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 "bifrost_ops.h"
35 #include "disassemble.h"
36 #include "util/macros.h"
38 // return bits (high, lo]
39 static uint64_t bits(uint32_t word
, unsigned lo
, unsigned high
)
43 return (word
& ((1 << high
) - 1)) >> lo
;
46 // each of these structs represents an instruction that's dispatched in one
47 // cycle. Note that these instructions are packed in funny ways within the
48 // clause, hence the need for a separate struct.
49 struct bifrost_alu_inst
{
56 unsigned uniform_const
: 8;
64 static unsigned get_reg0(struct bifrost_regs regs
)
67 return regs
.reg0
| ((regs
.reg1
& 0x1) << 5);
69 return regs
.reg0
<= regs
.reg1
? regs
.reg0
: 63 - regs
.reg0
;
72 static unsigned get_reg1(struct bifrost_regs regs
)
74 return regs
.reg0
<= regs
.reg1
? regs
.reg1
: 63 - regs
.reg1
;
77 enum bifrost_reg_write_unit
{
78 REG_WRITE_NONE
= 0, // don't write
79 REG_WRITE_TWO
, // write using reg2
80 REG_WRITE_THREE
, // write using reg3
83 // this represents the decoded version of the ctrl register field.
84 struct bifrost_reg_ctrl
{
88 enum bifrost_reg_write_unit fma_write_unit
;
89 enum bifrost_reg_write_unit add_write_unit
;
114 enum fma_src_type src_type
;
128 ADD_TEX_COMPACT
, // texture instruction with embedded sampler
129 ADD_TEX
, // texture instruction with sampler/etc. in uniform port
140 enum add_src_type src_type
;
144 struct bifrost_tex_ctrl
{
145 unsigned sampler_index
: 4; // also used to signal indirects
146 unsigned tex_index
: 7;
147 bool no_merge_index
: 1; // whether to merge (direct) sampler & texture indices
148 bool filter
: 1; // use the usual filtering pipeline (0 for texelFetch & textureGather)
150 bool texel_offset
: 1; // *Offset()
153 unsigned tex_type
: 2; // 2D, 3D, Cube, Buffer
154 bool compute_lod
: 1; // 0 for *Lod()
155 bool not_supply_lod
: 1; // 0 for *Lod() or when a bias is applied
156 bool calc_gradients
: 1; // 0 for *Grad()
158 unsigned result_type
: 4; // integer, unsigned, float TODO: why is this 4 bits?
162 struct bifrost_dual_tex_ctrl
{
163 unsigned sampler_index0
: 2;
165 unsigned tex_index0
: 2;
166 unsigned sampler_index1
: 2;
167 unsigned tex_index1
: 2;
171 enum branch_bit_size
{
175 // For the above combinations of bitsize and location, an extra bit is
176 // encoded via comparing the sources. The only possible source of ambiguity
177 // would be if the sources were the same, but then the branch condition
178 // would be always true or always false anyways, so we can ignore it. But
179 // this no longer works when comparing the y component to the x component,
180 // since it's valid to compare the y component of a source against its own
181 // x component. Instead, the extra bit is encoded via an extra bitsize.
184 BR_SIZE_32_AND_16X
= 5,
185 BR_SIZE_32_AND_16Y
= 6,
186 // Used for comparisons with zero and always-true, see below. I think this
187 // only works for integer comparisons.
191 void dump_header(FILE *fp
, struct bifrost_header header
, bool verbose
);
192 void dump_instr(FILE *fp
, const struct bifrost_alu_inst
*instr
,
193 struct bifrost_regs next_regs
, uint64_t *consts
,
194 unsigned data_reg
, unsigned offset
, bool verbose
);
195 bool dump_clause(FILE *fp
, uint32_t *words
, unsigned *size
, unsigned offset
, bool verbose
);
197 void dump_header(FILE *fp
, struct bifrost_header header
, bool verbose
)
199 if (header
.clause_type
!= 0) {
200 fprintf(fp
, "id(%du) ", header
.scoreboard_index
);
203 if (header
.scoreboard_deps
!= 0) {
204 fprintf(fp
, "next-wait(");
206 for (unsigned i
= 0; i
< 8; i
++) {
207 if (header
.scoreboard_deps
& (1 << i
)) {
211 fprintf(fp
, "%d", i
);
218 if (header
.datareg_writebarrier
)
219 fprintf(fp
, "data-reg-barrier ");
221 if (!header
.no_end_of_shader
)
224 if (!header
.back_to_back
) {
226 if (header
.branch_cond
)
227 fprintf(fp
, "branch-cond ");
229 fprintf(fp
, "branch-uncond ");
232 if (header
.elide_writes
)
235 if (header
.suppress_inf
)
236 fprintf(fp
, "suppress-inf ");
237 if (header
.suppress_nan
)
238 fprintf(fp
, "suppress-nan ");
241 fprintf(fp
, "unk0 ");
243 fprintf(fp
, "unk1 ");
245 fprintf(fp
, "unk2 ");
247 fprintf(fp
, "unk3 ");
249 fprintf(fp
, "unk4 ");
254 fprintf(fp
, "# clause type %d, next clause type %d\n",
255 header
.clause_type
, header
.next_clause_type
);
259 static struct bifrost_reg_ctrl
DecodeRegCtrl(FILE *fp
, struct bifrost_regs regs
)
261 struct bifrost_reg_ctrl decoded
= {};
263 if (regs
.ctrl
== 0) {
264 ctrl
= regs
.reg1
>> 2;
265 decoded
.read_reg0
= !(regs
.reg1
& 0x2);
266 decoded
.read_reg1
= false;
269 decoded
.read_reg0
= decoded
.read_reg1
= true;
273 decoded
.fma_write_unit
= REG_WRITE_TWO
;
277 decoded
.fma_write_unit
= REG_WRITE_TWO
;
278 decoded
.read_reg3
= true;
281 decoded
.read_reg3
= true;
284 decoded
.add_write_unit
= REG_WRITE_TWO
;
287 decoded
.add_write_unit
= REG_WRITE_TWO
;
288 decoded
.read_reg3
= true;
291 decoded
.clause_start
= true;
294 decoded
.fma_write_unit
= REG_WRITE_TWO
;
295 decoded
.clause_start
= true;
300 decoded
.read_reg3
= true;
301 decoded
.clause_start
= true;
304 decoded
.add_write_unit
= REG_WRITE_TWO
;
305 decoded
.clause_start
= true;
310 decoded
.fma_write_unit
= REG_WRITE_THREE
;
311 decoded
.add_write_unit
= REG_WRITE_TWO
;
314 fprintf(fp
, "# unknown reg ctrl %d\n", ctrl
);
320 // Pass in the add_write_unit or fma_write_unit, and this returns which register
321 // the ADD/FMA units are writing to
322 static unsigned GetRegToWrite(enum bifrost_reg_write_unit unit
, struct bifrost_regs regs
)
327 case REG_WRITE_THREE
:
329 default: /* REG_WRITE_NONE */
335 static void dump_regs(FILE *fp
, struct bifrost_regs srcs
)
337 struct bifrost_reg_ctrl ctrl
= DecodeRegCtrl(fp
, srcs
);
340 fprintf(fp
, "port 0: R%d ", get_reg0(srcs
));
342 fprintf(fp
, "port 1: R%d ", get_reg1(srcs
));
344 if (ctrl
.fma_write_unit
== REG_WRITE_TWO
)
345 fprintf(fp
, "port 2: R%d (write FMA) ", srcs
.reg2
);
346 else if (ctrl
.add_write_unit
== REG_WRITE_TWO
)
347 fprintf(fp
, "port 2: R%d (write ADD) ", srcs
.reg2
);
349 if (ctrl
.fma_write_unit
== REG_WRITE_THREE
)
350 fprintf(fp
, "port 3: R%d (write FMA) ", srcs
.reg3
);
351 else if (ctrl
.add_write_unit
== REG_WRITE_THREE
)
352 fprintf(fp
, "port 3: R%d (write ADD) ", srcs
.reg3
);
353 else if (ctrl
.read_reg3
)
354 fprintf(fp
, "port 3: R%d (read) ", srcs
.reg3
);
356 if (srcs
.uniform_const
) {
357 if (srcs
.uniform_const
& 0x80) {
358 fprintf(fp
, "uniform: U%d", (srcs
.uniform_const
& 0x7f) * 2);
364 static void dump_const_imm(FILE *fp
, uint32_t imm
)
371 fprintf(fp
, "0x%08x /* %f */", imm
, fi
.f
);
374 static uint64_t get_const(uint64_t *consts
, struct bifrost_regs srcs
)
376 unsigned low_bits
= srcs
.uniform_const
& 0xf;
378 switch (srcs
.uniform_const
>> 4) {
401 return imm
| low_bits
;
404 static void dump_uniform_const_src(FILE *fp
, struct bifrost_regs srcs
, uint64_t *consts
, bool high32
)
406 if (srcs
.uniform_const
& 0x80) {
407 unsigned uniform
= (srcs
.uniform_const
& 0x7f) * 2;
408 fprintf(fp
, "U%d", uniform
+ (high32
? 1 : 0));
409 } else if (srcs
.uniform_const
>= 0x20) {
410 uint64_t imm
= get_const(consts
, srcs
);
412 dump_const_imm(fp
, imm
>> 32);
414 dump_const_imm(fp
, imm
);
416 switch (srcs
.uniform_const
) {
421 fprintf(fp
, "atest-data");
424 fprintf(fp
, "sample-ptr");
434 fprintf(fp
, "blend-descriptor%u", (unsigned) srcs
.uniform_const
- 8);
437 fprintf(fp
, "unkConst%u", (unsigned) srcs
.uniform_const
);
448 static void dump_src(FILE *fp
, unsigned src
, struct bifrost_regs srcs
, uint64_t *consts
, bool isFMA
)
452 fprintf(fp
, "R%d", get_reg0(srcs
));
455 fprintf(fp
, "R%d", get_reg1(srcs
));
458 fprintf(fp
, "R%d", srcs
.reg3
);
464 fprintf(fp
, "T"); // i.e. the output of FMA this cycle
467 dump_uniform_const_src(fp
, srcs
, consts
, false);
470 dump_uniform_const_src(fp
, srcs
, consts
, true);
481 static void dump_output_mod(FILE *fp
, unsigned mod
)
487 fprintf(fp
, ".clamp_0_inf");
488 break; // max(out, 0)
490 fprintf(fp
, ".clamp_m1_1");
491 break; // clamp(out, -1, 1)
493 fprintf(fp
, ".clamp_0_1");
494 break; // clamp(out, 0, 1)
500 static void dump_minmax_mode(FILE *fp
, unsigned mod
)
504 /* Same as fmax() and fmin() -- return the other number if any
505 * number is NaN. Also always return +0 if one argument is +0 and
510 /* Instead of never returning a NaN, always return one. The
511 * "greater"/"lesser" NaN is always returned, first by checking the
512 * sign and then the mantissa bits.
