2 * Copyright (C) 2019 Connor Abbott <cwabbott0@gmail.com>
3 * Copyright (C) 2019 Lyude Paul <thatslyude@gmail.com>
4 * Copyright (C) 2019 Ryan Houdek <Sonicadvance1@gmail.com>
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
13 * The above copyright notice and this permission notice (including the next
14 * paragraph) shall be included in all copies or substantial portions of the
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
21 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
22 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
34 #include "disassemble.h"
35 #include "util/macros.h"
37 // return bits (high, lo]
38 static uint64_t bits(uint32_t word
, unsigned lo
, unsigned high
)
42 return (word
& ((1 << high
) - 1)) >> lo
;
45 // each of these structs represents an instruction that's dispatched in one
46 // cycle. Note that these instructions are packed in funny ways within the
47 // clause, hence the need for a separate struct.
48 struct bifrost_alu_inst
{
54 static unsigned get_reg0(struct bifrost_regs regs
)
57 return regs
.reg0
| ((regs
.reg1
& 0x1) << 5);
59 return regs
.reg0
<= regs
.reg1
? regs
.reg0
: 63 - regs
.reg0
;
62 static unsigned get_reg1(struct bifrost_regs regs
)
64 return regs
.reg0
<= regs
.reg1
? regs
.reg1
: 63 - regs
.reg1
;
67 // this represents the decoded version of the ctrl register field.
68 struct bifrost_reg_ctrl
{
72 enum bifrost_reg_write_unit fma_write_unit
;
73 enum bifrost_reg_write_unit add_write_unit
;
99 enum fma_src_type src_type
;
114 ADD_TEX_COMPACT
, // texture instruction with embedded sampler
115 ADD_TEX
, // texture instruction with sampler/etc. in uniform port
126 enum add_src_type src_type
;
130 void dump_header(FILE *fp
, struct bifrost_header header
, bool verbose
);
131 void dump_instr(FILE *fp
, const struct bifrost_alu_inst
*instr
,
132 struct bifrost_regs next_regs
, uint64_t *consts
,
133 unsigned data_reg
, unsigned offset
, bool verbose
);
134 bool dump_clause(FILE *fp
, uint32_t *words
, unsigned *size
, unsigned offset
, bool verbose
);
136 void dump_header(FILE *fp
, struct bifrost_header header
, bool verbose
)
138 if (header
.clause_type
!= 0) {
139 fprintf(fp
, "id(%du) ", header
.scoreboard_index
);
142 if (header
.scoreboard_deps
!= 0) {
143 fprintf(fp
, "next-wait(");
145 for (unsigned i
= 0; i
< 8; i
++) {
146 if (header
.scoreboard_deps
& (1 << i
)) {
150 fprintf(fp
, "%d", i
);
157 if (header
.datareg_writebarrier
)
158 fprintf(fp
, "data-reg-barrier ");
160 if (!header
.no_end_of_shader
)
163 if (!header
.back_to_back
) {
165 if (header
.branch_cond
)
166 fprintf(fp
, "branch-cond ");
168 fprintf(fp
, "branch-uncond ");
171 if (header
.elide_writes
)
174 if (header
.suppress_inf
)
175 fprintf(fp
, "suppress-inf ");
176 if (header
.suppress_nan
)
177 fprintf(fp
, "suppress-nan ");
180 fprintf(fp
, "unk0 ");
182 fprintf(fp
, "unk1 ");
184 fprintf(fp
, "unk2 ");
186 fprintf(fp
, "unk3 ");
188 fprintf(fp
, "unk4 ");
193 fprintf(fp
, "# clause type %d, next clause type %d\n",
194 header
.clause_type
, header
.next_clause_type
);
198 static struct bifrost_reg_ctrl
DecodeRegCtrl(FILE *fp
, struct bifrost_regs regs
)
200 struct bifrost_reg_ctrl decoded
= {};
202 if (regs
.ctrl
== 0) {
203 ctrl
= regs
.reg1
>> 2;
204 decoded
.read_reg0
= !(regs
.reg1
& 0x2);
205 decoded
.read_reg1
= false;
208 decoded
.read_reg0
= decoded
.read_reg1
= true;
212 decoded
.fma_write_unit
= REG_WRITE_TWO
;
216 decoded
.fma_write_unit
= REG_WRITE_TWO
;
217 decoded
.read_reg3
= true;
220 decoded
.read_reg3
= true;
223 decoded
.add_write_unit
= REG_WRITE_TWO
;
226 decoded
.add_write_unit
= REG_WRITE_TWO
;
227 decoded
.read_reg3
= true;
230 decoded
.clause_start
= true;
233 decoded
.fma_write_unit
= REG_WRITE_TWO
;
234 decoded
.clause_start
= true;
239 decoded
.read_reg3
= true;
240 decoded
.clause_start
= true;
243 decoded
.add_write_unit
= REG_WRITE_TWO
;
244 decoded
.clause_start
= true;
249 decoded
.fma_write_unit
= REG_WRITE_THREE
;
250 decoded
.add_write_unit
= REG_WRITE_TWO
;
253 fprintf(fp
, "# unknown reg ctrl %d\n", ctrl
);
259 // Pass in the add_write_unit or fma_write_unit, and this returns which register
260 // the ADD/FMA units are writing to
261 static unsigned GetRegToWrite(enum bifrost_reg_write_unit unit
, struct bifrost_regs regs
)
266 case REG_WRITE_THREE
:
268 default: /* REG_WRITE_NONE */
274 static void dump_regs(FILE *fp
, struct bifrost_regs srcs
)
276 struct bifrost_reg_ctrl ctrl
= DecodeRegCtrl(fp
, srcs
);
279 fprintf(fp
, "port 0: R%d ", get_reg0(srcs
));
281 fprintf(fp
, "port 1: R%d ", get_reg1(srcs
));
283 if (ctrl
.fma_write_unit
== REG_WRITE_TWO
)
284 fprintf(fp
, "port 2: R%d (write FMA) ", srcs
.reg2
);
285 else if (ctrl
.add_write_unit
== REG_WRITE_TWO
)
286 fprintf(fp
, "port 2: R%d (write ADD) ", srcs
.reg2
);
288 if (ctrl
.fma_write_unit
== REG_WRITE_THREE
)
289 fprintf(fp
, "port 3: R%d (write FMA) ", srcs
.reg3
);
290 else if (ctrl
.add_write_unit
== REG_WRITE_THREE
)
291 fprintf(fp
, "port 3: R%d (write ADD) ", srcs
.reg3
);
292 else if (ctrl
.read_reg3
)
293 fprintf(fp
, "port 3: R%d (read) ", srcs
.reg3
);
295 if (srcs
.uniform_const
) {
296 if (srcs
.uniform_const
& 0x80) {
297 fprintf(fp
, "uniform: U%d", (srcs
.uniform_const
& 0x7f) * 2);
303 static void dump_const_imm(FILE *fp
, uint32_t imm
)
310 fprintf(fp
, "0x%08x /* %f */", imm
, fi
.f
);
313 static uint64_t get_const(uint64_t *consts
, struct bifrost_regs srcs
)
315 unsigned low_bits
= srcs
.uniform_const
& 0xf;
317 switch (srcs
.uniform_const
>> 4) {
340 return imm
| low_bits
;
343 static void dump_uniform_const_src(FILE *fp
, struct bifrost_regs srcs
, uint64_t *consts
, bool high32
)
345 if (srcs
.uniform_const
& 0x80) {
346 unsigned uniform
= (srcs
.uniform_const
& 0x7f) * 2;
347 fprintf(fp
, "U%d", uniform
+ (high32
? 1 : 0));
348 } else if (srcs
.uniform_const
>= 0x20) {
349 uint64_t imm
= get_const(consts
, srcs
);
351 dump_const_imm(fp
, imm
>> 32);
353 dump_const_imm(fp
, imm
);
355 switch (srcs
.uniform_const
) {
360 fprintf(fp
, "atest-data");
363 fprintf(fp
, "sample-ptr");
373 fprintf(fp
, "blend-descriptor%u", (unsigned) srcs
.uniform_const
- 8);
376 fprintf(fp
, "unkConst%u", (unsigned) srcs
.uniform_const
);
387 static void dump_src(FILE *fp
, unsigned src
, struct bifrost_regs srcs
, uint64_t *consts
, bool isFMA
)
391 fprintf(fp
, "R%d", get_reg0(srcs
));
394 fprintf(fp
, "R%d", get_reg1(srcs
));
397 fprintf(fp
, "R%d", srcs
.reg3
);
403 fprintf(fp
, "T"); // i.e. the output of FMA this cycle
406 dump_uniform_const_src(fp
, srcs
, consts
, false);
409 dump_uniform_const_src(fp
, srcs
, consts
, true);
420 static void dump_output_mod(FILE *fp
, unsigned mod
)
426 fprintf(fp
, ".clamp_0_inf");
427 break; // max(out, 0)
428 case BIFROST_SAT_SIGNED
:
429 fprintf(fp
, ".clamp_m1_1");
430 break; // clamp(out, -1, 1)
432 fprintf(fp
, ".clamp_0_1");
433 break; // clamp(out, 0, 1)
439 static void dump_minmax_mode(FILE *fp
, unsigned mod
)
443 /* Same as fmax() and fmin() -- return the other number if any
444 * number is NaN. Also always return +0 if one argument is +0 and
449 /* Instead of never returning a NaN, always return one. The
450 * "greater"/"lesser" NaN is always returned, first by checking the
451 * sign and then the mantissa bits.
