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
;
115 enum fma_src_type src_type
;
129 ADD_TEX_COMPACT
, // texture instruction with embedded sampler
130 ADD_TEX
, // texture instruction with sampler/etc. in uniform port
141 enum add_src_type src_type
;
145 struct bifrost_tex_ctrl
{
146 unsigned sampler_index
: 4; // also used to signal indirects
147 unsigned tex_index
: 7;
148 bool no_merge_index
: 1; // whether to merge (direct) sampler & texture indices
149 bool filter
: 1; // use the usual filtering pipeline (0 for texelFetch & textureGather)
151 bool texel_offset
: 1; // *Offset()
154 unsigned tex_type
: 2; // 2D, 3D, Cube, Buffer
155 bool compute_lod
: 1; // 0 for *Lod()
156 bool not_supply_lod
: 1; // 0 for *Lod() or when a bias is applied
157 bool calc_gradients
: 1; // 0 for *Grad()
159 unsigned result_type
: 4; // integer, unsigned, float TODO: why is this 4 bits?
163 struct bifrost_dual_tex_ctrl
{
164 unsigned sampler_index0
: 2;
166 unsigned tex_index0
: 2;
167 unsigned sampler_index1
: 2;
168 unsigned tex_index1
: 2;
172 enum branch_bit_size
{
176 // For the above combinations of bitsize and location, an extra bit is
177 // encoded via comparing the sources. The only possible source of ambiguity
178 // would be if the sources were the same, but then the branch condition
179 // would be always true or always false anyways, so we can ignore it. But
180 // this no longer works when comparing the y component to the x component,
181 // since it's valid to compare the y component of a source against its own
182 // x component. Instead, the extra bit is encoded via an extra bitsize.
185 BR_SIZE_32_AND_16X
= 5,
186 BR_SIZE_32_AND_16Y
= 6,
187 // Used for comparisons with zero and always-true, see below. I think this
188 // only works for integer comparisons.
192 void dump_header(FILE *fp
, struct bifrost_header header
, bool verbose
);
193 void dump_instr(FILE *fp
, const struct bifrost_alu_inst
*instr
,
194 struct bifrost_regs next_regs
, uint64_t *consts
,
195 unsigned data_reg
, unsigned offset
, bool verbose
);
196 bool dump_clause(FILE *fp
, uint32_t *words
, unsigned *size
, unsigned offset
, bool verbose
);
198 void dump_header(FILE *fp
, struct bifrost_header header
, bool verbose
)
200 if (header
.clause_type
!= 0) {
201 fprintf(fp
, "id(%du) ", header
.scoreboard_index
);
204 if (header
.scoreboard_deps
!= 0) {
205 fprintf(fp
, "next-wait(");
207 for (unsigned i
= 0; i
< 8; i
++) {
208 if (header
.scoreboard_deps
& (1 << i
)) {
212 fprintf(fp
, "%d", i
);
219 if (header
.datareg_writebarrier
)
220 fprintf(fp
, "data-reg-barrier ");
222 if (!header
.no_end_of_shader
)
225 if (!header
.back_to_back
) {
227 if (header