514 fprintf(fp
, ".nan_wins");
517 /* For max, implement src0 > src1 ? src0 : src1
518 * For min, implement src0 < src1 ? src0 : src1
520 * This includes handling NaN's and signedness of 0 differently
521 * from above, since +0 and -0 compare equal and comparisons always
522 * return false for NaN's. As a result, this mode is *not*
525 fprintf(fp
, ".src1_wins");
528 /* For max, implement src0 < src1 ? src1 : src0
529 * For min, implement src0 > src1 ? src1 : src0
531 fprintf(fp
, ".src0_wins");
538 static void dump_round_mode(FILE *fp
, unsigned mod
)
542 /* roundTiesToEven, the IEEE default. */
545 /* roundTowardPositive in the IEEE spec. */
546 fprintf(fp
, ".round_pos");
549 /* roundTowardNegative in the IEEE spec. */
550 fprintf(fp
, ".round_neg");
553 /* roundTowardZero in the IEEE spec. */
554 fprintf(fp
, ".round_zero");
562 csel_cond_name(enum bifrost_csel_cond cond
)
565 case BIFROST_FEQ_F
: return "feq.f";
566 case BIFROST_FGT_F
: return "fgt.f";
567 case BIFROST_FGE_F
: return "fge.f";
568 case BIFROST_IEQ_F
: return "ieq.f";
569 case BIFROST_IGT_I
: return "igt.i";
570 case BIFROST_IGE_I
: return "uge.i";
571 case BIFROST_UGT_I
: return "ugt.i";
572 case BIFROST_UGE_I
: return "uge.i";
573 default: return "invalid";
577 static const struct fma_op_info FMAOpInfos
[] = {
578 { 0x00000, "FMA.f32", FMA_FMA
},
579 { 0x40000, "MAX.f32", FMA_FMINMAX
},
580 { 0x44000, "MIN.f32", FMA_FMINMAX
},
581 { 0x48000, "FCMP.GL", FMA_FCMP
},
582 { 0x4c000, "FCMP.D3D", FMA_FCMP
},
583 { 0x4ff98, "ADD.i32", FMA_TWO_SRC
},
584 { 0x4ffd8, "SUB.i32", FMA_TWO_SRC
},
585 { 0x4fff0, "SUBB.i32", FMA_TWO_SRC
},
586 { 0x50000, "FMA_MSCALE", FMA_FMA_MSCALE
},
587 { 0x58000, "ADD.f32", FMA_FADD
},
588 { 0x5c000, "CSEL4", FMA_CSEL4
},
589 { 0x5d8d0, "ICMP.D3D.GT.v2i16", FMA_TWO_SRC
},
590 { 0x5d9d0, "UCMP.D3D.GT.v2i16", FMA_TWO_SRC
},
591 { 0x5dad0, "ICMP.D3D.GE.v2i16", FMA_TWO_SRC
},
592 { 0x5dbd0, "UCMP.D3D.GE.v2i16", FMA_TWO_SRC
},
593 { 0x5dcd0, "ICMP.D3D.EQ.v2i16", FMA_TWO_SRC
},
594 { 0x5de40, "ICMP.GL.GT.i32", FMA_TWO_SRC
}, // src0 > src1 ? 1 : 0
595 { 0x5de48, "ICMP.GL.GE.i32", FMA_TWO_SRC
},
596 { 0x5de50, "UCMP.GL.GT.i32", FMA_TWO_SRC
},
597 { 0x5de58, "UCMP.GL.GE.i32", FMA_TWO_SRC
},
598 { 0x5de60, "ICMP.GL.EQ.i32", FMA_TWO_SRC
},
599 { 0x5dec0, "ICMP.D3D.GT.i32", FMA_TWO_SRC
}, // src0 > src1 ? ~0 : 0
600 { 0x5dec8, "ICMP.D3D.GE.i32", FMA_TWO_SRC
},
601 { 0x5ded0, "UCMP.D3D.GT.i32", FMA_TWO_SRC
},
602 { 0x5ded8, "UCMP.D3D.GE.i32", FMA_TWO_SRC
},
603 { 0x5dee0, "ICMP.D3D.EQ.i32", FMA_TWO_SRC
},
604 { 0x60000, "RSHIFT_NAND", FMA_SHIFT
},
605 { 0x61000, "RSHIFT_AND", FMA_SHIFT
},
606 { 0x62000, "LSHIFT_NAND", FMA_SHIFT
},
607 { 0x63000, "LSHIFT_AND", FMA_SHIFT
}, // (src0 << src2) & src1
608 { 0x64000, "RSHIFT_XOR", FMA_SHIFT
},
609 { 0x65200, "LSHIFT_ADD.i32", FMA_THREE_SRC
},
610 { 0x65600, "LSHIFT_SUB.i32", FMA_THREE_SRC
}, // (src0 << src2) - src1
611 { 0x65a00, "LSHIFT_RSUB.i32", FMA_THREE_SRC
}, // src1 - (src0 << src2)
612 { 0x65e00, "RSHIFT_ADD.i32", FMA_THREE_SRC
},
613 { 0x66200, "RSHIFT_SUB.i32", FMA_THREE_SRC
},
614 { 0x66600, "RSHIFT_RSUB.i32", FMA_THREE_SRC
},
615 { 0x66a00, "ARSHIFT_ADD.i32", FMA_THREE_SRC
},
616 { 0x66e00, "ARSHIFT_SUB.i32", FMA_THREE_SRC
},
617 { 0x67200, "ARSHIFT_RSUB.i32", FMA_THREE_SRC
},
618 { 0x80000, "FMA.v2f16", FMA_FMA16
},
619 { 0xc0000, "MAX.v2f16", FMA_FMINMAX16
},
620 { 0xc4000, "MIN.v2f16", FMA_FMINMAX16
},
621 { 0xc8000, "FCMP.GL", FMA_FCMP16
},
622 { 0xcc000, "FCMP.D3D", FMA_FCMP16
},
623 { 0xcf900, "ADD.v2i16", FMA_TWO_SRC
},
624 { 0xcfc10, "ADDC.i32", FMA_TWO_SRC
},
625 { 0xcfd80, "ADD.i32.i16.X", FMA_TWO_SRC
},
626 { 0xcfd90, "ADD.i32.u16.X", FMA_TWO_SRC
},
627 { 0xcfdc0, "ADD.i32.i16.Y", FMA_TWO_SRC
},
628 { 0xcfdd0, "ADD.i32.u16.Y", FMA_TWO_SRC
},
629 { 0xd8000, "ADD.v2f16", FMA_FADD16
},
630 { 0xdc000, "CSEL4.v16", FMA_CSEL4
},
631 { 0xdd000, "F32_TO_F16", FMA_TWO_SRC
},
632 { 0xe0046, "F16_TO_I16.XX", FMA_ONE_SRC
},
633 { 0xe0047, "F16_TO_U16.XX", FMA_ONE_SRC
},
634 { 0xe004e, "F16_TO_I16.YX", FMA_ONE_SRC
},
635 { 0xe004f, "F16_TO_U16.YX", FMA_ONE_SRC
},
636 { 0xe0056, "F16_TO_I16.XY", FMA_ONE_SRC
},
637 { 0xe0057, "F16_TO_U16.XY", FMA_ONE_SRC
},
638 { 0xe005e, "F16_TO_I16.YY", FMA_ONE_SRC
},
639 { 0xe005f, "F16_TO_U16.YY", FMA_ONE_SRC
},
640 { 0xe00c0, "I16_TO_F16.XX", FMA_ONE_SRC
},
641 { 0xe00c1, "U16_TO_F16.XX", FMA_ONE_SRC
},
642 { 0xe00c8, "I16_TO_F16.YX", FMA_ONE_SRC
},
643 { 0xe00c9, "U16_TO_F16.YX", FMA_ONE_SRC
},
644 { 0xe00d0, "I16_TO_F16.XY", FMA_ONE_SRC
},