453 fprintf(fp
, ".nan_wins");
456 /* For max, implement src0 > src1 ? src0 : src1
457 * For min, implement src0 < src1 ? src0 : src1
459 * This includes handling NaN's and signedness of 0 differently
460 * from above, since +0 and -0 compare equal and comparisons always
461 * return false for NaN's. As a result, this mode is *not*
464 fprintf(fp
, ".src1_wins");
467 /* For max, implement src0 < src1 ? src1 : src0
468 * For min, implement src0 > src1 ? src1 : src0
470 fprintf(fp
, ".src0_wins");
477 static void dump_round_mode(FILE *fp
, unsigned mod
)
481 /* roundTiesToEven, the IEEE default. */
484 /* roundTowardPositive in the IEEE spec. */
485 fprintf(fp
, ".round_pos");
488 /* roundTowardNegative in the IEEE spec. */
489 fprintf(fp
, ".round_neg");
492 /* roundTowardZero in the IEEE spec. */
493 fprintf(fp
, ".round_zero");
501 csel_cond_name(enum bifrost_csel_cond cond
)
504 case BIFROST_FEQ_F
: return "feq.f";
505 case BIFROST_FGT_F
: return "fgt.f";
506 case BIFROST_FGE_F
: return "fge.f";
507 case BIFROST_IEQ_F
: return "ieq.f";
508 case BIFROST_IGT_I
: return "igt.i";
509 case BIFROST_IGE_I
: return "uge.i";
510 case BIFROST_UGT_I
: return "ugt.i";
511 case BIFROST_UGE_I
: return "uge.i";
512 default: return "invalid";
516 static const struct fma_op_info FMAOpInfos
[] = {
517 { false, 0x00000, "FMA.f32", FMA_FMA
},
518 { false, 0x40000, "MAX.f32", FMA_FMINMAX
},
519 { false, 0x44000, "MIN.f32", FMA_FMINMAX
},
520 { false, 0x48000, "FCMP.GL", FMA_FCMP
},
521 { false, 0x4c000, "FCMP.D3D", FMA_FCMP
},
522 { false, 0x4ff98, "ADD.i32", FMA_TWO_SRC
},
523 { false, 0x4ffd8, "SUB.i32", FMA_TWO_SRC
},
524 { false, 0x4fff0, "SUBB.i32", FMA_TWO_SRC
},
525 { false, 0x50000, "FMA_MSCALE", FMA_FMA_MSCALE
},
526 { false, 0x58000, "ADD.f32", FMA_FADD
},
527 { false, 0x5c000, "CSEL4", FMA_CSEL4
},
528 { false, 0x5d8d0, "ICMP.D3D.GT.v2i16", FMA_TWO_SRC
},
529 { false, 0x5d9d0, "UCMP.D3D.GT.v2i16", FMA_TWO_SRC
},
530 { false, 0x5dad0, "ICMP.D3D.GE.v2i16", FMA_TWO_SRC
},
531 { false, 0x5dbd0, "UCMP.D3D.GE.v2i16", FMA_TWO_SRC
},
532 { false, 0x5dcd0, "ICMP.D3D.EQ.v2i16", FMA_TWO_SRC
},
533 { false, 0x5de40, "ICMP.GL.GT.i32", FMA_TWO_SRC
}, // src0 > src1 ? 1 : 0
534 { false, 0x5de48, "ICMP.GL.GE.i32", FMA_TWO_SRC
},
535 { false, 0x5de50, "UCMP.GL.GT.i32", FMA_TWO_SRC
},
536 { false, 0x5de58, "UCMP.GL.GE.i32", FMA_TWO_SRC
},
537 { false, 0x5de60, "ICMP.GL.EQ.i32", FMA_TWO_SRC
},
538 { false, 0x5dec0, "ICMP.D3D.GT.i32", FMA_TWO_SRC
}, // src0 > src1 ? ~0 : 0
539 { false, 0x5dec8, "ICMP.D3D.GE.i32", FMA_TWO_SRC
},
540 { false, 0x5ded0, "UCMP.D3D.GT.i32", FMA_TWO_SRC
},
541 { false, 0x5ded8, "UCMP.D3D.GE.i32", FMA_TWO_SRC
},
542 { false, 0x5dee0, "ICMP.D3D.EQ.i32", FMA_TWO_SRC
},
543 { false, 0x60000, "RSHIFT_NAND", FMA_SHIFT
},
544 { false, 0x61000, "RSHIFT_AND", FMA_SHIFT
},
545 { false, 0x62000, "LSHIFT_NAND", FMA_SHIFT
},
546 { false, 0x63000, "LSHIFT_AND", FMA_SHIFT
}, // (src0 << src2) & src1
547 { false, 0x64000, "RSHIFT_XOR", FMA_SHIFT
},
548 { false, 0x65200, "LSHIFT_ADD.i32", FMA_THREE_SRC
},
549 { false, 0x65600, "LSHIFT_SUB.i32", FMA_THREE_SRC
}, // (src0 << src2) - src1
550 { false, 0x65a00, "LSHIFT_RSUB.i32", FMA_THREE_SRC
}, // src1 - (src0 << src2)
551 { false, 0x65e00, "RSHIFT_ADD.i32", FMA_THREE_SRC
},
552 { false, 0x66200, "RSHIFT_SUB.i32", FMA_THREE_SRC
},
553 { false, 0x66600, "RSHIFT_RSUB.i32", FMA_THREE_SRC
},
554 { false, 0x66a00, "ARSHIFT_ADD.i32", FMA_THREE_SRC
},
555 { false, 0x66e00, "ARSHIFT_SUB.i32", FMA_THREE_SRC
},
556 { false, 0x67200, "ARSHIFT_RSUB.i32", FMA_THREE_SRC
},
557 { false, 0x80000, "FMA.v2f16", FMA_FMA16
},
558 { false, 0xc0000, "MAX.v2f16", FMA_FMINMAX16
},
559 { false, 0xc4000, "MIN.v2f16", FMA_FMINMAX16
},
560 { false, 0xc8000, "FCMP.GL", FMA_FCMP16
},
561 { false, 0xcc000, "FCMP.D3D", FMA_FCMP16
},
562 { false, 0xcf900, "ADD.v2i16", FMA_TWO_SRC
},
563 { false, 0xcfc10, "ADDC.i32", FMA_TWO_SRC
},
564 { false, 0xcfd80, "ADD.i32.i16.X", FMA_TWO_SRC
},
565 { false, 0xcfd90, "ADD.i32.u16.X", FMA_TWO_SRC
},
566 { false, 0xcfdc0, "ADD.i32.i16.Y", FMA_TWO_SRC
},
567 { false, 0xcfdd0, "ADD.i32.u16.Y", FMA_TWO_SRC
},
568 { false, 0xd8000, "ADD.v2f16", FMA_FADD16
},
569 { false, 0xdc000, "CSEL4.v16", FMA_CSEL4
},
570 { false, 0xdd000, "F32_TO_F16", FMA_TWO_SRC
},
571 { true, 0x00046, "F16_TO_I16.XX", FMA_ONE_SRC
},
572 { true, 0x00047, "F16_TO_U16.XX", FMA_ONE_SRC
},
573 { true, 0x0004e, "F16_TO_I16.YX", FMA_ONE_SRC
},
574 { true, 0x0004f, "F16_TO_U16.YX", FMA_ONE_SRC
},
575 { true, 0x00056, "F16_TO_I16.XY", FMA_ONE_SRC
},
576 { true, 0x00057, "F16_TO_U16.XY", FMA_ONE_SRC
},
577 { true, 0x0005e, "F16_TO_I16.YY", FMA_ONE_SRC
},
578 { true, 0x0005f, "F16_TO_U16.YY", FMA_ONE_SRC
},
579 { true, 0x000c0, "I16_TO_F16.XX", FMA_ONE_SRC
},
580 { true, 0x000c1, "U16_TO_F16.XX", FMA_ONE_SRC
},
581 { true, 0x000c8, "I16_TO_F16.YX", FMA_ONE_SRC
},
582 { true, 0x000c9, "U16_TO_F16.YX", FMA_ONE_SRC
},
583 { true, 0x000d0, "I16_TO_F16.XY", FMA_ONE_SRC
},
584 { true, 0x000d1, "U16_TO_F16.XY", FMA_ONE_SRC
},
585 { true, 0x000d8, "I16_TO_F16.YY", FMA_ONE_SRC
},
