
== REG_P0
);
292 typedef struct PACKED
{
297 uint32_t dummy1
: 16;
301 uint32_t dummy1
: 12;
317 uint32_t jmp_tgt
: 1;
319 uint32_t opc_cat
: 3;
322 typedef struct PACKED
{
325 /* for normal src register: */
328 /* at least low bit of pad must be zero or it will
329 * look like a address relative src
333 /* for address relative: */
336 uint32_t src_rel_c
: 1;
337 uint32_t src_rel
: 1;
338 uint32_t unknown
: 20;
352 uint32_t dst_type
: 3;
353 uint32_t dst_rel
: 1;
354 uint32_t src_type
: 3;
358 uint32_t pos_inf
: 1;
359 uint32_t must_be_0
: 2;
360 uint32_t jmp_tgt
: 1;
362 uint32_t opc_cat
: 3;
365 typedef struct PACKED
{
370 uint32_t must_be_zero1
: 2;
371 uint32_t src1_im
: 1; /* immediate */
372 uint32_t src1_neg
: 1; /* negate */
373 uint32_t src1_abs
: 1; /* absolute value */
377 uint32_t src1_c
: 1; /* relative-const */
378 uint32_t src1_rel
: 1; /* relative address */
379 uint32_t must_be_zero
: 1;
384 uint32_t src1_c
: 1; /* const */
392 uint32_t must_be_zero2
: 2;
393 uint32_t src2_im
: 1; /* immediate */
394 uint32_t src2_neg
: 1; /* negate */
395 uint32_t src2_abs
: 1; /* absolute value */
399 uint32_t src2_c
: 1; /* relative-const */
400 uint32_t src2_rel
: 1; /* relative address */
401 uint32_t must_be_zero
: 1;
406 uint32_t src2_c
: 1; /* const */
417 uint32_t ul
: 1; /* dunno */
418 uint32_t dst_half
: 1; /* or widen/narrow.. ie. dst hrN <-> rN */
422 uint32_t full
: 1; /* not half */
424 uint32_t jmp_tgt
: 1;
426 uint32_t opc_cat
: 3;
429 typedef struct PACKED
{
434 uint32_t must_be_zero1
: 2;
436 uint32_t src1_neg
: 1;
442 uint32_t src1_rel
: 1;
443 uint32_t must_be_zero
: 1;
456 uint32_t must_be_zero2
: 2;
458 uint32_t src2_neg
: 1;
459 uint32_t src3_neg
: 1;
464 uint32_t src3_rel
: 1;
465 uint32_t must_be_zero
: 1;
482 uint32_t dst_half
: 1; /* or widen/narrow.. ie. dst hrN <-> rN */
485 uint32_t jmp_tgt
: 1;
487 uint32_t opc_cat
: 3;
490 static inline bool instr_cat3_full(instr_cat3_t
*cat3
)
492 switch (_OPC(3, cat3
->opc
)) {
500 case OPC_SAD_S32
: // really??
507 typedef struct PACKED
{
512 uint32_t must_be_zero1
: 2;
513 uint32_t src_im
: 1; /* immediate */
514 uint32_t src_neg
: 1; /* negate */
515 uint32_t src_abs
: 1; /* absolute value */
519 uint32_t src_c
: 1; /* relative-const */
520 uint32_t src_rel
: 1; /* relative address */
521 uint32_t must_be_zero
: 1;
526 uint32_t src_c
: 1; /* const */
530 uint32_t dummy1
: 16; /* seem to be ignored */
539 uint32_t dst_half
: 1; /* or widen/narrow.. ie. dst hrN <-> rN */
540 uint32_t dummy2
: 5; /* seem to be ignored */
541 uint32_t full
: 1; /* not half */
543 uint32_t jmp_tgt
: 1;
545 uint32_t opc_cat
: 3;
548 typedef struct PACKED
{
553 uint32_t full
: 1; /* not half */
556 uint32_t dummy1
: 4; /* seem to be ignored */
562 uint32_t full
: 1; /* not half */
569 /* same in either case: */
570 // XXX I think, confirm this
572 uint32_t full
: 1; /* not half */
580 uint32_t wrmask
: 4; /* write-mask */
582 uint32_t dummy2
: 1; /* seems to be ignored */
587 uint32_t is_s2en
: 1;
592 uint32_t jmp_tgt
: 1;
594 uint32_t opc_cat
: 3;
597 /* dword0 encoding for src_off: [src1 + off], src2: */
598 typedef struct PACKED
{
600 uint32_t mustbe1
: 1;
603 uint32_t src1_im
: 1;
604 uint32_t src2_im
: 1;
611 /* dword0 encoding for !src_off: [src1], src2 */
612 typedef struct PACKED
{
614 uint32_t mustbe0
: 1;
616 uint32_t ignore0
: 8;
617 uint32_t src1_im
: 1;
618 uint32_t src2_im
: 1;
625 /* dword1 encoding for dst_off: */
626 typedef struct PACKED
{
630 /* note: there is some weird stuff going on where sometimes
631 * cat6->a.off is involved.. but that seems like a bug in
632 * the blob, since it is used even if !cat6->src_off
633 * It would make sense for there to be some more bits to
634 * bring us to 11 bits worth of offset, but not sure..