.branch_cond
)
228 fprintf(fp
, "branch-cond ");
230 fprintf(fp
, "branch-uncond ");
233 if (header
.elide_writes
)
236 if (header
.suppress_inf
)
237 fprintf(fp
, "suppress-inf ");
238 if (header
.suppress_nan
)
239 fprintf(fp
, "suppress-nan ");
242 fprintf(fp
, "unk0 ");
244 fprintf(fp
, "unk1 ");
246 fprintf(fp
, "unk2 ");
248 fprintf(fp
, "unk3 ");
250 fprintf(fp
, "unk4 ");
255 fprintf(fp
, "# clause type %d, next clause type %d\n",
256 header
.clause_type
, header
.next_clause_type
);
260 static struct bifrost_reg_ctrl
DecodeRegCtrl(FILE *fp
, struct bifrost_regs regs
)
262 struct bifrost_reg_ctrl decoded
= {};
264 if (regs
.ctrl
== 0) {
265 ctrl
= regs
.reg1
>> 2;
266 decoded
.read_reg0
= !(regs
.reg1
& 0x2);
267 decoded
.read_reg1
= false;
270 decoded
.read_reg0
= decoded
.read_reg1
= true;
274 decoded
.fma_write_unit
= REG_WRITE_TWO
;
278 decoded
.fma_write_unit
= REG_WRITE_TWO
;
279 decoded
.read_reg3
= true;
282 decoded
.read_reg3
= true;
285 decoded
.add_write_unit
= REG_WRITE_TWO
;
288 decoded
.add_write_unit
= REG_WRITE_TWO
;
289 decoded
.read_reg3
= true;
292 decoded
.clause_start
= true;
295 decoded
.fma_write_unit
= REG_WRITE_TWO
;
296 decoded
.clause_start
= true;
301 decoded
.read_reg3
= true;
302 decoded
.clause_start
= true;
305 decoded
.add_write_unit
= REG_WRITE_TWO
;
306 decoded
.clause_start
= true;
311 decoded
.fma_write_unit
= REG_WRITE_THREE
;
312 decoded
.add_write_unit
= REG_WRITE_TWO
;
315 fprintf(fp
, "# unknown reg ctrl %d\n", ctrl
);
321 // Pass in the add_write_unit or fma_write_unit, and this returns which register
322 // the ADD/FMA units are writing to
323 static unsigned GetRegToWrite(enum bifrost_reg_write_unit unit
, struct bifrost_regs regs
)
328 case REG_WRITE_THREE
:
330 default: /* REG_WRITE_NONE */
336 static void dump_regs(FILE *fp
, struct bifrost_regs srcs
)
338 struct bifrost_reg_ctrl ctrl
= DecodeRegCtrl(fp
, srcs
);
341 fprintf(fp
, "port 0: R%d ", get_reg0(srcs
));
343 fprintf(fp
, "port 1: R%d ", get_reg1(srcs
));
345 if (ctrl
.fma_write_unit
== REG_WRITE_TWO
)
346 fprintf(fp
, "port 2: R%d (write FMA) ", srcs
.reg2
);
347 else if (ctrl
.add_write_unit
== REG_WRITE_TWO
)
348 fprintf(fp
, "port 2: R%d (write ADD) ", srcs
.reg2
);
350 if (ctrl
.fma_write_unit
== REG_WRITE_THREE
)
351 fprintf(fp
, "port 3: R%d (write FMA) ", srcs
.reg3
);
352 else if (ctrl
.add_write_unit
== REG_WRITE_THREE
)
353 fprintf(fp
, "port 3: R%d (write ADD) ", srcs
.reg3
);
354 else if (ctrl
.read_reg3
)
355 fprintf(fp
, "port 3: R%d (read) ", srcs
.reg3
);
357 if (srcs
.uniform_const
) {
358 if (srcs
.uniform_const
& 0x80) {
359 fprintf(fp
, "uniform: U%d", (srcs
.uniform_const
& 0x7f) * 2);
365 static void dump_const_imm(FILE *fp