586 { true, 0x000d9, "U16_TO_F16.YY", FMA_ONE_SRC
},
587 { true, 0x00136, "F32_TO_I32", FMA_ONE_SRC
},
588 { true, 0x00137, "F32_TO_U32", FMA_ONE_SRC
},
589 { true, 0x00178, "I32_TO_F32", FMA_ONE_SRC
},
590 { true, 0x00179, "U32_TO_F32", FMA_ONE_SRC
},
591 { true, 0x00198, "I16_TO_I32.X", FMA_ONE_SRC
},
592 { true, 0x00199, "U16_TO_U32.X", FMA_ONE_SRC
},
593 { true, 0x0019a, "I16_TO_I32.Y", FMA_ONE_SRC
},
594 { true, 0x0019b, "U16_TO_U32.Y", FMA_ONE_SRC
},
595 { true, 0x0019c, "I16_TO_F32.X", FMA_ONE_SRC
},
596 { true, 0x0019d, "U16_TO_F32.X", FMA_ONE_SRC
},
597 { true, 0x0019e, "I16_TO_F32.Y", FMA_ONE_SRC
},
598 { true, 0x0019f, "U16_TO_F32.Y", FMA_ONE_SRC
},
599 { true, 0x001a2, "F16_TO_F32.X", FMA_ONE_SRC
},
600 { true, 0x001a3, "F16_TO_F32.Y", FMA_ONE_SRC
},
601 { true, 0x0032c, "NOP", FMA_ONE_SRC
},
602 { true, 0x0032d, "MOV", FMA_ONE_SRC
},
603 { true, 0x0032f, "SWZ.YY.v2i16", FMA_ONE_SRC
},
604 { true, 0x00345, "LOG_FREXPM", FMA_ONE_SRC
},
605 { true, 0x00365, "FRCP_FREXPM", FMA_ONE_SRC
},
606 { true, 0x00375, "FSQRT_FREXPM", FMA_ONE_SRC
},
607 { true, 0x0038d, "FRCP_FREXPE", FMA_ONE_SRC
},
608 { true, 0x003a5, "FSQRT_FREXPE", FMA_ONE_SRC
},
609 { true, 0x003ad, "FRSQ_FREXPE", FMA_ONE_SRC
},
610 { true, 0x003c5, "LOG_FREXPE", FMA_ONE_SRC
},
611 { true, 0x003fa, "CLZ", FMA_ONE_SRC
},
612 { true, 0x00b80, "IMAX3", FMA_THREE_SRC
},
613 { true, 0x00bc0, "UMAX3", FMA_THREE_SRC
},
614 { true, 0x00c00, "IMIN3", FMA_THREE_SRC
},
615 { true, 0x00c40, "UMIN3", FMA_THREE_SRC
},
616 { true, 0x00ec5, "ROUND", FMA_ONE_SRC
},
617 { true, 0x00f40, "CSEL", FMA_THREE_SRC
}, // src2 != 0 ? src1 : src0
618 { true, 0x00fc0, "MUX.i32", FMA_THREE_SRC
}, // see ADD comment
619 { true, 0x01805, "ROUNDEVEN", FMA_ONE_SRC
},
620 { true, 0x01845, "CEIL", FMA_ONE_SRC
},
621 { true, 0x01885, "FLOOR", FMA_ONE_SRC
},
622 { true, 0x018c5, "TRUNC", FMA_ONE_SRC
},
623 { true, 0x019b0, "ATAN_LDEXP.Y.f32", FMA_TWO_SRC
},
624 { true, 0x019b8, "ATAN_LDEXP.X.f32", FMA_TWO_SRC
},
625 { true, 0x01c80, "LSHIFT_ADD_LOW32.u32", FMA_SHIFT_ADD64
},
626 { true, 0x01cc0, "LSHIFT_ADD_LOW32.i64", FMA_SHIFT_ADD64
},
627 { true, 0x01d80, "LSHIFT_ADD_LOW32.i32", FMA_SHIFT_ADD64
},
628 { true, 0x01e00, "SEL.XX.i16", FMA_TWO_SRC
},
629 { true, 0x01e08, "SEL.YX.i16", FMA_TWO_SRC
},
630 { true, 0x01e10, "SEL.XY.i16", FMA_TWO_SRC
},
631 { true, 0x01e18, "SEL.YY.i16", FMA_TWO_SRC
},
632 { true, 0x00800, "IMAD", FMA_THREE_SRC
},
633 { true, 0x078db, "POPCNT", FMA_ONE_SRC
},
636 static struct fma_op_info
find_fma_op_info(unsigned op
, bool extended
)
638 for (unsigned i
= 0; i
< ARRAY_SIZE(FMAOpInfos
); i
++) {
641 if (FMAOpInfos
[i
].extended
!= extended
)
647 switch (FMAOpInfos
[i
].src_type
) {
656 opCmp
= op
& ~0x1fff;
659 case FMA_SHIFT_ADD64
:
666 opCmp
= op
& ~0x3fff;
670 opCmp
= op
& ~0x3ffff;
677 opCmp
= op
& ~0x7fff;
683 if (FMAOpInfos
[i
].op
== opCmp
)
684 return FMAOpInfos
[i
];
687 struct fma_op_info info
;
688 snprintf(info
.name
, sizeof(info
.name
), "op%04x", op
);
690 info
.src_type
= FMA_THREE_SRC
;
694 static void dump_fcmp(FILE *fp
, unsigned op
)
716 fprintf(fp
, ".unk%d", op
);
721 static void dump_16swizzle(FILE *fp
, unsigned swiz
)
725 fprintf(fp
, ".%c%c", "xy"[swiz
& 1], "xy"[(swiz
>> 1) & 1]);
728 static void dump_fma_expand_src0(FILE *fp
, unsigned ctrl
)
750 static void dump_fma_expand_src1(FILE *fp
, unsigned ctrl
)
773 bi_ldst_type_name(enum bifrost_ldst_type type
)
776 case BIFROST_LDST_F16
: return "f16";
777 case BIFROST_LDST_F32
: return "f32";
778 case BIFROST_LDST_I32
: return "i32";
779 case BIFROST_LDST_U32
: return "u32";
780 default: return "invalid";
784 static void dump_fma(FILE *fp
, uint64_t word
, struct bifrost_regs regs
, struct bifrost_regs next_regs
, uint64_t *consts
, bool verbose
)
787 fprintf(fp
, "# FMA: %016" PRIx64
"\n", word
);
789 struct bifrost_fma_inst FMA
;
790 memcpy((char *) &FMA
, (char *) &word
, sizeof(struct bifrost_fma_inst
));
791 struct fma_op_info info
= find_fma_op_info(FMA
.op
, (FMA
.op
& 0xe0000) == 0xe0000);
793 fprintf(fp
, "%s", info
.name
);
794 if (info
.src_type
== FMA_FADD
||
795 info
.src_type
== FMA_FMINMAX
||
796 info
.src_type
== FMA_FMA
||
797 info
.src_type
== FMA_FADD16
||
798 info
.src_type
== FMA_FMINMAX16
||
799 info
.src_type
== FMA_FMA16
) {
800 dump_output_mod(fp
, bits(FMA
.op
, 12, 14));
801 switch (info
.src_type
) {
806 dump_round_mode(fp
, bits(FMA
.op
, 10, 12));
810 dump_minmax_mode(fp
, bits(FMA
.op
, 10, 12));
815 } else if (info
.src_type
== FMA_FCMP
|| info
.src_type
== FMA_FCMP16
) {
816 dump_fcmp(fp
, bits(FMA
.op
, 10, 13));
817 if (info
.src_type
== FMA_FCMP
)
820 fprintf(fp
, ".v2f16");
821 } else if (info
.src_type
== FMA_FMA_MSCALE
) {
822 if (FMA
.op
& (1 << 11)) {
823 switch ((FMA
.op
>> 9) & 0x3) {
825 /* This mode seems to do a few things:
826 * - Makes 0 * infinity (and incidentally 0 * nan) return 0,
827 * since generating a nan would poison the result of
828 * 1/infinity and 1/0.
829 * - Fiddles with which nan is returned in nan * nan,
830 * presumably to make sure that the same exact nan is
831 * returned for 1/nan.