637 uint32_t mustbe1
: 1;
642 /* dword1 encoding for !dst_off: */
643 typedef struct PACKED
{
648 uint32_t mustbe0
: 1;
653 /* ldgb and atomics..
655 * ldgb: pad0=0, pad3=1
656 * atomic .g: pad0=1, pad3=1
659 typedef struct PACKED
{
665 uint32_t type_size
: 2;
667 uint32_t src1_im
: 1;
668 uint32_t src2_im
: 1;
673 uint32_t mustbe0
: 1;
674 uint32_t src_ssbo
: 8;
675 uint32_t pad2
: 3; // type
678 uint32_t pad4
: 10; // opc/jmp_tgt/sync/opc_cat
681 /* stgb, pad0=0, pad3=2
683 typedef struct PACKED
{
685 uint32_t mustbe1
: 1; // ???
689 uint32_t type_size
: 2;
691 uint32_t src2_im
: 1;
696 uint32_t src3_im
: 1;
697 uint32_t dst_ssbo
: 8;
698 uint32_t pad2
: 3; // type
700 uint32_t pad4
: 10; // opc/jmp_tgt/sync/opc_cat
703 typedef union PACKED
{
708 instr_cat6ldgb_t ldgb
;
709 instr_cat6stgb_t stgb
;
712 uint32_t src_off
: 1;
717 uint32_t dst_off
: 1;
720 uint32_t g
: 1; /* or in some cases it means dst immed */
723 uint32_t jmp_tgt
: 1;
725 uint32_t opc_cat
: 3;
729 typedef struct PACKED
{
735 uint32_t ss
: 1; /* maybe in the encoding, but blob only uses (sy) */
737 uint32_t w
: 1; /* write */
738 uint32_t r
: 1; /* read */
739 uint32_t l
: 1; /* local */
740 uint32_t g
: 1; /* global */
741 uint32_t opc
: 4; /* presumed, but only a couple known OPCs */
742 uint32_t jmp_tgt
: 1; /* (jp) */
743 uint32_t sync
: 1; /* (sy) */
744 uint32_t opc_cat
: 3;
747 typedef union PACKED
{
762 uint32_t ss
: 1; /* cat1-cat4 (cat0??) and cat7 (?) */
763 uint32_t ul
: 1; /* cat2-cat4 (and cat1 in blob.. which may be bug??) */
765 uint32_t jmp_tgt
: 1;
767 uint32_t opc_cat
: 3;
772 static inline uint32_t instr_repeat(instr_t
*instr
)
774 switch (instr
->opc_cat
) {
775 case 0: return instr
->cat0
.repeat
;
776 case 1: return instr
->cat1
.repeat
;
777 case 2: return instr
->cat2
.repeat
;
778 case 3: return instr
->cat3
.repeat
;
779 case 4: return instr
->cat4
.repeat
;
784 static inline bool instr_sat(instr_t
*instr
)
786 switch (instr
->opc_cat
) {
787 case 2: return instr
->cat2
.sat
;
788 case 3: return instr
->cat3
.sat
;
789 case 4: return instr
->cat4
.sat
;
790 default: return false;
794 static inline uint32_t instr_opc(instr_t
*instr
)
796 switch (instr
->opc_cat
) {
797 case 0: return instr
->cat0
.opc
;
799 case 2: return instr
->cat2
.opc
;
800 case 3: return instr
->cat3
.opc
;
801 case 4: return instr
->cat4
.opc
;
802 case 5: return instr
->cat5
.opc
;
803 case 6: return instr
->cat6
.opc
;
804 case 7: return instr
->cat7
.opc
;
809 static inline bool is_mad(opc_t opc
)
824 static inline bool is_madsh(opc_t opc
)
835 static inline bool is_atomic(opc_t opc
)
840 case OPC_ATOMIC_XCHG
:
843 case OPC_ATOMIC_CMPXCHG
:
855 static inline bool is_ssbo(opc_t opc
)
869 #endif /* INSTR_A3XX_H_ */