645 { 0xe00d1, "U16_TO_F16.XY", FMA_ONE_SRC
},
646 { 0xe00d8, "I16_TO_F16.YY", FMA_ONE_SRC
},
647 { 0xe00d9, "U16_TO_F16.YY", FMA_ONE_SRC
},
648 { 0xe0136, "F32_TO_I32", FMA_ONE_SRC
},
649 { 0xe0137, "F32_TO_U32", FMA_ONE_SRC
},
650 { 0xe0178, "I32_TO_F32", FMA_ONE_SRC
},
651 { 0xe0179, "U32_TO_F32", FMA_ONE_SRC
},
652 { 0xe0198, "I16_TO_I32.X", FMA_ONE_SRC
},
653 { 0xe0199, "U16_TO_U32.X", FMA_ONE_SRC
},
654 { 0xe019a, "I16_TO_I32.Y", FMA_ONE_SRC
},
655 { 0xe019b, "U16_TO_U32.Y", FMA_ONE_SRC
},
656 { 0xe019c, "I16_TO_F32.X", FMA_ONE_SRC
},
657 { 0xe019d, "U16_TO_F32.X", FMA_ONE_SRC
},
658 { 0xe019e, "I16_TO_F32.Y", FMA_ONE_SRC
},
659 { 0xe019f, "U16_TO_F32.Y", FMA_ONE_SRC
},
660 { 0xe01a2, "F16_TO_F32.X", FMA_ONE_SRC
},
661 { 0xe01a3, "F16_TO_F32.Y", FMA_ONE_SRC
},
662 { 0xe032c, "NOP", FMA_ONE_SRC
},
663 { 0xe032d, "MOV", FMA_ONE_SRC
},
664 { 0xe032f, "SWZ.YY.v2i16", FMA_ONE_SRC
},
665 { 0xe0345, "LOG_FREXPM", FMA_ONE_SRC
},
666 { 0xe0365, "FRCP_FREXPM", FMA_ONE_SRC
},
667 { 0xe0375, "FSQRT_FREXPM", FMA_ONE_SRC
},
668 { 0xe038d, "FRCP_FREXPE", FMA_ONE_SRC
},
669 { 0xe03a5, "FSQRT_FREXPE", FMA_ONE_SRC
},
670 { 0xe03ad, "FRSQ_FREXPE", FMA_ONE_SRC
},
671 { 0xe03c5, "LOG_FREXPE", FMA_ONE_SRC
},
672 { 0xe03fa, "CLZ", FMA_ONE_SRC
},
673 { 0xe0b80, "IMAX3", FMA_THREE_SRC
},
674 { 0xe0bc0, "UMAX3", FMA_THREE_SRC
},
675 { 0xe0c00, "IMIN3", FMA_THREE_SRC
},
676 { 0xe0c40, "UMIN3", FMA_THREE_SRC
},
677 { 0xe0ec5, "ROUND", FMA_ONE_SRC
},
678 { 0xe0f40, "CSEL", FMA_THREE_SRC
}, // src2 != 0 ? src1 : src0
679 { 0xe0fc0, "MUX.i32", FMA_THREE_SRC
}, // see ADD comment
680 { 0xe1805, "ROUNDEVEN", FMA_ONE_SRC
},
681 { 0xe1845, "CEIL", FMA_ONE_SRC
},
682 { 0xe1885, "FLOOR", FMA_ONE_SRC
},
683 { 0xe18c5, "TRUNC", FMA_ONE_SRC
},
684 { 0xe19b0, "ATAN_LDEXP.Y.f32", FMA_TWO_SRC
},
685 { 0xe19b8, "ATAN_LDEXP.X.f32", FMA_TWO_SRC
},
686 { 0xe1c80, "LSHIFT_ADD_LOW32.u32", FMA_SHIFT_ADD64
},
687 { 0xe1cc0, "LSHIFT_ADD_LOW32.i64", FMA_SHIFT_ADD64
},
688 { 0xe1d80, "LSHIFT_ADD_LOW32.i32", FMA_SHIFT_ADD64
},
689 { 0xe1e00, "SEL.XX.i16", FMA_TWO_SRC
},
690 { 0xe1e08, "SEL.YX.i16", FMA_TWO_SRC
},
691 { 0xe1e10, "SEL.XY.i16", FMA_TWO_SRC
},
692 { 0xe1e18, "SEL.YY.i16", FMA_TWO_SRC
},
693 { 0xe7800, "IMAD", FMA_THREE_SRC
},
694 { 0xe78db, "POPCNT", FMA_ONE_SRC
},
697 static struct fma_op_info
find_fma_op_info(unsigned op
)
699 for (unsigned i
= 0; i
< ARRAY_SIZE(FMAOpInfos
); i
++) {
701 switch (FMAOpInfos
[i
].src_type
) {
710 opCmp
= op
& ~0x1fff;
713 case FMA_SHIFT_ADD64
:
720 opCmp
= op
& ~0x3fff;
724 opCmp
= op
& ~0x3ffff;
731 opCmp
= op
& ~0x7fff;
737 if (FMAOpInfos
[i
].op
== opCmp
)
738 return FMAOpInfos
[i
];
741 struct fma_op_info info
;
742 snprintf(info
.name
, sizeof(info
.name
), "op%04x", op
);
744 info
.src_type
= FMA_THREE_SRC
;
748 static void dump_fcmp(FILE *fp
, unsigned op
)
770 fprintf(fp
, ".unk%d", op
);
775 static void dump_16swizzle(FILE *fp
, unsigned swiz
)
779 fprintf(fp
, ".%c%c", "xy"[swiz
& 1], "xy"[(swiz
>> 1) & 1]);
782 static void dump_fma_expand_src0(FILE *fp
, unsigned ctrl
)
804 static void dump_fma_expand_src1(FILE *fp
, unsigned ctrl
)
826 static void dump_fma(FILE *fp
, uint64_t word
, struct bifrost_regs regs
, struct bifrost_regs next_regs
, uint64_t *consts
, bool verbose
)
829 fprintf(fp
, "# FMA: %016" PRIx64
"\n", word
);
831 struct bifrost_fma_inst FMA
;
832 memcpy((char *) &FMA
, (char *) &word
, sizeof(struct bifrost_fma_inst
));
833 struct fma_op_info info
= find_fma_op_info(FMA
.op
);
835 fprintf(fp
, "%s", info
.name
);
836 if (info
.src_type
== FMA_FADD
||
837 info
.src_type
== FMA_FMINMAX
||
838 info
.src_type
== FMA_FMA
||
839 info
.src_type
== FMA_FADD16
||
840 info
.src_type
== FMA_FMINMAX16
||
841 info
.src_type
== FMA_FMA16
) {
842 dump_output_mod(fp
, bits(FMA
.op
, 12, 14));
843 switch (info
.src_type
) {
848 dump_round_mode(fp
, bits(FMA
.op
, 10, 12));
852 dump_minmax_mode(fp
, bits(FMA
.op
, 10, 12));
857 } else if (info
.src_type
== FMA_FCMP
|| info
.src_type
== FMA_FCMP16
) {
858 dump_fcmp(fp
, bits(FMA
.op
, 10, 13));
859 if (info
.src_type
== FMA_FCMP
)
862 fprintf(fp
, ".v2f16");
863 } else if (info
.src_type
== FMA_FMA_MSCALE
) {
864 if (FMA
.op
& (1 << 11)) {
865 switch ((FMA
.op
>> 9) & 0x3) {
867 /* This mode seems to do a few things:
868 * - Makes 0 * infinity (and incidentally 0 * nan) return 0,
869 * since generating a nan would poison the result of
870 * 1/infinity and 1/0.
871 * - Fiddles with which nan is returned in nan * nan,
872 * presumably to make sure that the same exact nan is
873 * returned for 1/nan.