833 fprintf(fp
, ".rcp_mode");
836 /* Similar to the above, but src0 always wins when multiplying
839 fprintf(fp
, ".sqrt_mode");
842 fprintf(fp
, ".unk%d_mode", (int) (FMA
.op
>> 9) & 0x3);
845 dump_output_mod(fp
, bits(FMA
.op
, 9, 11));
847 } else if (info
.src_type
== FMA_SHIFT
) {
848 struct bifrost_shift_fma shift
;
849 memcpy(&shift
, &FMA
, sizeof(shift
));
851 if (shift
.half
== 0x7)
852 fprintf(fp
, ".v2i16");
853 else if (shift
.half
== 0)
855 else if (shift
.half
== 0x4)
856 fprintf(fp
, ".v4i8");
858 fprintf(fp
, ".unk%u", shift
.half
);
861 fprintf(fp
, ".no_unk");
864 fprintf(fp
, ".invert_1");
867 fprintf(fp
, ".invert_2");
872 struct bifrost_reg_ctrl next_ctrl
= DecodeRegCtrl(fp
, next_regs
);
873 if (next_ctrl
.fma_write_unit
!= REG_WRITE_NONE
) {
874 fprintf(fp
, "{R%d, T0}, ", GetRegToWrite(next_ctrl
.fma_write_unit
, next_regs
));
879 switch (info
.src_type
) {
881 dump_src(fp
, FMA
.src0
, regs
, consts
, true);
884 dump_src(fp
, FMA
.src0
, regs
, consts
, true);
886 dump_src(fp
, FMA
.op
& 0x7, regs
, consts
, true);
894 dump_src(fp
, FMA
.src0
, regs
, consts
, true);
895 dump_fma_expand_src0(fp
, (FMA
.op
>> 6) & 0x7);
903 dump_src(fp
, FMA
.op
& 0x7, regs
, consts
, true);
904 dump_fma_expand_src1(fp
, (FMA
.op
>> 6) & 0x7);
909 case FMA_FMINMAX16
: {
910 bool abs1
= FMA
.op
& 0x8;
911 bool abs2
= (FMA
.op
& 0x7) < FMA
.src0
;
916 dump_src(fp
, FMA
.src0
, regs
, consts
, true);
917 dump_16swizzle(fp
, (FMA
.op
>> 6) & 0x3);
925 dump_src(fp
, FMA
.op
& 0x7, regs
, consts
, true);
926 dump_16swizzle(fp
, (FMA
.op
>> 8) & 0x3);
934 dump_src(fp
, FMA
.src0
, regs
, consts
, true);
935 dump_fma_expand_src0(fp
, (FMA
.op
>> 6) & 0x7);
943 dump_src(fp
, FMA
.op
& 0x7, regs
, consts
, true);
944 dump_fma_expand_src1(fp
, (FMA
.op
>> 6) & 0x7);
949 dump_src(fp
, FMA
.src0
, regs
, consts
, true);
950 // Note: this is kinda a guess, I haven't seen the blob set this to
951 // anything other than the identity, but it matches FMA_TWO_SRCFmod16
952 dump_16swizzle(fp
, (FMA
.op
>> 6) & 0x3);
954 dump_src(fp
, FMA
.op
& 0x7, regs
, consts
, true);
955 dump_16swizzle(fp
, (FMA
.op
>> 8) & 0x3);
957 case FMA_SHIFT_ADD64
:
958 dump_src(fp
, FMA
.src0
, regs
, consts
, true);
960 dump_src(fp
, FMA
.op
& 0x7, regs
, consts
, true);
962 fprintf(fp
, "shift:%u", (FMA
.op
>> 3) & 0x7);
965 dump_src(fp
, FMA
.src0
, regs
, consts
, true);
967 dump_src(fp
, FMA
.op
& 0x7, regs
, consts
, true);
969 dump_src(fp
, (FMA
.op
>> 3) & 0x7, regs
, consts
, true);
972 struct bifrost_shift_fma shift
;
973 memcpy(&shift
, &FMA
, sizeof(shift
));
975 dump_src(fp
, shift
.src0
, regs
, consts
, true);
977 dump_src(fp
, shift
.src1
, regs
, consts
, true);
979 dump_src(fp
, shift
.src2
, regs
, consts
, true);
983 if (FMA
.op
& (1 << 14))
985 if (FMA
.op
& (1 << 9))
987 dump_src(fp
, FMA
.src0
, regs
, consts
, true);
988 dump_fma_expand_src0(fp
, (FMA
.op
>> 6) & 0x7);
989 if (FMA
.op
& (1 << 9))
992 if (FMA
.op
& (1 << 16))
994 dump_src(fp
, FMA
.op
& 0x7, regs
, consts
, true);
995 dump_fma_expand_src1(fp
, (FMA
.op
>> 6) & 0x7);
996 if (FMA
.op
& (1 << 16))
999 if (FMA
.op
& (1 << 15))
1001 if (FMA
.op
& (1 << 17))
1002 fprintf(fp
, "abs(");
1003 dump_src(fp
, (FMA
.op
>> 3) & 0x7, regs
, consts
, true);
1004 if (FMA
.op
& (1 << 17))
1008 if (FMA
.op
& (1 << 14))
1010 dump_src(fp
, FMA
.src0
, regs
, consts
, true);
1011 dump_16swizzle(fp
, (FMA
.op
>> 6) & 0x3);
1013 dump_src(fp
, FMA
.op
& 0x7, regs
, consts
, true);
1014 dump_16swizzle(fp
, (FMA
.op
>> 8) & 0x3);
1016 if (FMA
.op
& (1 << 15))
1018 dump_src(fp
, (FMA
.op
>> 3) & 0x7, regs
, consts
, true);
1019 dump_16swizzle(fp
, (FMA
.op
>> 16) & 0x3);
1022 struct bifrost_csel4 csel
;
1023 memcpy(&csel
, &FMA
, sizeof(csel
));
1024 fprintf(fp
, ".%s ", csel_cond_name(csel
.cond
));
1026 dump_src(fp
, csel
.src0
, regs
, consts
, true);
1028 dump_src(fp
, csel
.src1
, regs
, consts
, true);
1030 dump_src(fp
, csel
.src2
, regs
, consts
, true);
1032 dump_src(fp
, csel
.src3
, regs
, consts
, true);
1035 case FMA_FMA_MSCALE
:
1036 if (FMA
.op
& (1 << 12))
1037 fprintf(fp
, "abs(");
1038 dump_src(fp
, FMA
.src0
, regs
, consts
, true);
1039 if (FMA
.op
& (1 << 12))
1042 if (FMA
.op
& (1 << 13))
1044 dump_src(fp
, FMA
.op
& 0x7, regs
, consts
, true);
1046 if (FMA
.op
& (1 << 14))
1048 dump_src(fp
, (FMA
.op
>> 3) & 0x7, regs
, consts
, true);
1050 dump_src(fp
, (FMA
.op
>> 6) & 0x7, regs
, consts
, true);
1056 static const struct add_op_info add_op_infos
[] = {
1057 { 0x00000, "MAX.f32", ADD_FMINMAX
},
1058 { 0x02000, "MIN.f32", ADD_FMINMAX
},
1059 { 0x04000, "ADD.f32", ADD_FADD
},
1060 { 0x06000, "FCMP.GL", ADD_FCMP
},
1061 { 0x07000, "FCMP.D3D", ADD_FCMP
},
1062 { 0x07856, "F16_TO_I16", ADD_ONE_SRC
},
1063 { 0x07857, "F16_TO_U16", ADD_ONE_SRC
},
1064 { 0x078c0, "I16_TO_F16.XX", ADD_ONE_SRC
},
1065 { 0x078c1, "U16_TO_F16.XX", ADD_ONE_SRC
},
1066 { 0x078c8, "I16_TO_F16.YX", ADD_ONE_SRC
},
1067 { 0x078c9, "U16_TO_F16.YX", ADD_ONE_SRC
},
1068 { 0x078d0, "I16_TO_F16.XY", ADD_ONE_SRC
},
1069 { 0x078d1, "U16_TO_F16.XY", ADD_ONE_SRC
},
1070 { 0x078d8, "I16_TO_F16.YY", ADD_ONE_SRC
},
1071 { 0x078d9, "U16_TO_F16.YY", ADD_ONE_SRC
},
1072 { 0x07936, "F32_TO_I32", ADD_ONE_SRC
},
1073 { 0x07937, "F32_TO_U32", ADD_ONE_SRC
},
1074 { 0x07978, "I32_TO_F32", ADD_ONE_SRC
},
1075 { 0x07979, "U32_TO_F32", ADD_ONE_SRC
},
1076 { 0x07998, "I16_TO_I32.X", ADD_ONE_SRC
},
1077 { 0x07999, "U16_TO_U32.X", ADD_ONE_SRC
},
1078 { 0x0799a, "I16_TO_I32.Y", ADD_ONE_SRC
},
1079 { 0x0799b, "U16_TO_U32.Y", ADD_ONE_SRC
},
1080 { 0x0799c, "I16_TO_F32.X", ADD_ONE_SRC
},
1081 { 0x0799d, "U16_TO_F32.X", ADD_ONE_SRC
},
1082 { 0x0799e, "I16_TO_F32.Y", ADD_ONE_SRC
},
1083 { 0x0799f, "U16_TO_F32.Y", ADD_ONE_SRC
},