, uint32_t imm
)
372 fprintf(fp
, "0x%08x /* %f */", imm
, fi
.f
);
375 static uint64_t get_const(uint64_t *consts
, struct bifrost_regs srcs
)
377 unsigned low_bits
= srcs
.uniform_const
& 0xf;
379 switch (srcs
.uniform_const
>> 4) {
402 return imm
| low_bits
;
405 static void dump_uniform_const_src(FILE *fp
, struct bifrost_regs srcs
, uint64_t *consts
, bool high32
)
407 if (srcs
.uniform_const
& 0x80) {
408 unsigned uniform
= (srcs
.uniform_const
& 0x7f) * 2;
409 fprintf(fp
, "U%d", uniform
+ (high32
? 1 : 0));
410 } else if (srcs
.uniform_const
>= 0x20) {
411 uint64_t imm
= get_const(consts
, srcs
);
413 dump_const_imm(fp
, imm
>> 32);
415 dump_const_imm(fp
, imm
);
417 switch (srcs
.uniform_const
) {
422 fprintf(fp
, "atest-data");
425 fprintf(fp
, "sample-ptr");
435 fprintf(fp
, "blend-descriptor%u", (unsigned) srcs
.uniform_const
- 8);
438 fprintf(fp
, "unkConst%u", (unsigned) srcs
.uniform_const
);
449 static void dump_src(FILE *fp
, unsigned src
, struct bifrost_regs srcs
, uint64_t *consts
, bool isFMA
)
453 fprintf(fp
, "R%d", get_reg0(srcs
));
456 fprintf(fp
, "R%d", get_reg1(srcs
));
459 fprintf(fp
, "R%d", srcs
.reg3
);
465 fprintf(fp
, "T"); // i.e. the output of FMA this cycle
468 dump_uniform_const_src(fp
, srcs
, consts
, false);
471 dump_uniform_const_src(fp
, srcs
, consts
, true);
482 static void dump_output_mod(FILE *fp
, unsigned mod
)
488 fprintf(fp
, ".clamp_0_inf");
489 break; // max(out, 0)
491 fprintf(fp
, ".clamp_m1_1");
492 break; // clamp(out, -1, 1)
494 fprintf(fp
, ".clamp_0_1");
495 break; // clamp(out, 0, 1)
501 static void dump_minmax_mode(FILE *fp
, unsigned mod
)
505 /* Same as fmax() and fmin() -- return the other number if any
506 * number is NaN. Also always return +0 if one argument is +0 and
511 /* Instead of never returning a NaN, always return one. The
512 * "greater"/"lesser" NaN is always returned, first by checking the
513 * sign and then the mantissa bits.
515 fprintf(fp
, ".nan_wins");
518 /* For max, implement src0 > src1 ? src0 : src1
519 * For min, implement src0 < src1 ? src0 : src1
521 * This includes handling NaN's and signedness of 0 differently
522 * from above, since +0 and -0 compare equal and comparisons always
523 * return false for NaN's. As a result, this mode is *not*
526 fprintf(fp
, ".src1_wins");
529 /* For max, implement src0 < src1 ? src1 : src0
530 * For min, implement src0 > src1 ? src1 : src0
532 fprintf(fp
, ".src0_wins");
539 static void dump_round_mode(FILE *fp
, unsigned mod
)
543 /* roundTiesToEven, the IEEE default. */
546 /* roundTowardPositive in the IEEE spec. */
547 fprintf(fp
, ".round_pos");
550 /* roundTowardNegative in the IEEE spec. */
551 fprintf(fp