875 fprintf(fp
, ".rcp_mode");
878 /* Similar to the above, but src0 always wins when multiplying
881 fprintf(fp
, ".sqrt_mode");
884 fprintf(fp
, ".unk%d_mode", (int) (FMA
.op
>> 9) & 0x3);
887 dump_output_mod(fp
, bits(FMA
.op
, 9, 11));
889 } else if (info
.src_type
== FMA_SHIFT
) {
890 struct bifrost_shift_fma shift
;
891 memcpy(&shift
, &FMA
, sizeof(shift
));
893 if (shift
.half
== 0x7)
894 fprintf(fp
, ".v2i16");
895 else if (shift
.half
== 0)
897 else if (shift
.half
== 0x4)
898 fprintf(fp
, ".v4i8");
900 fprintf(fp
, ".unk%u", shift
.half
);
903 fprintf(fp
, ".no_unk");
906 fprintf(fp
, ".invert_1");
909 fprintf(fp
, ".invert_2");
914 struct bifrost_reg_ctrl next_ctrl
= DecodeRegCtrl(fp
, next_regs
);
915 if (next_ctrl
.fma_write_unit
!= REG_WRITE_NONE
) {
916 fprintf(fp
, "{R%d, T0}, ", GetRegToWrite(next_ctrl
.fma_write_unit
, next_regs
));
921 switch (info
.src_type
) {
923 dump_src(fp
, FMA
.src0
, regs
, consts
, true);
926 dump_src(fp
, FMA
.src0
, regs
, consts
, true);
928 dump_src(fp
, FMA
.op
& 0x7, regs
, consts
, true);
936 dump_src(fp
, FMA
.src0
, regs
, consts
, true);
937 dump_fma_expand_src0(fp
, (FMA
.op
>> 6) & 0x7);
945 dump_src(fp
, FMA
.op
& 0x7, regs
, consts
, true);
946 dump_fma_expand_src1(fp
, (FMA
.op
>> 6) & 0x7);
951 case FMA_FMINMAX16
: {
952 bool abs1
= FMA
.op
& 0x8;
953 bool abs2
= (FMA
.op
& 0x7) < FMA
.src0
;
958 dump_src(fp
, FMA
.src0
, regs
, consts
, true);
959 dump_16swizzle(fp
, (FMA
.op
>> 6) & 0x3);
967 dump_src(fp
, FMA
.op
& 0x7, regs
, consts
, true);
968 dump_16swizzle(fp
, (FMA
.op
>> 8) & 0x3);
976 dump_src(fp
, FMA
.src0
, regs
, consts
, true);
977 dump_fma_expand_src0(fp
, (FMA
.op
>> 6) & 0x7);
985 dump_src(fp
, FMA
.op
& 0x7, regs
, consts
, true);
986 dump_fma_expand_src1(fp
, (FMA
.op
>> 6) & 0x7);
991 dump_src(fp
, FMA
.src0
, regs
, consts
, true);
992 // Note: this is kinda a guess, I haven't seen the blob set this to
993 // anything other than the identity, but it matches FMA_TWO_SRCFmod16
994 dump_16swizzle(fp
, (FMA
.op
>> 6) & 0x3);
996 dump_src(fp
, FMA
.op
& 0x7, regs
, consts
, true);
997 dump_16swizzle(fp
, (FMA
.op
>> 8) & 0x3);
999 case FMA_SHIFT_ADD64
:
1000 dump_src(fp
, FMA
.src0
, regs
, consts
, true);
1002 dump_src(fp
, FMA
.op
& 0x7, regs
, consts
, true);
1004 fprintf(fp
, "shift:%u", (FMA
.op
>> 3) & 0x7);
1007 dump_src(fp
, FMA
.src0
, regs
, consts
, true);
1009 dump_src(fp
, FMA
.op
& 0x7, regs
, consts
, true);
1011 dump_src(fp
, (FMA
.op
>> 3) & 0x7, regs
, consts
, true);
1014 struct bifrost_shift_fma shift
;
1015 memcpy(&shift
, &FMA
, sizeof(shift
));
1017 dump_src(fp
, shift
.src0
, regs
, consts
, true);
1019 dump_src(fp
, shift
.src1
, regs
, consts
, true);
1021 dump_src(fp
, shift
.src2
, regs
, consts
, true);
1025 if (FMA
.op
& (1 << 14))
1027 if (FMA
.op
& (1 << 9))
1028 fprintf(fp
, "abs(");
1029 dump_src(fp
, FMA
.src0
, regs
, consts
, true);
1030 dump_fma_expand_src0(fp
, (FMA
.op
>> 6) & 0x7);
1031 if (FMA
.op
& (1 << 9))
1034 if (FMA
.op
& (1 << 16))
1035 fprintf(fp
, "abs(");
1036 dump_src(fp
, FMA
.op
& 0x7, regs
, consts
, true);
1037 dump_fma_expand_src1(fp
, (FMA
.op
>> 6) & 0x7);
1038 if (FMA
.op
& (1 << 16))
1041 if (FMA
.op
& (1 << 15))
1043 if (FMA
.op
& (1 << 17))
1044 fprintf(fp
, "abs(");
1045 dump_src(fp
, (FMA
.op
>> 3) & 0x7, regs
, consts
, true);
1046 if (FMA
.op
& (1 << 17))
1050 if (FMA
.op
& (1 << 14))
1052 dump_src(fp
, FMA
.src0
, regs
, consts
, true);
1053 dump_16swizzle(fp
, (FMA
.op
>> 6) & 0x3);
1055 dump_src(fp
, FMA
.op
& 0x7, regs
, consts
, true);
1056 dump_16swizzle(fp
, (FMA
.op
>> 8) & 0x3);
1058 if (FMA
.op
& (1 << 15))
1060 dump_src(fp
, (FMA
.op
>> 3) & 0x7, regs
, consts
, true);
1061 dump_16swizzle(fp
, (FMA
.op
>> 16) & 0x3);
1064 struct bifrost_csel4 csel
;
1065 memcpy(&csel
, &FMA
, sizeof(csel
));
1066 fprintf(fp
, ".%s ", csel_cond_name(csel
.cond
));
1068 dump_src(fp
, csel
.src0
, regs
, consts
, true);
1070 dump_src(fp
, csel
.src1
, regs
, consts
, true);
1072 dump_src(fp
, csel
.src2
, regs
, consts
, true);
1074 dump_src(fp
, csel
.src3
, regs
, consts
, true);
1077 case FMA_FMA_MSCALE
:
1078 if (FMA
.op
& (1 << 12))
1079 fprintf(fp
, "abs(");
1080 dump_src(fp
, FMA
.src0
, regs
, consts
, true);
1081 if (FMA
.op
& (1 << 12))
1084 if (FMA
.op
& (1 << 13))
1086 dump_src(fp
, FMA
.op
& 0x7, regs
, consts
, true);
1088 if (FMA
.op
& (1 << 14))
1090 dump_src(fp
, (FMA
.op
>> 3) & 0x7, regs
, consts
, true);
1092 dump_src(fp
, (FMA
.op
>> 6) & 0x7, regs
, consts
, true);
1098 static const struct add_op_info add_op_infos
[] = {
1099 { 0x00000, "MAX.f32", ADD_FMINMAX
},
1100 { 0x02000, "MIN.f32", ADD_FMINMAX
},
1101 { 0x04000, "ADD.f32", ADD_FADD
},
1102 { 0x06000, "FCMP.GL", ADD_FCMP
},
1103 { 0x07000, "FCMP.D3D", ADD_FCMP
},
1104 { 0x07856, "F16_TO_I16", ADD_ONE_SRC
},
1105 { 0x07857, "F16_TO_U16", ADD_ONE_SRC
},
1106 { 0x078c0, "I16_TO_F16.XX", ADD_ONE_SRC
},
1107 { 0x078c1, "U16_TO_F16.XX", ADD_ONE_SRC
},
1108 { 0x078c8, "I16_TO_F16.YX", ADD_ONE_SRC
},
1109 { 0x078c9, "U16_TO_F16.YX", ADD_ONE_SRC
},
1110 { 0x078d0, "I16_TO_F16.XY", ADD_ONE_SRC
},
1111 { 0x078d1, "U16_TO_F16.XY", ADD_ONE_SRC
},
1112 { 0x078d8, "I16_TO_F16.YY", ADD_ONE_SRC
},
1113 { 0x078d9, "U16_TO_F16.YY", ADD_ONE_SRC
},
1114 { 0x07936, "F32_TO_I32", ADD_ONE_SRC
},
1115 { 0x07937, "F32_TO_U32", ADD_ONE_SRC
},
1116 { 0x07978, "I32_TO_F32", ADD_ONE_SRC
},
1117 { 0x07979, "U32_TO_F32", ADD_ONE_SRC
},
1118 { 0x07998, "I16_TO_I32.X", ADD_ONE_SRC
},
1119 { 0x07999, "U16_TO_U32.X", ADD_ONE_SRC
},
1120 { 0x0799a, "I16_TO_I32.Y", ADD_ONE_SRC
},
1121 { 0x0799b, "U16_TO_U32.Y", ADD_ONE_SRC
},
1122 { 0x0799c, "I16_TO_F32.X", ADD_ONE_SRC
},
1123 { 0x0799d, "U16_TO_F32.X", ADD_ONE_SRC
},
1124 { 0x0799e, "I16_TO_F32.Y", ADD_ONE_SRC
},
1125 { 0x0799f, "U16_TO_F32.Y", ADD_ONE_SRC
},
1126 { 0x079a2, "F16_TO_F32.X", ADD_ONE_SRC
},
1127 { 0x079a3, "F16_TO_F32.Y", ADD_ONE_SRC
},
1128 { 0x07b2b, "SWZ.YX.v2i16", ADD_ONE_SRC
},
1129 { 0x07b2c, "NOP", ADD_ONE_SRC
},
1130 { 0x07b29, "SWZ.XX.v2i16", ADD_ONE_SRC
},
1131 { 0x07b2d, "MOV", ADD_ONE_SRC
},
1132 { 0x07b2f, "SWZ.YY.v2i16", ADD_ONE_SRC