1084 { 0x079a2, "F16_TO_F32.X", ADD_ONE_SRC
},
1085 { 0x079a3, "F16_TO_F32.Y", ADD_ONE_SRC
},
1086 { 0x07b2b, "SWZ.YX.v2i16", ADD_ONE_SRC
},
1087 { 0x07b2c, "NOP", ADD_ONE_SRC
},
1088 { 0x07b29, "SWZ.XX.v2i16", ADD_ONE_SRC
},
1089 { 0x07b2d, "MOV", ADD_ONE_SRC
},
1090 { 0x07b2f, "SWZ.YY.v2i16", ADD_ONE_SRC
},
1091 { 0x07b65, "FRCP_FREXPM", ADD_ONE_SRC
},
1092 { 0x07b75, "FSQRT_FREXPM", ADD_ONE_SRC
},
1093 { 0x07b8d, "FRCP_FREXPE", ADD_ONE_SRC
},
1094 { 0x07ba5, "FSQRT_FREXPE", ADD_ONE_SRC
},
1095 { 0x07bad, "FRSQ_FREXPE", ADD_ONE_SRC
},
1096 { 0x07bc5, "FLOG_FREXPE", ADD_ONE_SRC
},
1097 { 0x07d45, "CEIL", ADD_ONE_SRC
},
1098 { 0x07d85, "FLOOR", ADD_ONE_SRC
},
1099 { 0x07dc5, "TRUNC", ADD_ONE_SRC
},
1100 { 0x07f18, "LSHIFT_ADD_HIGH32.i32", ADD_TWO_SRC
},
1101 { 0x08000, "LD_ATTR", ADD_LOAD_ATTR
, true },
1102 { 0x0a000, "LD_VAR.32", ADD_VARYING_INTERP
, true },
1103 { 0x0b000, "TEX", ADD_TEX_COMPACT
, true },
1104 { 0x0c188, "LOAD.i32", ADD_TWO_SRC
, true },
1105 { 0x0c1a0, "LD_UBO.i32", ADD_TWO_SRC
, true },
1106 { 0x0c1b8, "LD_SCRATCH.v2i32", ADD_TWO_SRC
, true },
1107 { 0x0c1c8, "LOAD.v2i32", ADD_TWO_SRC
, true },
1108 { 0x0c1e0, "LD_UBO.v2i32", ADD_TWO_SRC
, true },
1109 { 0x0c1f8, "LD_SCRATCH.v2i32", ADD_TWO_SRC
, true },
1110 { 0x0c208, "LOAD.v4i32", ADD_TWO_SRC
, true },
1111 { 0x0c220, "LD_UBO.v4i32", ADD_TWO_SRC
, true },
1112 { 0x0c238, "LD_SCRATCH.v4i32", ADD_TWO_SRC
, true },
1113 { 0x0c248, "STORE.v4i32", ADD_TWO_SRC
, true },
1114 { 0x0c278, "ST_SCRATCH.v4i32", ADD_TWO_SRC
, true },
1115 { 0x0c588, "STORE.i32", ADD_TWO_SRC
, true },
1116 { 0x0c5b8, "ST_SCRATCH.i32", ADD_TWO_SRC
, true },
1117 { 0x0c5c8, "STORE.v2i32", ADD_TWO_SRC
, true },
1118 { 0x0c5f8, "ST_SCRATCH.v2i32", ADD_TWO_SRC
, true },
1119 { 0x0c648, "LOAD.u16", ADD_TWO_SRC
, true }, // zero-extends
1120 { 0x0ca88, "LOAD.v3i32", ADD_TWO_SRC
, true },
1121 { 0x0caa0, "LD_UBO.v3i32", ADD_TWO_SRC
, true },
1122 { 0x0cab8, "LD_SCRATCH.v3i32", ADD_TWO_SRC
, true },
1123 { 0x0cb88, "STORE.v3i32", ADD_TWO_SRC
, true },
1124 { 0x0cbb8, "ST_SCRATCH.v3i32", ADD_TWO_SRC
, true },
1125 { 0x0cc00, "FRCP_FAST.f32", ADD_ONE_SRC
},
1126 { 0x0cc20, "FRSQ_FAST.f32", ADD_ONE_SRC
},
1127 { 0x0ce00, "FRCP_TABLE", ADD_ONE_SRC
},
1128 { 0x0ce10, "FRCP_FAST.f16.X", ADD_ONE_SRC
},
1129 { 0x0ce20, "FRSQ_TABLE", ADD_ONE_SRC
},
1130 { 0x0ce30, "FRCP_FAST.f16.Y", ADD_ONE_SRC
},
1131 { 0x0ce50, "FRSQ_FAST.f16.X", ADD_ONE_SRC
},
1132 { 0x0ce60, "FRCP_APPROX", ADD_ONE_SRC
},
1133 { 0x0ce70, "FRSQ_FAST.f16.Y", ADD_ONE_SRC
},
1134 { 0x0cf40, "ATAN_ASSIST", ADD_TWO_SRC
},
1135 { 0x0cf48, "ATAN_TABLE", ADD_TWO_SRC
},
1136 { 0x0cf50, "SIN_TABLE", ADD_ONE_SRC
},
1137 { 0x0cf51, "COS_TABLE", ADD_ONE_SRC
},
1138 { 0x0cf58, "EXP_TABLE", ADD_ONE_SRC
},
1139 { 0x0cf60, "FLOG2_TABLE", ADD_ONE_SRC
},
1140 { 0x0cf64, "FLOGE_TABLE", ADD_ONE_SRC
},
1141 { 0x0d000, "BRANCH", ADD_BRANCH
},
1142 { 0x0e8c0, "MUX", ADD_THREE_SRC
},
1143 { 0x0e9b0, "ATAN_LDEXP.Y.f32", ADD_TWO_SRC
},
1144 { 0x0e9b8, "ATAN_LDEXP.X.f32", ADD_TWO_SRC
},
1145 { 0x0ea60, "SEL.XX.i16", ADD_TWO_SRC
},
1146 { 0x0ea70, "SEL.XY.i16", ADD_TWO_SRC
},
1147 { 0x0ea68, "SEL.YX.i16", ADD_TWO_SRC
},
1148 { 0x0ea78, "SEL.YY.i16", ADD_TWO_SRC
},
1149 { 0x0ec00, "F32_TO_F16", ADD_TWO_SRC
},
1150 { 0x0f640, "ICMP.GL.GT", ADD_TWO_SRC
}, // src0 > src1 ? 1 : 0
1151 { 0x0f648, "ICMP.GL.GE", ADD_TWO_SRC
},
1152 { 0x0f650, "UCMP.GL.GT", ADD_TWO_SRC
},
1153 { 0x0f658, "UCMP.GL.GE", ADD_TWO_SRC
},
1154 { 0x0f660, "ICMP.GL.EQ", ADD_TWO_SRC
},
1155 { 0x0f669, "ICMP.GL.NEQ", ADD_TWO_SRC
},
1156 { 0x0f6c0, "ICMP.D3D.GT", ADD_TWO_SRC
}, // src0 > src1 ? ~0 : 0
1157 { 0x0f6c8, "ICMP.D3D.GE", ADD_TWO_SRC
},
1158 { 0x0f6d0, "UCMP.D3D.GT", ADD_TWO_SRC
},
1159 { 0x0f6d8, "UCMP.D3D.GE", ADD_TWO_SRC
},
1160 { 0x0f6e0, "ICMP.D3D.EQ", ADD_TWO_SRC
},
1161 { 0x10000, "MAX.v2f16", ADD_FMINMAX16
},
1162 { 0x11000, "ADD_MSCALE.f32", ADD_FADDMscale
},
1163 { 0x12000, "MIN.v2f16", ADD_FMINMAX16
},
1164 { 0x14000, "ADD.v2f16", ADD_FADD16
},
1165 { 0x17000, "FCMP.D3D", ADD_FCMP16
},
1166 { 0x178c0, "ADD.i32", ADD_TWO_SRC
},
1167 { 0x17900, "ADD.v2i16", ADD_TWO_SRC
},
1168 { 0x17ac0, "SUB.i32", ADD_TWO_SRC
},
1169 { 0x17c10, "ADDC.i32", ADD_TWO_SRC
}, // adds src0 to the bottom bit of src1
1170 { 0x17d80, "ADD.i32.i16.X", ADD_TWO_SRC
},
1171 { 0x17d90, "ADD.i32.u16.X", ADD_TWO_SRC
},
1172 { 0x17dc0, "ADD.i32.i16.Y", ADD_TWO_SRC
},
1173 { 0x17dd0, "ADD.i32.u16.Y", ADD_TWO_SRC
},
1174 { 0x18000, "LD_VAR_ADDR", ADD_VARYING_ADDRESS
, true },
1175 { 0x19181, "DISCARD.FEQ.f32", ADD_TWO_SRC
, true },
1176 { 0x19189, "DISCARD.FNE.f32", ADD_TWO_SRC
, true },
1177 { 0x1918C, "DISCARD.GL.f32", ADD_TWO_SRC
, true }, /* Consumes ICMP.GL/etc with fixed 0 argument */
1178 { 0x19190, "DISCARD.FLE.f32", ADD_TWO_SRC
, true },
1179 { 0x19198, "DISCARD.FLT.f32", ADD_TWO_SRC
, true },
1180 { 0x191e8, "ATEST.f32", ADD_TWO_SRC
, true },
1181 { 0x191f0, "ATEST.X.f16", ADD_TWO_SRC
, true },
1182 { 0x191f8, "ATEST.Y.f16", ADD_TWO_SRC
, true },
1183 { 0x19300, "ST_VAR.v1", ADD_THREE_SRC
, true },
1184 { 0x19340, "ST_VAR.v2", ADD_THREE_SRC
, true },
1185 { 0x19380, "ST_VAR.v3", ADD_THREE_SRC
, true },
1186 { 0x193c0, "ST_VAR.v4", ADD_THREE_SRC
, true },
1187 { 0x1952c, "BLEND", ADD_BLENDING
, true },
1188 { 0x1a000, "LD_VAR.16", ADD_VARYING_INTERP
, true },
1189 { 0x1ae60, "TEX", ADD_TEX
, true },
1190 { 0x1c000, "RSHIFT_NAND.i32", ADD_SHIFT
},
1191 { 0x1c400, "RSHIFT_AND.i32", ADD_SHIFT
},
1192 { 0x1c800, "LSHIFT_NAND.i32", ADD_SHIFT
},
1193 { 0x1cc00, "LSHIFT_AND.i32", ADD_SHIFT
},
1194 { 0x1d000, "RSHIFT_XOR.i32", ADD_SHIFT
},
1195 { 0x1d400, "LSHIFT_ADD.i32", ADD_SHIFT
},
1196 { 0x1d800, "RSHIFT_SUB.i32", ADD_SHIFT
},
1197 { 0x1dd18, "OR.i32", ADD_TWO_SRC
},
1198 { 0x1dd20, "AND.i32", ADD_TWO_SRC
},