, ".round_neg");
554 /* roundTowardZero in the IEEE spec. */
555 fprintf(fp
, ".round_zero");
563 csel_cond_name(enum bifrost_csel_cond cond
)
566 case BIFROST_FEQ_F
: return "feq.f";
567 case BIFROST_FGT_F
: return "fgt.f";
568 case BIFROST_FGE_F
: return "fge.f";
569 case BIFROST_IEQ_F
: return "ieq.f";
570 case BIFROST_IGT_I
: return "igt.i";
571 case BIFROST_IGE_I
: return "uge.i";
572 case BIFROST_UGT_I
: return "ugt.i";
573 case BIFROST_UGE_I
: return "uge.i";
574 default: return "invalid";
578 static const struct fma_op_info FMAOpInfos
[] = {
579 { false, 0x00000, "FMA.f32", FMA_FMA
},
580 { false, 0x40000, "MAX.f32", FMA_FMINMAX
},
581 { false, 0x44000, "MIN.f32", FMA_FMINMAX
},
582 { false, 0x48000, "FCMP.GL", FMA_FCMP
},
583 { false, 0x4c000, "FCMP.D3D", FMA_FCMP
},
584 { false, 0x4ff98, "ADD.i32", FMA_TWO_SRC
},
585 { false, 0x4ffd8, "SUB.i32", FMA_TWO_SRC
},
586 { false, 0x4fff0, "SUBB.i32", FMA_TWO_SRC
},
587 { false, 0x50000, "FMA_MSCALE", FMA_FMA_MSCALE
},
588 { false, 0x58000, "ADD.f32", FMA_FADD
},
589 { false, 0x5c000, "CSEL4", FMA_CSEL4
},
590 { false, 0x5d8d0, "ICMP.D3D.GT.v2i16", FMA_TWO_SRC
},
591 { false, 0x5d9d0, "UCMP.D3D.GT.v2i16", FMA_TWO_SRC
},
592 { false, 0x5dad0, "ICMP.D3D.GE.v2i16", FMA_TWO_SRC
},
593 { false, 0x5dbd0, "UCMP.D3D.GE.v2i16", FMA_TWO_SRC
},
594 { false, 0x5dcd0, "ICMP.D3D.EQ.v2i16", FMA_TWO_SRC
},
595 { false, 0x5de40, "ICMP.GL.GT.i32", FMA_TWO_SRC
}, // src0 > src1 ? 1 : 0
596 { false, 0x5de48, "ICMP.GL.GE.i32", FMA_TWO_SRC
},
597 { false, 0x5de50, "UCMP.GL.GT.i32", FMA_TWO_SRC
},
598 { false, 0x5de58, "UCMP.GL.GE.i32", FMA_TWO_SRC
},
599 { false, 0x5de60, "ICMP.GL.EQ.i32", FMA_TWO_SRC
},
600 { false, 0x5dec0, "ICMP.D3D.GT.i32", FMA_TWO_SRC
}, // src0 > src1 ? ~0 : 0
601 { false, 0x5dec8, "ICMP.D3D.GE.i32", FMA_TWO_SRC
},
602 { false, 0x5ded0, "UCMP.D3D.GT.i32", FMA_TWO_SRC
},
603 { false, 0x5ded8, "UCMP.D3D.GE.i32", FMA_TWO_SRC
},
604 { false, 0x5dee0, "ICMP.D3D.EQ.i32", FMA_TWO_SRC
},
605 { false, 0x60000, "RSHIFT_NAND", FMA_SHIFT
},
606 { false, 0x61000, "RSHIFT_AND", FMA_SHIFT
},
607 { false, 0x62000, "LSHIFT_NAND", FMA_SHIFT
},
608 { false, 0x63000, "LSHIFT_AND", FMA_SHIFT
}, // (src0 << src2) & src1
609 { false, 0x64000, "RSHIFT_XOR", FMA_SHIFT
},
610 { false, 0x65200, "LSHIFT_ADD.i32", FMA_THREE_SRC
},
611 { false, 0x65600, "LSHIFT_SUB.i32", FMA_THREE_SRC
}, // (src0 << src2) - src1
612 { false, 0x65a00, "LSHIFT_RSUB.i32", FMA_THREE_SRC
}, // src1 - (src0 << src2)
613 { false, 0x65e00, "RSHIFT_ADD.i32", FMA_THREE_SRC
},
614 { false, 0x66200, "RSHIFT_SUB.i32", FMA_THREE_SRC
},
615 { false, 0x66600, "RSHIFT_RSUB.i32", FMA_THREE_SRC
},
616 { false, 0x66a00, "ARSHIFT_ADD.i32", FMA_THREE_SRC