},
1133 { 0x07b65, "FRCP_FREXPM", ADD_ONE_SRC
},
1134 { 0x07b75, "FSQRT_FREXPM", ADD_ONE_SRC
},
1135 { 0x07b8d, "FRCP_FREXPE", ADD_ONE_SRC
},
1136 { 0x07ba5, "FSQRT_FREXPE", ADD_ONE_SRC
},
1137 { 0x07bad, "FRSQ_FREXPE", ADD_ONE_SRC
},
1138 { 0x07bc5, "FLOG_FREXPE", ADD_ONE_SRC
},
1139 { 0x07d45, "CEIL", ADD_ONE_SRC
},
1140 { 0x07d85, "FLOOR", ADD_ONE_SRC
},
1141 { 0x07dc5, "TRUNC", ADD_ONE_SRC
},
1142 { 0x07f18, "LSHIFT_ADD_HIGH32.i32", ADD_TWO_SRC
},
1143 { 0x08000, "LD_ATTR.f16", ADD_LOAD_ATTR
, true },
1144 { 0x08100, "LD_ATTR.v2f16", ADD_LOAD_ATTR
, true },
1145 { 0x08200, "LD_ATTR.v3f16", ADD_LOAD_ATTR
, true },
1146 { 0x08300, "LD_ATTR.v4f16", ADD_LOAD_ATTR
, true },
1147 { 0x08400, "LD_ATTR.f32", ADD_LOAD_ATTR
, true },
1148 { 0x08500, "LD_ATTR.v3f32", ADD_LOAD_ATTR
, true },
1149 { 0x08600, "LD_ATTR.v3f32", ADD_LOAD_ATTR
, true },
1150 { 0x08700, "LD_ATTR.v4f32", ADD_LOAD_ATTR
, true },
1151 { 0x08800, "LD_ATTR.i32", ADD_LOAD_ATTR
, true },
1152 { 0x08900, "LD_ATTR.v3i32", ADD_LOAD_ATTR
, true },
1153 { 0x08a00, "LD_ATTR.v3i32", ADD_LOAD_ATTR
, true },
1154 { 0x08b00, "LD_ATTR.v4i32", ADD_LOAD_ATTR
, true },
1155 { 0x08c00, "LD_ATTR.u32", ADD_LOAD_ATTR
, true },
1156 { 0x08d00, "LD_ATTR.v3u32", ADD_LOAD_ATTR
, true },
1157 { 0x08e00, "LD_ATTR.v3u32", ADD_LOAD_ATTR
, true },
1158 { 0x08f00, "LD_ATTR.v4u32", ADD_LOAD_ATTR
, true },
1159 { 0x0a000, "LD_VAR.32", ADD_VARYING_INTERP
, true },
1160 { 0x0b000, "TEX", ADD_TEX_COMPACT
, true },
1161 { 0x0c188, "LOAD.i32", ADD_TWO_SRC
, true },
1162 { 0x0c1a0, "LD_UBO.i32", ADD_TWO_SRC
, true },
1163 { 0x0c1b8, "LD_SCRATCH.v2i32", ADD_TWO_SRC
, true },
1164 { 0x0c1c8, "LOAD.v2i32", ADD_TWO_SRC
, true },
1165 { 0x0c1e0, "LD_UBO.v2i32", ADD_TWO_SRC
, true },
1166 { 0x0c1f8, "LD_SCRATCH.v2i32", ADD_TWO_SRC
, true },
1167 { 0x0c208, "LOAD.v4i32", ADD_TWO_SRC
, true },
1168 { 0x0c220, "LD_UBO.v4i32", ADD_TWO_SRC
, true },
1169 { 0x0c238, "LD_SCRATCH.v4i32", ADD_TWO_SRC
, true },
1170 { 0x0c248, "STORE.v4i32", ADD_TWO_SRC
, true },
1171 { 0x0c278, "ST_SCRATCH.v4i32", ADD_TWO_SRC
, true },
1172 { 0x0c588, "STORE.i32", ADD_TWO_SRC
, true },
1173 { 0x0c5b8, "ST_SCRATCH.i32", ADD_TWO_SRC
, true },
1174 { 0x0c5c8, "STORE.v2i32", ADD_TWO_SRC
, true },
1175 { 0x0c5f8, "ST_SCRATCH.v2i32", ADD_TWO_SRC
, true },
1176 { 0x0c648, "LOAD.u16", ADD_TWO_SRC
, true }, // zero-extends
1177 { 0x0ca88, "LOAD.v3i32", ADD_TWO_SRC
, true },
1178 { 0x0caa0, "LD_UBO.v3i32", ADD_TWO_SRC
, true },
1179 { 0x0cab8, "LD_SCRATCH.v3i32", ADD_TWO_SRC
, true },
1180 { 0x0cb88, "STORE.v3i32", ADD_TWO_SRC
, true },
1181 { 0x0cbb8, "ST_SCRATCH.v3i32", ADD_TWO_SRC
, true },
1182 { 0x0cc00, "FRCP_FAST.f32", ADD_ONE_SRC
},
1183 { 0x0cc20, "FRSQ_FAST.f32", ADD_ONE_SRC
},
1184 { 0x0ce00, "FRCP_TABLE", ADD_ONE_SRC
},
1185 { 0x0ce10, "FRCP_FAST.f16.X", ADD_ONE_SRC
},
1186 { 0x0ce20, "FRSQ_TABLE", ADD_ONE_SRC
},
1187 { 0x0ce30, "FRCP_FAST.f16.Y", ADD_ONE_SRC
},
1188 { 0x0ce50, "FRSQ_FAST.f16.X", ADD_ONE_SRC
},
1189 { 0x0ce60, "FRCP_APPROX", ADD_ONE_SRC
},
1190 { 0x0ce70, "FRSQ_FAST.f16.Y", ADD_ONE_SRC
},
1191 { 0x0cf40, "ATAN_ASSIST", ADD_TWO_SRC
},
1192 { 0x0cf48, "ATAN_TABLE", ADD_TWO_SRC
},
1193 { 0x0cf50, "SIN_TABLE", ADD_ONE_SRC
},
1194 { 0x0cf51, "COS_TABLE", ADD_ONE_SRC
},
1195 { 0x0cf58, "EXP_TABLE", ADD_ONE_SRC
},
1196 { 0x0cf60, "FLOG2_TABLE", ADD_ONE_SRC
},
1197 { 0x0cf64, "FLOGE_TABLE", ADD_ONE_SRC
},
1198 { 0x0d000, "BRANCH", ADD_BRANCH
},
1199 { 0x0e8c0, "MUX", ADD_THREE_SRC
},
1200 { 0x0e9b0, "ATAN_LDEXP.Y.f32", ADD_TWO_SRC
},
1201 { 0x0e9b8, "ATAN_LDEXP.X.f32", ADD_TWO_SRC
},
1202 { 0x0ea60, "SEL.XX.i16", ADD_TWO_SRC
},
1203 { 0x0ea70, "SEL.XY.i16", ADD_TWO_SRC
},
1204 { 0x0ea68, "SEL.YX.i16", ADD_TWO_SRC
},
1205 { 0x0ea78, "SEL.YY.i16", ADD_TWO_SRC
},
1206 { 0x0ec00, "F32_TO_F16", ADD_TWO_SRC
},
1207 { 0x0f640, "ICMP.GL.GT", ADD_TWO_SRC
}, // src0 > src1 ? 1 : 0
1208 { 0x0f648, "ICMP.GL.GE", ADD_TWO_SRC
},
1209 { 0x0f650, "UCMP.GL.GT", ADD_TWO_SRC
},
1210 { 0x0f658, "UCMP.GL.GE", ADD_TWO_SRC
},
1211 { 0x0f660, "ICMP.GL.EQ", ADD_TWO_SRC
},
1212 { 0x0f669, "ICMP.GL.NEQ", ADD_TWO_SRC
},
1213 { 0x0f6c0, "ICMP.D3D.GT", ADD_TWO_SRC
}, // src0 > src1 ? ~0 : 0
1214 { 0x0f6c8, "ICMP.D3D.GE", ADD_TWO_SRC
},
1215 { 0x0f6d0, "UCMP.D3D.GT", ADD_TWO_SRC
},
1216 { 0x0f6d8, "UCMP.D3D.GE", ADD_TWO_SRC
},
1217 { 0x0f6e0, "ICMP.D3D.EQ", ADD_TWO_SRC
},
1218 { 0x10000, "MAX.v2f16", ADD_FMINMAX16
},
1219 { 0x11000, "ADD_MSCALE.f32", ADD_FADDMscale
},
1220 { 0x12000, "MIN.v2f16", ADD_FMINMAX16
},
1221 { 0x14000, "ADD.v2f16", ADD_FADD16
},
1222 { 0x17000, "FCMP.D3D", ADD_FCMP16
},
1223 { 0x178c0, "ADD.i32", ADD_TWO_SRC
},
1224 { 0x17900, "ADD.v2i16", ADD_TWO_SRC
},
1225 { 0x17ac0, "SUB.i32", ADD_TWO_SRC
},
1226 { 0x17c10, "ADDC.i32", ADD_TWO_SRC
}, // adds src0 to the bottom bit of src1
1227 { 0x17d80, "ADD.i32.i16.X", ADD_TWO_SRC
},
1228 { 0x17d90, "ADD.i32.u16.X", ADD_TWO_SRC
},
1229 { 0x17dc0, "ADD.i32.i16.Y", ADD_TWO_SRC
},
1230 { 0x17dd0, "ADD.i32.u16.Y", ADD_TWO_SRC
},
1231 { 0x18000, "LD_VAR_ADDR.f16", ADD_VARYING_ADDRESS
, true },
1232 { 0x18100, "LD_VAR_ADDR.f32", ADD_VARYING_ADDRESS
, true },
1233 { 0x18200, "LD_VAR_ADDR.i32", ADD_VARYING_ADDRESS
, true },
1234 { 0x18300, "LD_VAR_ADDR.u32", ADD_VARYING_ADDRESS
, true },
1235 { 0x19181, "DISCARD.FEQ.f32", ADD_TWO_SRC
, true },
1236 { 0x19189, "DISCARD.FNE.f32", ADD_TWO_SRC
, true },
1237 { 0x1918C, "DISCARD.GL.f32", ADD_TWO_SRC
, true }, /* Consumes ICMP.GL/etc with fixed 0 argument */
1238 { 0x19190, "DISCARD.FLE.f32", ADD_TWO_SRC
, true },
1239 { 0x19198, "DISCARD.FLT.f32", ADD_TWO_SRC
, true },
1240 { 0x191e8, "ATEST.f32", ADD_TWO_SRC
, true },
1241 { 0x191f0, "ATEST.X.f16", ADD_TWO_SRC
, true },
1242 { 0x191f8, "ATEST.Y.f16", ADD_TWO_SRC
, true },
1243 { 0x19300, "ST_VAR.v1", ADD_THREE_SRC
, true },
1244 { 0x19340, "ST_VAR.v2", ADD_THREE_SRC
, true },
1245 { 0x19380, "ST_VAR.v3", ADD_THREE_SRC
, true },
1246 { 0x193c0, "ST_VAR.v4", ADD_THREE_SRC
, true },
1247 { 0x1952c, "BLEND", ADD_BLENDING
, true },
1248 { 0x1a000, "LD_VAR.16", ADD_VARYING_INTERP
, true },
1249 { 0x1ae60, "TEX", ADD_TEX
, true },
1250 { 0x1c000, "RSHIFT_NAND.i32", ADD_THREE_SRC