1199 { 0x1dd60, "LSHIFT.i32", ADD_TWO_SRC
},
1200 { 0x1dd50, "XOR.i32", ADD_TWO_SRC
},
1201 { 0x1dd80, "RSHIFT.i32", ADD_TWO_SRC
},
1202 { 0x1dda0, "ARSHIFT.i32", ADD_TWO_SRC
},
1205 static struct add_op_info
find_add_op_info(unsigned op
)
1207 for (unsigned i
= 0; i
< ARRAY_SIZE(add_op_infos
); i
++) {
1208 unsigned opCmp
= ~0;
1209 switch (add_op_infos
[i
].src_type
) {
1221 opCmp
= op
& ~0x3ff;
1229 opCmp
= op
& ~0x1fff;
1232 case ADD_FADDMscale
:
1233 opCmp
= op
& ~0xfff;
1237 opCmp
= op
& ~0x7ff;
1239 case ADD_TEX_COMPACT
:
1240 opCmp
= op
& ~0x3ff;
1242 case ADD_VARYING_INTERP
:
1243 opCmp
= op
& ~0x7ff;
1245 case ADD_VARYING_ADDRESS
:
1246 opCmp
= op
& ~0xfff;
1250 opCmp
= op
& ~0xfff;
1256 if (add_op_infos
[i
].op
== opCmp
)
1257 return add_op_infos
[i
];
1260 struct add_op_info info
;
1261 snprintf(info
.name
, sizeof(info
.name
), "op%04x", op
);
1263 info
.src_type
= ADD_TWO_SRC
;
1264 info
.has_data_reg
= true;
1268 static void dump_add(FILE *fp
, uint64_t word
, struct bifrost_regs regs
,
1269 struct bifrost_regs next_regs
, uint64_t *consts
,
1270 unsigned data_reg
, unsigned offset
, bool verbose
)
1273 fprintf(fp
, "# ADD: %016" PRIx64
"\n", word
);
1275 struct bifrost_add_inst ADD
;
1276 memcpy((char *) &ADD
, (char *) &word
, sizeof(ADD
));
1277 struct add_op_info info
= find_add_op_info(ADD
.op
);
1279 fprintf(fp
, "%s", info
.name
);
1281 // float16 seems like it doesn't support output modifiers
1282 if (info
.src_type
== ADD_FADD
|| info
.src_type
== ADD_FMINMAX
) {
1284 dump_output_mod(fp
, bits(ADD
.op
, 8, 10));
1285 if (info
.src_type
== ADD_FADD
)
1286 dump_round_mode(fp
, bits(ADD
.op
, 10, 12));
1288 dump_minmax_mode(fp
, bits(ADD
.op
, 10, 12));
1289 } else if (info
.src_type
== ADD_FCMP
|| info
.src_type
== ADD_FCMP16
) {
1290 dump_fcmp(fp
, bits(ADD
.op
, 3, 6));
1291 if (info
.src_type
== ADD_FCMP
)
1292 fprintf(fp
, ".f32");
1294 fprintf(fp
, ".v2f16");
1295 } else if (info
.src_type
== ADD_FADDMscale
) {
1296 switch ((ADD
.op
>> 6) & 0x7) {
1299 // causes GPU hangs on G71
1301 fprintf(fp
, ".invalid");
1303 // Same as usual outmod value.
1305 fprintf(fp
, ".clamp_0_1");
1307 // If src0 is infinite or NaN, flush it to zero so that the other
1308 // source is passed through unmodified.
1310 fprintf(fp
, ".flush_src0_inf_nan");
1314 fprintf(fp
, ".flush_src1_inf_nan");
1316 // Every other case seems to behave the same as the above?
1318 fprintf(fp
, ".unk%d", (ADD
.op
>> 6) & 0x7);
1321 } else if (info
.src_type
== ADD_VARYING_INTERP
) {
1323 fprintf(fp
, ".reuse");
1325 fprintf(fp
, ".flat");
1326 switch ((ADD
.op
>> 7) & 0x3) {
1328 fprintf(fp
, ".per_frag");
1331 fprintf(fp
, ".centroid");
1336 fprintf(fp
, ".explicit");
1339 fprintf(fp
, ".v%d", ((ADD
.op
>> 5) & 0x3) + 1);
1340 } else if (info
.src_type
== ADD_BRANCH
) {
1341 enum bifrost_branch_code branchCode
= (enum bifrost_branch_code
) ((ADD
.op
>> 6) & 0x3f);
1342 if (branchCode
== BR_ALWAYS
) {
1343 // unconditional branch
1345 enum bifrost_branch_cond cond
= (enum bifrost_branch_cond
) ((ADD
.op
>> 6) & 0x7);
1346 enum branch_bit_size size
= (enum branch_bit_size
) ((ADD
.op
>> 9) & 0x7);
1347 bool portSwapped
= (ADD
.op
& 0x7) < ADD
.src0
;
1348 // See the comment in branch_bit_size
1349 if (size
== BR_SIZE_16YX0
)
1351 if (size
== BR_SIZE_16YX1
)
1352 portSwapped
= false;
1353 // These sizes are only for floating point comparisons, so the
1354 // non-floating-point comparisons are reused to encode the flipped
1356 if (size
== BR_SIZE_32_AND_16X
|| size
== BR_SIZE_32_AND_16Y
)
1357 portSwapped
= false;
1358 // There's only one argument, so we reuse the extra argument to
1360 if (size
== BR_SIZE_ZERO
)
1361 portSwapped
= !(ADD
.op
& 1);
1366 fprintf(fp
, ".LT.u");
1368 fprintf(fp
, ".LT.i");
1371 if (size
== BR_SIZE_32_AND_16X
|| size
== BR_SIZE_32_AND_16Y
) {
1372 fprintf(fp
, ".UNE.f");
1375 fprintf(fp
, ".LE.u");
1377 fprintf(fp
, ".LE.i");
1382 fprintf(fp
, ".GT.u");
1384 fprintf(fp
, ".GT.i");
1388 fprintf(fp
, ".GE.u");
1390 fprintf(fp
, ".GE.i");
1394 fprintf(fp
, ".NE.i");
1396 fprintf(fp
, ".EQ.i");
1400 fprintf(fp
, ".UNE.f");
1402 fprintf(fp
, ".OEQ.f");
1406 fprintf(fp
, ".OGT.unk.f");
1408 fprintf(fp
, ".OGT.f");
1412 fprintf(fp
, ".OLT.unk.f");
1414 fprintf(fp
, ".OLT.f");
1419 case BR_SIZE_32_AND_16X
:
1420 case BR_SIZE_32_AND_16Y
:
1429 case BR_SIZE_ZERO
: {
1430 unsigned ctrl
= (ADD
.op
>> 1) & 0x3;
1432 fprintf(fp
, "32.Z");
1434 fprintf(fp
, "16.Z");
1439 } else if (info
.src_type
== ADD_SHIFT
) {
1440 struct bifrost_shift_add shift
;
1441 memcpy(&shift
, &ADD
, sizeof(ADD
));
1444 fprintf(fp
, ".invert_1");
1447 fprintf(fp
, ".invert_2");
1450 fprintf(fp
, ".unk%u", shift
.zero
);
1451 } else if (info
.src_type
== ADD_VARYING_ADDRESS
) {
1452 struct bifrost_ld_var_addr ld
;
1453 memcpy(&ld
, &ADD
, sizeof(ADD
));
1454 fprintf(fp
, ".%s", bi_ldst_type_name(ld
.type
));
1455 } else if (info
.src_type
== ADD_LOAD_ATTR
) {
1456 struct bifrost_ld_attr ld
;
1457 memcpy(&ld
, &ADD
, sizeof(ADD
));
1460 fprintf(fp
, ".v%d%s", ld
.channels
+ 1, bi_ldst_type_name(ld
.type
));
1462 fprintf(fp
, ".%s", bi_ldst_type_name(ld
.type
));
1467 struct bifrost_reg_ctrl next_ctrl
= DecodeRegCtrl(fp
, next_regs
);
1468 if (next_ctrl
.add_write_unit
!= REG_WRITE_NONE
) {
1469 fprintf(fp
, "{R%d, T1}, ", GetRegToWrite(next_ctrl
.add_write_unit
, next_regs
));
1471 fprintf(fp
, "T1, ");
1474 switch (info
.src_type
) {
1476 // Note: in this case, regs.uniform_const == location | 0x8
1477 // This probably means we can't load uniforms or immediates in the
1478 // same instruction. This re-uses the encoding that normally means
1479 // "disabled", where the low 4 bits are ignored. Perhaps the extra
1480 // 0x8 or'd in indicates this is happening.