},
617 { false, 0x66e00, "ARSHIFT_SUB.i32", FMA_THREE_SRC
},
618 { false, 0x67200, "ARSHIFT_RSUB.i32", FMA_THREE_SRC
},
619 { false, 0x80000, "FMA.v2f16", FMA_FMA16
},
620 { false, 0xc0000, "MAX.v2f16", FMA_FMINMAX16
},
621 { false, 0xc4000, "MIN.v2f16", FMA_FMINMAX16
},
622 { false, 0xc8000, "FCMP.GL", FMA_FCMP16
},
623 { false, 0xcc000, "FCMP.D3D", FMA_FCMP16
},
624 { false, 0xcf900, "ADD.v2i16", FMA_TWO_SRC
},
625 { false, 0xcfc10, "ADDC.i32", FMA_TWO_SRC
},
626 { false, 0xcfd80, "ADD.i32.i16.X", FMA_TWO_SRC
},
627 { false, 0xcfd90, "ADD.i32.u16.X", FMA_TWO_SRC
},
628 { false, 0xcfdc0, "ADD.i32.i16.Y", FMA_TWO_SRC
},
629 { false, 0xcfdd0, "ADD.i32.u16.Y", FMA_TWO_SRC
},
630 { false, 0xd8000, "ADD.v2f16", FMA_FADD16
},
631 { false, 0xdc000, "CSEL4.v16", FMA_CSEL4
},
632 { false, 0xdd000, "F32_TO_F16", FMA_TWO_SRC
},
633 { true, 0x00046, "F16_TO_I16.XX", FMA_ONE_SRC
},
634 { true, 0x00047, "F16_TO_U16.XX", FMA_ONE_SRC
},
635 { true, 0x0004e, "F16_TO_I16.YX", FMA_ONE_SRC
},
636 { true, 0x0004f, "F16_TO_U16.YX", FMA_ONE_SRC
},
637 { true, 0x00056, "F16_TO_I16.XY", FMA_ONE_SRC
},
638 { true, 0x00057, "F16_TO_U16.XY", FMA_ONE_SRC
},
639 { true, 0x0005e, "F16_TO_I16.YY", FMA_ONE_SRC
},
640 { true, 0x0005f, "F16_TO_U16.YY", FMA_ONE_SRC
},
641 { true, 0x000c0, "I16_TO_F16.XX", FMA_ONE_SRC
},
642 { true, 0x000c1, "U16_TO_F16.XX", FMA_ONE_SRC
},
643 { true, 0x000c8, "I16_TO_F16.YX", FMA_ONE_SRC
},
644 { true, 0x000c9, "U16_TO_F16.YX", FMA_ONE_SRC
},
645 { true, 0x000d0, "I16_TO_F16.XY", FMA_ONE_SRC
},
646 { true, 0x000d1, "U16_TO_F16.XY", FMA_ONE_SRC
},
647 { true, 0x000d8, "I16_TO_F16.YY", FMA_ONE_SRC
},
648 { true, 0x000d9, "U16_TO_F16.YY", FMA_ONE_SRC
},
649 { true, 0x00136, "F32_TO_I32", FMA_ONE_SRC
},
650 { true, 0x00137, "F32_TO_U32", FMA_ONE_SRC
},
651 { true, 0x00178, "I32_TO_F32", FMA_ONE_SRC
},
652 { true, 0x00179, "U32_TO_F32", FMA_ONE_SRC
},
653 { true, 0x00198, "I16_TO_I32.X", FMA_ONE_SRC
},
654 { true, 0x00199, "U16_TO_U32.X", FMA_ONE_SRC
},
655 { true, 0x0019a, "I16_TO_I32.Y", FMA_ONE_SRC
},
656 { true, 0x0019b, "U16_TO_U32.Y", FMA_ONE_SRC
},
657 { true, 0x0019c, "I16_TO_F32.X", FMA_ONE_SRC
},
658 { true, 0x0019d, "U16_TO_F32.X", FMA_ONE_SRC
},
659 { true, 0x0019e, "I16_TO_F32.Y", FMA_ONE_SRC
},
660 { true, 0x0019f, "U16_TO_F32.Y", FMA_ONE_SRC
},
661 { true, 0x001a2, "F16_TO_F32.X", FMA_ONE_SRC
},
662 { true, 0x001a3, "F16_TO_F32.Y", FMA_ONE_SRC
},
663 { true, 0x0032c, "NOP", FMA_ONE_SRC
},
664 { true, 0x0032d, "MOV", FMA_ONE_SRC
},
665 { true, 0x0032f, "SWZ.YY.v2i16", FMA_ONE_SRC
},
666 { true, 0x00345, "LOG_FREXPM", FMA_ONE_SRC
},
667 { true, 0x00365, "FRCP_FREXPM", FMA_ONE_SRC
},
668 { true, 0x00375, "FSQRT_FREXPM", FMA_ONE_SRC