},
1251 { 0x1c300, "RSHIFT_OR.i32", ADD_THREE_SRC
},
1252 { 0x1c400, "RSHIFT_AND.i32", ADD_THREE_SRC
},
1253 { 0x1c700, "RSHIFT_NOR.i32", ADD_THREE_SRC
},
1254 { 0x1c800, "LSHIFT_NAND.i32", ADD_THREE_SRC
},
1255 { 0x1cb00, "LSHIFT_OR.i32", ADD_THREE_SRC
},
1256 { 0x1cc00, "LSHIFT_AND.i32", ADD_THREE_SRC
},
1257 { 0x1cf00, "LSHIFT_NOR.i32", ADD_THREE_SRC
},
1258 { 0x1d000, "RSHIFT_XOR.i32", ADD_THREE_SRC
},
1259 { 0x1d100, "RSHIFT_XNOR.i32", ADD_THREE_SRC
},
1260 { 0x1d200, "LSHIFT_XOR.i32", ADD_THREE_SRC
},
1261 { 0x1d300, "LSHIFT_XNOR.i32", ADD_THREE_SRC
},
1262 { 0x1d400, "LSHIFT_ADD.i32", ADD_THREE_SRC
},
1263 { 0x1d500, "LSHIFT_SUB.i32", ADD_THREE_SRC
},
1264 { 0x1d500, "LSHIFT_RSUB.i32", ADD_THREE_SRC
},
1265 { 0x1d700, "RSHIFT_ADD.i32", ADD_THREE_SRC
},
1266 { 0x1d800, "RSHIFT_SUB.i32", ADD_THREE_SRC
},
1267 { 0x1d900, "RSHIFT_RSUB.i32", ADD_THREE_SRC
},
1268 { 0x1da00, "ARSHIFT_ADD.i32", ADD_THREE_SRC
},
1269 { 0x1db00, "ARSHIFT_SUB.i32", ADD_THREE_SRC
},
1270 { 0x1dc00, "ARSHIFT_RSUB.i32", ADD_THREE_SRC
},
1271 { 0x1dd18, "OR.i32", ADD_TWO_SRC
},
1272 { 0x1dd20, "AND.i32", ADD_TWO_SRC
},
1273 { 0x1dd60, "LSHIFT.i32", ADD_TWO_SRC
},
1274 { 0x1dd50, "XOR.i32", ADD_TWO_SRC
},
1275 { 0x1dd80, "RSHIFT.i32", ADD_TWO_SRC
},
1276 { 0x1dda0, "ARSHIFT.i32", ADD_TWO_SRC
},
1279 static struct add_op_info
find_add_op_info(unsigned op
)
1281 for (unsigned i
= 0; i
< ARRAY_SIZE(add_op_infos
); i
++) {
1282 unsigned opCmp
= ~0;
1283 switch (add_op_infos
[i
].src_type
) {
1300 opCmp
= op
& ~0x1fff;
1303 case ADD_FADDMscale
:
1304 opCmp
= op
& ~0xfff;
1308 opCmp
= op
& ~0x7ff;
1310 case ADD_TEX_COMPACT
:
1311 opCmp
= op
& ~0x3ff;
1313 case ADD_VARYING_INTERP
:
1314 opCmp
= op
& ~0x7ff;
1316 case ADD_VARYING_ADDRESS
:
1323 opCmp
= op
& ~0xfff;
1329 if (add_op_infos
[i
].op
== opCmp
)
1330 return add_op_infos
[i
];
1333 struct add_op_info info
;
1334 snprintf(info
.name
, sizeof(info
.name
), "op%04x", op
);
1336 info
.src_type
= ADD_TWO_SRC
;
1337 info
.has_data_reg
= true;
1341 static void dump_add(FILE *fp
, uint64_t word
, struct bifrost_regs regs
,
1342 struct bifrost_regs next_regs
, uint64_t *consts
,
1343 unsigned data_reg
, unsigned offset
, bool verbose
)
1346 fprintf(fp
, "# ADD: %016" PRIx64
"\n", word
);
1348 struct bifrost_add_inst ADD
;
1349 memcpy((char *) &ADD
, (char *) &word
, sizeof(ADD
));
1350 struct add_op_info info
= find_add_op_info(ADD
.op
);
1352 fprintf(fp
, "%s", info
.name
);
1354 // float16 seems like it doesn't support output modifiers
1355 if (info
.src_type
== ADD_FADD
|| info
.src_type
== ADD_FMINMAX
) {
1357 dump_output_mod(fp
, bits(ADD
.op
, 8, 10));
1358 if (info
.src_type
== ADD_FADD
)
1359 dump_round_mode(fp
, bits(ADD
.op
, 10, 12));
1361 dump_minmax_mode(fp
, bits(ADD
.op
, 10, 12));
1362 } else if (info
.src_type
== ADD_FCMP
|| info
.src_type
== ADD_FCMP16
) {
1363 dump_fcmp(fp
, bits(ADD
.op
, 3, 6));
1364 if (info
.src_type
== ADD_FCMP
)
1365 fprintf(fp
, ".f32");
1367 fprintf(fp
, ".v2f16");
1368 } else if (info
.src_type
== ADD_FADDMscale
) {
1369 switch ((ADD
.op
>> 6) & 0x7) {
1372 // causes GPU hangs on G71
1374 fprintf(fp
, ".invalid");
1376 // Same as usual outmod value.
1378 fprintf(fp
, ".clamp_0_1");
1380 // If src0 is infinite or NaN, flush it to zero so that the other
1381 // source is passed through unmodified.
1383 fprintf(fp
, ".flush_src0_inf_nan");
1387 fprintf(fp
, ".flush_src1_inf_nan");
1389 // Every other case seems to behave the same as the above?
1391 fprintf(fp
, ".unk%d", (ADD
.op
>> 6) & 0x7);
1394 } else if (info
.src_type
== ADD_VARYING_INTERP
) {
1396 fprintf(fp
, ".reuse");
1398 fprintf(fp
, ".flat");
1399 switch ((ADD
.op
>> 7) & 0x3) {
1401 fprintf(fp
, ".per_frag");
1404 fprintf(fp
, ".centroid");
1409 fprintf(fp
, ".explicit");
1412 fprintf(fp
, ".v%d", ((ADD
.op
>> 5) & 0x3) + 1);
1413 } else if (info
.src_type
== ADD_BRANCH
) {
1414 enum branch_code branchCode
= (enum branch_code
) ((ADD
.op
>> 6) & 0x3f);
1415 if (branchCode
== BR_ALWAYS
) {
1416 // unconditional branch
1418 enum branch_cond cond
= (enum branch_cond
) ((ADD
.op
>> 6) & 0x7);
1419 enum branch_bit_size size
= (enum branch_bit_size
) ((ADD
.op
>> 9) & 0x7);
1420 bool portSwapped
= (ADD
.op
& 0x7) < ADD
.src0
;
1421 // See the comment in branch_bit_size
1422 if (size
== BR_SIZE_16YX0
)
1424 if (size
== BR_SIZE_16YX1
)
1425 portSwapped
= false;
1426 // These sizes are only for floating point comparisons, so the
1427 // non-floating-point comparisons are reused to encode the flipped
1429 if (size
== BR_SIZE_32_AND_16X
|| size
== BR_SIZE_32_AND_16Y
)
1430 portSwapped
= false;
1431 // There's only one argument, so we reuse the extra argument to
1433 if (size
== BR_SIZE_ZERO
)
1434 portSwapped
= !(ADD
.op
& 1);
1439 fprintf(fp
, ".LT.u");
1441 fprintf(fp
, ".LT.i");
1444 if (size
== BR_SIZE_32_AND_16X
|| size
== BR_SIZE_32_AND_16Y
) {
1445 fprintf(fp
, ".UNE.f");
1448 fprintf(fp
, ".LE.u");
1450 fprintf(fp
, ".LE.i");
1455 fprintf(fp
, ".GT.u");
1457 fprintf(fp
, ".GT.i");
1461 fprintf(fp
, ".GE.u");
1463 fprintf(fp
, ".GE.i");
1467 fprintf(fp
, ".NE.i");
1469 fprintf(fp
, ".EQ.i");
1473 fprintf(fp
, ".UNE.f");
1475 fprintf(fp
, ".OEQ.f");
1479 fprintf(fp
, ".OGT.unk.f");
1481 fprintf(fp
, ".OGT.f");
1485 fprintf(fp
, ".OLT.unk.f");
1487 fprintf(fp
, ".OLT.f");
1492 case BR_SIZE_32_AND_16X
:
1493 case BR_SIZE_32_AND_16Y
:
1502 case BR_SIZE_ZERO
: {
1503 unsigned ctrl
= (ADD
.op
>> 1) & 0x3;
1505 fprintf(fp
, "32.Z");
1507 fprintf(fp
, "16.Z");
1515 struct bifrost_reg_ctrl next_ctrl
= DecodeRegCtrl(fp
, next_regs
);
1516 if (next_ctrl
.add_write_unit
!= REG_WRITE_NONE
) {
1517 fprintf(fp
, "{R%d, T1}, ", GetRegToWrite(next_ctrl
.add_write_unit
, next_regs
));
1519 fprintf(fp
, "T1, ");
1522 switch (info
.src_type
) {
1524 // Note: in this case, regs.uniform_const == location | 0x8
1525 // This probably means we can't load uniforms or immediates in the
1526 // same instruction. This re-uses the encoding that normally means
1527 // "disabled", where the low 4 bits are ignored. Perhaps the extra
1528 // 0x8 or'd in indicates this is happening.