1481 fprintf(fp
, "location:%d, ", regs
.uniform_const
& 0x7);
1484 dump_src(fp
, ADD
.src0
, regs
, consts
, false);
1487 case ADD_TEX_COMPACT
: {
1490 bool dualTex
= false;
1491 if (info
.src_type
== ADD_TEX_COMPACT
) {
1492 tex_index
= (ADD
.op
>> 3) & 0x7;
1493 sampler_index
= (ADD
.op
>> 7) & 0x7;
1494 bool unknown
= (ADD
.op
& 0x40);
1495 // TODO: figure out if the unknown bit is ever 0
1497 fprintf(fp
, "unknown ");
1499 uint64_t constVal
= get_const(consts
, regs
);
1500 uint32_t controlBits
= (ADD
.op
& 0x8) ? (constVal
>> 32) : constVal
;
1501 struct bifrost_tex_ctrl ctrl
;
1502 memcpy((char *) &ctrl
, (char *) &controlBits
, sizeof(ctrl
));
1504 // TODO: figure out what actually triggers dual-tex
1505 if (ctrl
.result_type
== 9) {
1506 struct bifrost_dual_tex_ctrl dualCtrl
;
1507 memcpy((char *) &dualCtrl
, (char *) &controlBits
, sizeof(ctrl
));
1508 fprintf(fp
, "(dualtex) tex0:%d samp0:%d tex1:%d samp1:%d ",
1509 dualCtrl
.tex_index0
, dualCtrl
.sampler_index0
,
1510 dualCtrl
.tex_index1
, dualCtrl
.sampler_index1
);
1511 if (dualCtrl
.unk0
!= 3)
1512 fprintf(fp
, "unk:%d ", dualCtrl
.unk0
);
1515 if (ctrl
.no_merge_index
) {
1516 tex_index
= ctrl
.tex_index
;
1517 sampler_index
= ctrl
.sampler_index
;
1519 tex_index
= sampler_index
= ctrl
.tex_index
;
1520 unsigned unk
= ctrl
.sampler_index
>> 2;
1522 fprintf(fp
, "unk:%d ", unk
);
1523 if (ctrl
.sampler_index
& 1)
1525 if (ctrl
.sampler_index
& 2)
1530 fprintf(fp
, "unk0:%d ", ctrl
.unk0
);
1532 fprintf(fp
, "unk1 ");
1533 if (ctrl
.unk2
!= 0xf)
1534 fprintf(fp
, "unk2:%x ", ctrl
.unk2
);
1536 switch (ctrl
.result_type
) {
1538 fprintf(fp
, "f32 ");
1541 fprintf(fp
, "i32 ");
1544 fprintf(fp
, "u32 ");
1547 fprintf(fp
, "unktype(%x) ", ctrl
.result_type
);
1550 switch (ctrl
.tex_type
) {
1552 fprintf(fp
, "cube ");
1555 fprintf(fp
, "buffer ");
1566 fprintf(fp
, "shadow ");
1568 fprintf(fp
, "array ");
1571 if (ctrl
.calc_gradients
) {
1572 int comp
= (controlBits
>> 20) & 0x3;
1573 fprintf(fp
, "txg comp:%d ", comp
);
1575 fprintf(fp
, "txf ");
1578 if (!ctrl
.not_supply_lod
) {
1579 if (ctrl
.compute_lod
)
1580 fprintf(fp
, "lod_bias ");
1582 fprintf(fp
, "lod ");
1585 if (!ctrl
.calc_gradients
)
1586 fprintf(fp
, "grad ");
1589 if (ctrl
.texel_offset
)
1590 fprintf(fp
, "offset ");
1595 if (tex_index
== -1)
1596 fprintf(fp
, "tex:indirect ");
1598 fprintf(fp
, "tex:%d ", tex_index
);
1600 if (sampler_index
== -1)
1601 fprintf(fp
, "samp:indirect ");
1603 fprintf(fp
, "samp:%d ", sampler_index
);
1607 case ADD_VARYING_INTERP
: {
1608 unsigned addr
= ADD
.op
& 0x1f;
1609 if (addr
< 0b10100) {
1611 fprintf(fp
, "%d", addr
);
1612 } else if (addr
< 0b11000) {
1614 fprintf(fp
, "fragw");
1615 else if (addr
== 23)
1616 fprintf(fp
, "fragz");
1618 fprintf(fp
, "unk%d", addr
);
1620 dump_src(fp
, ADD
.op
& 0x7, regs
, consts
, false);
1623 dump_src(fp
, ADD
.src0
, regs
, consts
, false);
1626 case ADD_VARYING_ADDRESS
: {
1627 dump_src(fp
, ADD
.src0
, regs
, consts
, false);
1629 dump_src(fp
, ADD
.op
& 0x7, regs
, consts
, false);
1631 unsigned location
= (ADD
.op
>> 3) & 0x1f;
1632 if (location
< 16) {
1633 fprintf(fp
, "location:%d", location
);
1634 } else if (location
== 20) {
1635 fprintf(fp
, "location:%u", (uint32_t) get_const(consts
, regs
));
1636 } else if (location
== 21) {
1637 fprintf(fp
, "location:%u", (uint32_t) (get_const(consts
, regs
) >> 32));
1639 fprintf(fp
, "location:%d(unk)", location
);
1644 fprintf(fp
, "location:%d, ", (ADD
.op
>> 3) & 0x1f);
1646 dump_src(fp
, ADD
.src0
, regs
, consts
, false);
1648 dump_src(fp
, ADD
.op
& 0x7, regs
, consts
, false);
1651 dump_src(fp
, ADD
.src0
, regs
, consts
, false);
1653 dump_src(fp
, ADD
.op
& 0x7, regs
, consts
, false);
1655 dump_src(fp
, (ADD
.op
>> 3) & 0x7, regs
, consts
, false);
1658 struct bifrost_shift_add shift
;
1659 memcpy(&shift
, &ADD
, sizeof(ADD
));
1660 dump_src(fp
, shift
.src0
, regs
, consts
, false);
1662 dump_src(fp
, shift
.src1
, regs
, consts
, false);
1664 dump_src(fp
, shift
.src2
, regs
, consts
, false);
1671 if (ADD
.op
& 0x1000)
1672 fprintf(fp
, "abs(");
1673 dump_src(fp
, ADD
.src0
, regs
, consts
, false);
1674 switch ((ADD
.op
>> 6) & 0x3) {
1681 if (ADD
.op
& 0x1000)
1687 fprintf(fp
, "abs(");
1688 dump_src(fp
, ADD
.op
& 0x7, regs
, consts
, false);
1689 switch ((ADD
.op
>> 6) & 0x3) {
1700 fprintf(fp
, ".unk");
1709 if (ADD
.op
& 0x1000)
1710 fprintf(fp
, "abs(");
1711 dump_src(fp
, ADD
.src0
, regs
, consts
, false);
1712 if (ADD
.op
& 0x1000)
1714 dump_16swizzle(fp
, (ADD
.op
>> 6) & 0x3);
1719 fprintf(fp
, "abs(");
1720 dump_src(fp
, ADD
.op
& 0x7, regs
, consts
, false);
1721 dump_16swizzle(fp
, (ADD
.op
>> 8) & 0x3);
1725 case ADD_FMINMAX16
: {
1726 bool abs1
= ADD
.op
& 0x8;
1727 bool abs2
= (ADD
.op
& 0x7) < ADD
.src0
;
1731 fprintf(fp
, "abs(");
1732 dump_src(fp
, ADD
.src0
, regs
, consts
, false);
1733 dump_16swizzle(fp
, (ADD
.op
>> 6) & 0x3);
1740 fprintf(fp
, "abs(");
1741 dump_src(fp
, ADD
.op
& 0x7, regs
, consts
, false);
1742 dump_16swizzle(fp
, (ADD
.op
>> 8) & 0x3);
1747 case ADD_FADDMscale
: {
1751 fprintf(fp
, "abs(");
1752 dump_src(fp
, ADD
.src0
, regs
, consts
, false);
1760 dump_src(fp
, ADD
.op
& 0x7, regs
, consts
, false);
1764 dump_src(fp
, (ADD
.op
>> 3) & 0x7, regs
, consts
, false);
1768 if (ADD
.op
& 0x400) {
1771 if (ADD
.op
& 0x100) {
1772 fprintf(fp
, "abs(");
1774 dump_src(fp
, ADD
.src0
, regs
, consts
, false);
1775 switch ((ADD
.op
>> 6) & 0x3) {
1782 if (ADD
.op
& 0x100) {
1786 if (ADD
.op
& 0x200) {
1787 fprintf(fp
, "abs(");
1789 dump_src(fp
, ADD
.op
& 0x7, regs
, consts
, false);
1790 switch ((ADD
.op
>> 6) & 0x3) {
1801 fprintf(fp
, ".unk");
1804 if (ADD
.op
& 0x200) {
1809 dump_src(fp
, ADD
.src0
, regs
, consts
, false);
1810 dump_16swizzle(fp
, (ADD
.op
>> 6) & 0x3);
1812 dump_src(fp
, ADD
.op
& 0x7, regs
, consts
, false);
1813 dump_16swizzle(fp
, (ADD
.op
>> 8) & 0x3);
1816 enum bifrost_branch_code code
= (enum bifrost_branch_code
) ((ADD
.op
>> 6) & 0x3f);
1817 enum branch_bit_size size
= (enum branch_bit_size
) ((ADD
.op
>> 9) & 0x7);
1818 if (code
!= BR_ALWAYS
) {
1819 dump_src(fp
, ADD
.src0
, regs
, consts
, false);
1829 case BR_SIZE_ZERO
: {
1830 unsigned ctrl
= (ADD
.op
>> 1) & 0x3;
1847 if (code
!= BR_ALWAYS
&& size
!= BR_SIZE_ZERO
) {
1848 dump_src(fp
, ADD
.op
& 0x7, regs
, consts
, false);
1853 case BR_SIZE_32_AND_16X
:
1857 case BR_SIZE_32_AND_16Y
:
1865 // I haven't had the chance to test if this actually specifies the
1866 // branch offset, since I couldn't get it to produce values other
1867 // than 5 (uniform/const high), but these three bits are always
1868 // consistent across branch instructions, so it makes sense...