},
669 { true, 0x0038d, "FRCP_FREXPE", FMA_ONE_SRC
},
670 { true, 0x003a5, "FSQRT_FREXPE", FMA_ONE_SRC
},
671 { true, 0x003ad, "FRSQ_FREXPE", FMA_ONE_SRC
},
672 { true, 0x003c5, "LOG_FREXPE", FMA_ONE_SRC
},
673 { true, 0x003fa, "CLZ", FMA_ONE_SRC
},
674 { true, 0x00b80, "IMAX3", FMA_THREE_SRC
},
675 { true, 0x00bc0, "UMAX3", FMA_THREE_SRC
},
676 { true, 0x00c00, "IMIN3", FMA_THREE_SRC
},
677 { true, 0x00c40, "UMIN3", FMA_THREE_SRC
},
678 { true, 0x00ec5, "ROUND", FMA_ONE_SRC
},
679 { true, 0x00f40, "CSEL", FMA_THREE_SRC
}, // src2 != 0 ? src1 : src0
680 { true, 0x00fc0, "MUX.i32", FMA_THREE_SRC
}, // see ADD comment
681 { true, 0x01805, "ROUNDEVEN", FMA_ONE_SRC
},
682 { true, 0x01845, "CEIL", FMA_ONE_SRC
},
683 { true, 0x01885, "FLOOR", FMA_ONE_SRC
},
684 { true, 0x018c5, "TRUNC", FMA_ONE_SRC
},
685 { true, 0x019b0, "ATAN_LDEXP.Y.f32", FMA_TWO_SRC
},
686 { true, 0x019b8, "ATAN_LDEXP.X.f32", FMA_TWO_SRC
},
687 { true, 0x01c80, "LSHIFT_ADD_LOW32.u32", FMA_SHIFT_ADD64
},
688 { true, 0x01cc0, "LSHIFT_ADD_LOW32.i64", FMA_SHIFT_ADD64
},
689 { true, 0x01d80, "LSHIFT_ADD_LOW32.i32", FMA_SHIFT_ADD64
},
690 { true, 0x01e00, "SEL.XX.i16", FMA_TWO_SRC
},
691 { true, 0x01e08, "SEL.YX.i16", FMA_TWO_SRC
},
692 { true, 0x01e10, "SEL.XY.i16", FMA_TWO_SRC
},
693 { true, 0x01e18, "SEL.YY.i16", FMA_TWO_SRC
},
694 { true, 0x00800, "IMAD", FMA_THREE_SRC
},
695 { true, 0x078db, "POPCNT", FMA_ONE_SRC
},
698 static struct fma_op_info
find_fma_op_info(unsigned op
, bool extended
)
700 for (unsigned i
= 0; i
< ARRAY_SIZE(FMAOpInfos
); i
++) {
703 if (FMAOpInfos
[i
].extended
!= extended
)
709 switch (FMAOpInfos
[i
].src_type
) {
718 opCmp
= op
& ~0x1fff;
721 case FMA_SHIFT_ADD64
:
728 opCmp
= op
& ~0x3fff;
732 opCmp
= op
& ~0x3ffff;
739 opCmp
= op
& ~0x7fff;
745 if (FMAOpInfos
[i
].op
== opCmp
)
746 return FMAOpInfos
[i
];
749 struct fma_op_info info
;
750 snprintf(info
.name
, sizeof(info
.name
), "op%04x", op
);
752 info
.src_type
= FMA_THREE_SRC
;
756 static void dump_fcmp(FILE *fp
, unsigned op
)
778 fprintf(fp
, ".unk%d", op
);
783 static void dump_16swizzle(FILE *fp
, unsigned swiz
)
787 fprintf(fp
, ".%c%c", "xy"[swiz
& 1], "xy"[(swiz
>> 1) & 1]);
790 static void dump_fma_expand_src0(FILE *fp
, unsigned ctrl
)
812 static void dump_fma_expand_src1(FILE *fp
, unsigned ctrl
)
834 static void dump_fma(FILE *fp
, uint64_t word
, struct bifrost_regs regs
, struct bifrost_regs next_regs
, uint64_t *consts
, bool verbose
)
837 fprintf(fp
, "# FMA: %016" PRIx64
"\n", word
);
839 struct bifrost_fma_inst FMA
;
840 memcpy((char *) &FMA
, (char *) &word
, sizeof(struct bifrost_fma_inst
));
841 struct fma_op_info info
= find_fma_op_info(FMA
.op
, (FMA
.op
& 0xe0000) == 0xe0000);