1529 fprintf(fp
, "location:%d, ", regs
.uniform_const
& 0x7);
1532 dump_src(fp
, ADD
.src0
, regs
, consts
, false);
1535 case ADD_TEX_COMPACT
: {
1538 bool dualTex
= false;
1539 if (info
.src_type
== ADD_TEX_COMPACT
) {
1540 tex_index
= (ADD
.op
>> 3) & 0x7;
1541 sampler_index
= (ADD
.op
>> 7) & 0x7;
1542 bool unknown
= (ADD
.op
& 0x40);
1543 // TODO: figure out if the unknown bit is ever 0
1545 fprintf(fp
, "unknown ");
1547 uint64_t constVal
= get_const(consts
, regs
);
1548 uint32_t controlBits
= (ADD
.op
& 0x8) ? (constVal
>> 32) : constVal
;
1549 struct bifrost_tex_ctrl ctrl
;
1550 memcpy((char *) &ctrl
, (char *) &controlBits
, sizeof(ctrl
));
1552 // TODO: figure out what actually triggers dual-tex
1553 if (ctrl
.result_type
== 9) {
1554 struct bifrost_dual_tex_ctrl dualCtrl
;
1555 memcpy((char *) &dualCtrl
, (char *) &controlBits
, sizeof(ctrl
));
1556 fprintf(fp
, "(dualtex) tex0:%d samp0:%d tex1:%d samp1:%d ",
1557 dualCtrl
.tex_index0
, dualCtrl
.sampler_index0
,
1558 dualCtrl
.tex_index1
, dualCtrl
.sampler_index1
);
1559 if (dualCtrl
.unk0
!= 3)
1560 fprintf(fp
, "unk:%d ", dualCtrl
.unk0
);
1563 if (ctrl
.no_merge_index
) {
1564 tex_index
= ctrl
.tex_index
;
1565 sampler_index
= ctrl
.sampler_index
;
1567 tex_index
= sampler_index
= ctrl
.tex_index
;
1568 unsigned unk
= ctrl
.sampler_index
>> 2;
1570 fprintf(fp
, "unk:%d ", unk
);
1571 if (ctrl
.sampler_index
& 1)
1573 if (ctrl
.sampler_index
& 2)
1578 fprintf(fp
, "unk0:%d ", ctrl
.unk0
);
1580 fprintf(fp
, "unk1 ");
1581 if (ctrl
.unk2
!= 0xf)
1582 fprintf(fp
, "unk2:%x ", ctrl
.unk2
);
1584 switch (ctrl
.result_type
) {
1586 fprintf(fp
, "f32 ");
1589 fprintf(fp
, "i32 ");
1592 fprintf(fp
, "u32 ");
1595 fprintf(fp
, "unktype(%x) ", ctrl
.result_type
);
1598 switch (ctrl
.tex_type
) {
1600 fprintf(fp
, "cube ");
1603 fprintf(fp
, "buffer ");
1614 fprintf(fp
, "shadow ");
1616 fprintf(fp
, "array ");
1619 if (ctrl
.calc_gradients
) {
1620 int comp
= (controlBits
>> 20) & 0x3;
1621 fprintf(fp
, "txg comp:%d ", comp
);
1623 fprintf(fp
, "txf ");
1626 if (!ctrl
.not_supply_lod
) {
1627 if (ctrl
.compute_lod
)
1628 fprintf(fp
, "lod_bias ");
1630 fprintf(fp
, "lod ");
1633 if (!ctrl
.calc_gradients
)
1634 fprintf(fp
, "grad ");
1637 if (ctrl
.texel_offset
)
1638 fprintf(fp
, "offset ");
1643 if (tex_index
== -1)
1644 fprintf(fp
, "tex:indirect ");
1646 fprintf(fp
, "tex:%d ", tex_index
);
1648 if (sampler_index
== -1)
1649 fprintf(fp
, "samp:indirect ");
1651 fprintf(fp
, "samp:%d ", sampler_index
);
1655 case ADD_VARYING_INTERP
: {
1656 unsigned addr
= ADD
.op
& 0x1f;
1657 if (addr
< 0b10100) {
1659 fprintf(fp
, "%d", addr
);
1660 } else if (addr
< 0b11000) {
1662 fprintf(fp
, "fragw");
1663 else if (addr
== 23)
1664 fprintf(fp
, "fragz");
1666 fprintf(fp
, "unk%d", addr
);
1668 dump_src(fp
, ADD
.op
& 0x7, regs
, consts
, false);
1671 dump_src(fp
, ADD
.src0
, regs
, consts
, false);
1674 case ADD_VARYING_ADDRESS
: {
1675 dump_src(fp
, ADD
.src0
, regs
, consts
, false);
1677 dump_src(fp
, ADD
.op
& 0x7, regs
, consts
, false);
1679 unsigned location
= (ADD
.op
>> 3) & 0x1f;
1680 if (location
< 16) {
1681 fprintf(fp
, "location:%d", location
);
1682 } else if (location
== 20) {
1683 fprintf(fp
, "location:%u", (uint32_t) get_const(consts
, regs
));
1684 } else if (location
== 21) {
1685 fprintf(fp
, "location:%u", (uint32_t) (get_const(consts
, regs
) >> 32));
1687 fprintf(fp
, "location:%d(unk)", location
);
1692 fprintf(fp
, "location:%d, ", (ADD
.op
>> 3) & 0xf);
1694 dump_src(fp
, ADD
.src0
, regs
, consts
, false);
1696 dump_src(fp
, ADD
.op
& 0x7, regs
, consts
, false);
1699 dump_src(fp
, ADD
.src0
, regs
, consts
, false);
1701 dump_src(fp
, ADD
.op
& 0x7, regs
, consts
, false);
1703 dump_src(fp
, (ADD
.op
>> 3) & 0x7, regs
, consts
, false);
1709 if (ADD
.op
& 0x1000)
1710 fprintf(fp
, "abs(");
1711 dump_src(fp
, ADD
.src0
, regs
, consts
, false);
1712 switch ((ADD
.op
>> 6) & 0x3) {
1719 if (ADD
.op
& 0x1000)
1725 fprintf(fp
, "abs(");
1726 dump_src(fp
, ADD
.op
& 0x7, regs
, consts
, false);
1727 switch ((ADD
.op
>> 6) & 0x3) {
1738 fprintf(fp
, ".unk");
1747 if (ADD
.op
& 0x1000)
1748 fprintf(fp
, "abs(");
1749 dump_src(fp
, ADD
.src0
, regs
, consts
, false);
1750 if (ADD
.op
& 0x1000)
1752 dump_16swizzle(fp
, (ADD
.op
>> 6) & 0x3);
1757 fprintf(fp
, "abs(");
1758 dump_src(fp
, ADD
.op
& 0x7, regs
, consts
, false);
1759 dump_16swizzle(fp
, (ADD
.op
>> 8) & 0x3);
1763 case ADD_FMINMAX16
: {
1764 bool abs1
= ADD
.op
& 0x8;
1765 bool abs2
= (ADD
.op
& 0x7) < ADD
.src0
;
1769 fprintf(fp
, "abs(");
1770 dump_src(fp
, ADD
.src0
, regs
, consts
, false);
1771 dump_16swizzle(fp
, (ADD
.op
>> 6) & 0x3);
1778 fprintf(fp
, "abs(");
1779 dump_src(fp
, ADD
.op
& 0x7, regs
, consts
, false);
1780 dump_16swizzle(fp
, (ADD
.op
>> 8) & 0x3);
1785 case ADD_FADDMscale
: {
1789 fprintf(fp
, "abs(");
1790 dump_src(fp
, ADD
.src0
, regs
, consts
, false);
1798 dump_src(fp
, ADD
.op
& 0x7, regs
, consts
, false);
1802 dump_src(fp
, (ADD
.op
>> 3) & 0x7, regs
, consts
, false);
1806 if (ADD
.op
& 0x400) {
1809 if (ADD
.op
& 0x100) {
1810 fprintf(fp
, "abs(");
1812 dump_src(fp
, ADD
.src0
, regs
, consts
, false);
1813 switch ((ADD
.op
>> 6) & 0x3) {
1820 if (ADD
.op
& 0x100) {
1824 if (ADD
.op
& 0x200) {
1825 fprintf(fp
, "abs(");
1827 dump_src(fp
, ADD
.op
& 0x7, regs
, consts
, false);
1828 switch ((ADD
.op
>> 6) & 0x3) {
1839 fprintf(fp
, ".unk");
1842 if (ADD
.op
& 0x200) {
1847 dump_src(fp
, ADD
.src0
, regs
, consts
, false);
1848 dump_16swizzle(fp
, (ADD
.op
>> 6) & 0x3);
1850 dump_src(fp
, ADD
.op
& 0x7, regs
, consts
, false);
1851 dump_16swizzle(fp
, (ADD
.op
>> 8) & 0x3);
1854 enum branch_code code
= (enum branch_code
) ((ADD
.op
>> 6) & 0x3f);
1855 enum branch_bit_size size
= (enum branch_bit_size
) ((ADD
.op
>> 9) & 0x7);
1856 if (code
!= BR_ALWAYS
) {
1857 dump_src(fp
, ADD
.src0
, regs
, consts
, false);
1867 case BR_SIZE_ZERO
: {
1868 unsigned ctrl
= (ADD
.op
>> 1) & 0x3;
1885 if (code
!= BR_ALWAYS
&& size
!= BR_SIZE_ZERO
) {
1886 dump_src(fp
, ADD
.op
& 0x7, regs
, consts
, false);
1891 case BR_SIZE_32_AND_16X
:
1895 case BR_SIZE_32_AND_16Y
:
1903 // I haven't had the chance to test if this actually specifies the
1904 // branch offset, since I couldn't get it to produce values other
1905 // than 5 (uniform/const high), but these three bits are always
1906 // consistent across branch instructions, so it makes sense...