1869 int offsetSrc
= (ADD
.op
>> 3) & 0x7;
1870 if (offsetSrc
== 4 || offsetSrc
== 5) {
1871 // If the offset is known/constant, we can decode it
1872 uint32_t raw_offset
;
1874 raw_offset
= get_const(consts
, regs
);
1876 raw_offset
= get_const(consts
, regs
) >> 32;
1877 // The high 4 bits are flags, while the rest is the
1878 // twos-complement offset in bytes (here we convert to
1880 int32_t branch_offset
= ((int32_t) raw_offset
<< 4) >> 8;
1882 // If high4 is the high 4 bits of the last 64-bit constant,
1883 // this is calculated as (high4 + 4) & 0xf, or 0 if the branch
1884 // offset itself is the last constant. Not sure if this is
1885 // actually used, or just garbage in unused bits, but in any
1886 // case, we can just ignore it here since it's redundant. Note
1887 // that if there is any padding, this will be 4 since the
1888 // padding counts as the last constant.
1889 unsigned flags
= raw_offset
>> 28;
1892 // Note: the offset is in bytes, relative to the beginning of the
1893 // current clause, so a zero offset would be a loop back to the
1894 // same clause (annoyingly different from Midgard).
1895 fprintf(fp
, "clause_%d", offset
+ branch_offset
);
1897 dump_src(fp
, offsetSrc
, regs
, consts
, false);
1901 if (info
.has_data_reg
) {
1902 fprintf(fp
, ", R%d", data_reg
);
1907 void dump_instr(FILE *fp
, const struct bifrost_alu_inst
*instr
,
1908 struct bifrost_regs next_regs
, uint64_t *consts
,
1909 unsigned data_reg
, unsigned offset
, bool verbose
)
1911 struct bifrost_regs regs
;
1912 memcpy((char *) ®s
, (char *) &instr
->reg_bits
, sizeof(regs
));
1915 fprintf(fp
, "# regs: %016" PRIx64
"\n", instr
->reg_bits
);
1916 dump_regs(fp
, regs
);
1918 dump_fma(fp
, instr
->fma_bits
, regs
, next_regs
, consts
, verbose
);
1919 dump_add(fp
, instr
->add_bits
, regs
, next_regs
, consts
, data_reg
, offset
, verbose
);
1922 bool dump_clause(FILE *fp
, uint32_t *words
, unsigned *size
, unsigned offset
, bool verbose
)
1924 // State for a decoded clause
1925 struct bifrost_alu_inst instrs
[8] = {};
1926 uint64_t consts
[6] = {};
1927 unsigned num_instrs
= 0;
1928 unsigned num_consts
= 0;
1929 uint64_t header_bits
= 0;
1930 bool stopbit
= false;
1933 for (i
= 0; ; i
++, words
+= 4) {
1936 for (int j
= 0; j
< 4; j
++)
1937 fprintf(fp
, "%08x ", words
[3 - j
]); // low bit on the right
1940 unsigned tag
= bits(words
[0], 0, 8);
1942 // speculatively decode some things that are common between many formats, so we can share some code
1943 struct bifrost_alu_inst main_instr
= {};
1945 main_instr
.add_bits
= bits(words
[2], 2, 32 - 13);
1947 main_instr
.fma_bits
= bits(words
[1], 11, 32) | bits(words
[2], 0, 2) << (32 - 11);
1949 main_instr
.reg_bits
= ((uint64_t) bits(words
[1], 0, 11)) << 24 | (uint64_t) bits(words
[0], 8, 32);
1951 uint64_t const0
= bits(words
[0], 8, 32) << 4 | (uint64_t) words
[1] << 28 | bits(words
[2], 0, 4) << 60;
1952 uint64_t const1
= bits(words
[2], 4, 32) << 4 | (uint64_t) words
[3] << 32;
1954 bool stop
= tag
& 0x40;
1957 fprintf(fp
, "# tag: 0x%02x\n", tag
);
1960 unsigned idx
= stop
? 5 : 2;
1961 main_instr
.add_bits
|= ((tag
>> 3) & 0x7) << 17;
1962 instrs
[idx
+ 1] = main_instr
;
1963 instrs
[idx
].add_bits
= bits(words
[3], 0, 17) | ((tag
& 0x7) << 17);
1964 instrs
[idx
].fma_bits
|= bits(words
[2], 19, 32) << 10;
1965 consts
[0] = bits(words
[3], 17, 32) << 4;
1968 switch ((tag
>> 3) & 0x7) {
1970 switch (tag
& 0x7) {
1972 main_instr
.add_bits
|= bits(words
[3], 29, 32) << 17;
1973 instrs
[1] = main_instr
;
1978 instrs
[2].add_bits
= bits(words
[3], 0, 17) | bits(words
[3], 29, 32) << 17;
1979 instrs
[2].fma_bits
|= bits(words
[2], 19, 32) << 10;
1987 instrs
[2].add_bits
= bits(words
[3], 0, 17) | bits(words
[3], 29, 32) << 17;
1988 instrs
[2].fma_bits
|= bits(words
[2], 19, 32) << 10;
1989 main_instr
.add_bits
|= bits(words
[3], 26, 29) << 17;
1990 instrs
[3] = main_instr
;
1991 if ((tag
& 0x7) == 0x5) {
1997 instrs
[5].add_bits
= bits(words
[3], 0, 17) | bits(words
[3], 29, 32) << 17;
1998 instrs
[5].fma_bits
|= bits(words
[2], 19, 32) << 10;
2005 instrs
[5].add_bits
= bits(words
[3], 0, 17) | bits(words
[3], 29, 32) << 17;
2006 instrs
[5].fma_bits
|= bits(words
[2], 19, 32) << 10;
2007 main_instr
.add_bits
|= bits(words
[3], 26, 29) << 17;
2008 instrs
[6] = main_instr
;
2013 fprintf(fp
, "unknown tag bits 0x%02x\n", tag
);
2018 unsigned idx
= ((tag
>> 3) & 0x7) == 2 ? 4 : 7;
2019 main_instr
.add_bits
|= (tag
& 0x7) << 17;
2020 instrs
[idx
] = main_instr
;
2021 consts
[0] |= (bits(words
[2], 19, 32) | ((uint64_t) words
[3] << 13)) << 19;
2023 num_instrs
= idx
+ 1;
2028 unsigned idx
= stop
? 4 : 1;
2029 main_instr
.add_bits
|= (tag
& 0x7) << 17;
2030 instrs
[idx
] = main_instr
;
2031 instrs
[idx
+ 1].fma_bits
|= bits(words
[3], 22, 32);
2032 instrs
[idx
+ 1].reg_bits
= bits(words
[2], 19, 32) | (bits(words
[3], 0, 22) << (32 - 19));
2036 // only constants can come after this
2040 header_bits
= bits(words
[2], 19, 32) | ((uint64_t) words
[3] << (32 - 19));
2041 main_instr
.add_bits
|= (tag
& 0x7) << 17;
2042 instrs
[0] = main_instr
;
2046 unsigned pos
= tag
& 0xf;
2047 // note that `pos' encodes both the total number of
2048 // instructions and the position in the constant stream,
2049 // presumably because decoded constants and instructions
2050 // share a buffer in the decoder, but we only care about
2051 // the position in the constant stream; the total number of
2052 // instructions is redundant.
2053 unsigned const_idx
= 0;
2080 fprintf(fp
, "# unknown pos 0x%x\n", pos
);
2084 if (num_consts
< const_idx
+ 2)
2085 num_consts
= const_idx
+ 2;
2087 consts
[const_idx
] = const0
;
2088 consts
[const_idx
+ 1] = const1
;
2104 fprintf(fp
, "# header: %012" PRIx64
"\n", header_bits
);
2107 struct bifrost_header header
;
2108 memcpy((char *) &header
, (char *) &header_bits
, sizeof(struct bifrost_header
));
2109 dump_header(fp
, header
, verbose
);
2110 if (!header
.no_end_of_shader
)
2114 for (i
= 0; i
< num_instrs
; i
++) {
2115 struct bifrost_regs next_regs
;
2116 if (i
+ 1 == num_instrs
) {
2117 memcpy((char *) &next_regs
, (char *) &instrs
[0].reg_bits
,
2120 memcpy((char *) &next_regs
, (char *) &instrs
[i
+ 1].reg_bits
,
2124 dump_instr(fp
, &instrs
[i
], next_regs
, consts
, header
.datareg
, offset
, verbose
);
2129 for (unsigned i
= 0; i
< num_consts
; i
++) {
2130 fprintf(fp
, "# const%d: %08" PRIx64
"\n", 2 * i
, consts
[i
] & 0xffffffff);
2131 fprintf(fp
, "# const%d: %08" PRIx64
"\n", 2 * i
+ 1, consts
[i
] >> 32);
2137 void disassemble_bifrost(FILE *fp
, uint8_t *code
, size_t size
, bool verbose
)
2139 uint32_t *words
= (uint32_t *) code
;
2140 uint32_t *words_end
= words
+ (size
/ 4);
2141 // used for displaying branch targets
2142 unsigned offset
= 0;
2143 while (words
!= words_end
) {
2144 // we don't know what the program-end bit is quite yet, so for now just
2145 // assume that an all-0 quadword is padding
2146 uint32_t zero
[4] = {};
2147 if (memcmp(words
, zero
, 4 * sizeof(uint32_t)) == 0)
2149 fprintf(fp
, "clause_%d:\n", offset
);
2151 if (dump_clause(fp
, words
, &size
, offset
, verbose
) == true) {