1907 int offsetSrc
= (ADD
.op
>> 3) & 0x7;
1908 if (offsetSrc
== 4 || offsetSrc
== 5) {
1909 // If the offset is known/constant, we can decode it
1910 uint32_t raw_offset
;
1912 raw_offset
= get_const(consts
, regs
);
1914 raw_offset
= get_const(consts
, regs
) >> 32;
1915 // The high 4 bits are flags, while the rest is the
1916 // twos-complement offset in bytes (here we convert to
1918 int32_t branch_offset
= ((int32_t) raw_offset
<< 4) >> 8;
1920 // If high4 is the high 4 bits of the last 64-bit constant,
1921 // this is calculated as (high4 + 4) & 0xf, or 0 if the branch
1922 // offset itself is the last constant. Not sure if this is
1923 // actually used, or just garbage in unused bits, but in any
1924 // case, we can just ignore it here since it's redundant. Note
1925 // that if there is any padding, this will be 4 since the
1926 // padding counts as the last constant.
1927 unsigned flags
= raw_offset
>> 28;
1930 // Note: the offset is in bytes, relative to the beginning of the
1931 // current clause, so a zero offset would be a loop back to the
1932 // same clause (annoyingly different from Midgard).
1933 fprintf(fp
, "clause_%d", offset
+ branch_offset
);
1935 dump_src(fp
, offsetSrc
, regs
, consts
, false);
1939 if (info
.has_data_reg
) {
1940 fprintf(fp
, ", R%d", data_reg
);
1945 void dump_instr(FILE *fp
, const struct bifrost_alu_inst
*instr
,
1946 struct bifrost_regs next_regs
, uint64_t *consts
,
1947 unsigned data_reg
, unsigned offset
, bool verbose
)
1949 struct bifrost_regs regs
;
1950 memcpy((char *) ®s
, (char *) &instr
->reg_bits
, sizeof(regs
));
1953 fprintf(fp
, "# regs: %016" PRIx64
"\n", instr
->reg_bits
);
1954 dump_regs(fp
, regs
);
1956 dump_fma(fp
, instr
->fma_bits
, regs
, next_regs
, consts
, verbose
);
1957 dump_add(fp
, instr
->add_bits
, regs
, next_regs
, consts
, data_reg
, offset
, verbose
);
1960 bool dump_clause(FILE *fp
, uint32_t *words
, unsigned *size
, unsigned offset
, bool verbose
)
1962 // State for a decoded clause
1963 struct bifrost_alu_inst instrs
[8] = {};
1964 uint64_t consts
[6] = {};
1965 unsigned num_instrs
= 0;
1966 unsigned num_consts
= 0;
1967 uint64_t header_bits
= 0;
1968 bool stopbit
= false;
1971 for (i
= 0; ; i
++, words
+= 4) {
1974 for (int j
= 0; j
< 4; j
++)
1975 fprintf(fp
, "%08x ", words
[3 - j
]); // low bit on the right
1978 unsigned tag
= bits(words
[0], 0, 8);
1980 // speculatively decode some things that are common between many formats, so we can share some code
1981 struct bifrost_alu_inst main_instr
= {};
1983 main_instr
.add_bits
= bits(words
[2], 2, 32 - 13);
1985 main_instr
.fma_bits
= bits(words
[1], 11, 32) | bits(words
[2], 0, 2) << (32 - 11);
1987 main_instr
.reg_bits
= ((uint64_t) bits(words
[1], 0, 11)) << 24 | (uint64_t) bits(words
[0], 8, 32);
1989 uint64_t const0
= bits(words
[0], 8, 32) << 4 | (uint64_t) words
[1] << 28 | bits(words
[2], 0, 4) << 60;
1990 uint64_t const1
= bits(words
[2], 4, 32) << 4 | (uint64_t) words
[3] << 32;
1992 bool stop
= tag
& 0x40;
1995 fprintf(fp
, "# tag: 0x%02x\n", tag
);
1998 unsigned idx
= stop
? 5 : 2;
1999 main_instr
.add_bits
|= ((tag
>> 3) & 0x7) << 17;
2000 instrs
[idx
+ 1] = main_instr
;
2001 instrs
[idx
].add_bits
= bits(words
[3], 0, 17) | ((tag
& 0x7) << 17);
2002 instrs
[idx
].fma_bits
|= bits(words
[2], 19, 32) << 10;
2003 consts
[0] = bits(words
[3], 17, 32) << 4;
2006 switch ((tag
>> 3) & 0x7) {
2008 switch (tag
& 0x7) {
2010 main_instr
.add_bits
|= bits(words
[3], 29, 32) << 17;
2011 instrs
[1] = main_instr
;
2016 instrs
[2].add_bits
= bits(words
[3], 0, 17) | bits(words
[3], 29, 32) << 17;
2017 instrs
[2].fma_bits
|= bits(words
[2], 19, 32) << 10;
2025 instrs
[2].add_bits
= bits(words
[3], 0, 17) | bits(words
[3], 29, 32) << 17;
2026 instrs
[2].fma_bits
|= bits(words
[2], 19, 32) << 10;
2027 main_instr
.add_bits
|= bits(words
[3], 26, 29) << 17;
2028 instrs
[3] = main_instr
;
2029 if ((tag
& 0x7) == 0x5) {
2035 instrs
[5].add_bits
= bits(words
[3], 0, 17) | bits(words
[3], 29, 32) << 17;
2036 instrs
[5].fma_bits
|= bits(words
[2], 19, 32) << 10;
2043 instrs
[5].add_bits
= bits(words
[3], 0, 17) | bits(words
[3], 29, 32) << 17;
2044 instrs
[5].fma_bits
|= bits(words
[2], 19, 32) << 10;
2045 main_instr
.add_bits
|= bits(words
[3], 26, 29) << 17;
2046 instrs
[6] = main_instr
;
2051 fprintf(fp
, "unknown tag bits 0x%02x\n", tag
);
2056 unsigned idx
= ((tag
>> 3) & 0x7) == 2 ? 4 : 7;
2057 main_instr
.add_bits
|= (tag
& 0x7) << 17;
2058 instrs
[idx
] = main_instr
;
2059 consts
[0] |= (bits(words
[2], 19, 32) | ((uint64_t) words
[3] << 13)) << 19;
2061 num_instrs
= idx
+ 1;
2066 unsigned idx
= stop
? 4 : 1;
2067 main_instr
.add_bits
|= (tag
& 0x7) << 17;
2068 instrs
[idx
] = main_instr
;
2069 instrs
[idx
+ 1].fma_bits
|= bits(words
[3], 22, 32);
2070 instrs
[idx
+ 1].reg_bits
= bits(words
[2], 19, 32) | (bits(words
[3], 0, 22) << (32 - 19));
2074 // only constants can come after this
2078 header_bits
= bits(words
[2], 19, 32) | ((uint64_t) words
[3] << (32 - 19));
2079 main_instr
.add_bits
|= (tag
& 0x7) << 17;
2080 instrs
[0] = main_instr
;
2084 unsigned pos
= tag
& 0xf;
2085 // note that `pos' encodes both the total number of
2086 // instructions and the position in the constant stream,
2087 // presumably because decoded constants and instructions
2088 // share a buffer in the decoder, but we only care about
2089 // the position in the constant stream; the total number of
2090 // instructions is redundant.
2091 unsigned const_idx
= 0;
2118 fprintf(fp
, "# unknown pos 0x%x\n", pos
);
2122 if (num_consts
< const_idx
+ 2)
2123 num_consts
= const_idx
+ 2;
2125 consts
[const_idx
] = const0
;
2126 consts
[const_idx
+ 1] = const1
;
2142 fprintf(fp
, "# header: %012" PRIx64
"\n", header_bits
);
2145 struct bifrost_header header
;
2146 memcpy((char *) &header
, (char *) &header_bits
, sizeof(struct bifrost_header
));
2147 dump_header(fp
, header
, verbose
);
2148 if (!header
.no_end_of_shader
)
2152 for (i
= 0; i
< num_instrs
; i
++) {
2153 struct bifrost_regs next_regs
;
2154 if (i
+ 1 == num_instrs
) {
2155 memcpy((char *) &next_regs
, (char *) &instrs
[0].reg_bits
,
2158 memcpy((char *) &next_regs
, (char *) &instrs
[i
+ 1].reg_bits
,
2162 dump_instr(fp
, &instrs
[i
], next_regs
, consts
, header
.datareg
, offset
, verbose
);
2167 for (unsigned i
= 0; i
< num_consts
; i
++) {
2168 fprintf(fp
, "# const%d: %08" PRIx64
"\n", 2 * i
, consts
[i
] & 0xffffffff);
2169 fprintf(fp
, "# const%d: %08" PRIx64
"\n", 2 * i
+ 1, consts
[i
] >> 32);
2175 void disassemble_bifrost(FILE *fp
, uint8_t *code
, size_t size
, bool verbose
)
2177 uint32_t *words
= (uint32_t *) code
;
2178 uint32_t *words_end
= words
+ (size
/ 4);
2179 // used for displaying branch targets
2180 unsigned offset
= 0;
2181 while (words
!= words_end
) {
2182 // we don't know what the program-end bit is quite yet, so for now just
2183 // assume that an all-0 quadword is padding
2184 uint32_t zero
[4] = {};
2185 if (memcmp(words
, zero
, 4 * sizeof(uint32_t)) == 0)
2187 fprintf(fp
, "clause_%d:\n", offset
);
2189 if (dump_clause(fp
, words
, &size
, offset
, verbose
) == true) {