843 fprintf(fp
, "%s", info
.name
);
844 if (info
.src_type
== FMA_FADD
||
845 info
.src_type
== FMA_FMINMAX
||
846 info
.src_type
== FMA_FMA
||
847 info
.src_type
== FMA_FADD16
||
848 info
.src_type
== FMA_FMINMAX16
||
849 info
.src_type
== FMA_FMA16
) {
850 dump_output_mod(fp
, bits(FMA
.op
, 12, 14));
851 switch (info
.src_type
) {
856 dump_round_mode(fp
, bits(FMA
.op
, 10, 12));
860 dump_minmax_mode(fp
, bits(FMA
.op
, 10, 12));
865 } else if (info
.src_type
== FMA_FCMP
|| info
.src_type
== FMA_FCMP16
) {
866 dump_fcmp(fp
, bits(FMA
.op
, 10, 13));
867 if (info
.src_type
== FMA_FCMP
)
870 fprintf(fp
, ".v2f16");
871 } else if (info
.src_type
== FMA_FMA_MSCALE
) {
872 if (FMA
.op
& (1 << 11)) {
873 switch ((FMA
.op
>> 9) & 0x3) {
875 /* This mode seems to do a few things:
876 * - Makes 0 * infinity (and incidentally 0 * nan) return 0,
877 * since generating a nan would poison the result of
878 * 1/infinity and 1/0.
879 * - Fiddles with which nan is returned in nan * nan,
880 * presumably to make sure that the same exact nan is
881 * returned for 1/nan.
883 fprintf(fp
, ".rcp_mode");
886 /* Similar to the above, but src0 always wins when multiplying
889 fprintf(fp
, ".sqrt_mode");
892 fprintf(fp
, ".unk%d_mode", (int) (FMA
.op
>> 9) & 0x3);
895 dump_output_mod(fp
, bits(FMA
.op
, 9, 11));
897 } else if (info
.src_type
== FMA_SHIFT
) {
898 struct bifrost_shift_fma shift
;
899 memcpy(&shift
, &FMA
, sizeof(shift
));
901 if (shift
.half
== 0x7)
902 fprintf(fp
, ".v2i16");
903 else if (shift
.half
== 0)
905 else if (shift
.half
== 0x4)
906 fprintf(fp
, ".v4i8");
908 fprintf(fp
, ".unk%u", shift
.half
);
911 fprintf(fp
, ".no_unk");
914 fprintf(fp
, ".invert_1");
917 fprintf(fp
, ".invert_2");
922 struct bifrost_reg_ctrl next_ctrl
= DecodeRegCtrl(fp
, next_regs
);
923 if (next_ctrl
.fma_write_unit
!= REG_WRITE_NONE
) {
924 fprintf(fp
, "{R%d, T0}, ", GetRegToWrite(next_ctrl
.fma_write_unit
, next_regs
));
929 switch (info
.src_type
) {
931 dump_src(fp
, FMA
.src0
, regs
, consts
, true);
934 dump_src(fp
, FMA
.src0
, regs
, consts
, true);
936 dump_src(fp
, FMA
.op
& 0x7, regs
, consts
, true);
944 dump_src(fp
, FMA
.src0
, regs
, consts
, true);
945 dump_fma_expand_src0(fp
, (FMA
.op
>> 6) & 0x7);
953 dump_src(fp
, FMA
.op
& 0x7, regs
, consts
, true);
954 dump_fma_expand_src1(fp
, (FMA
.op
>> 6) & 0x7);
959 case FMA_FMINMAX16
: {
960 bool abs1
= FMA
.op
& 0x8;
961 bool abs2
= (FMA
.op
& 0x7) < FMA
.src0
;
966 dump_src(fp
, FMA
.src0
, regs
, consts
, true);
967 dump_16swizzle(fp
, (FMA
.op
>> 6) & 0x3);
975 dump_src(fp
, FMA
.op
& 0x7, regs
, consts
, true);
976 dump_16swizzle(fp
, (FMA
.op
>> 8) & 0x3);
984 dump_src(fp
, FMA
.src0
, regs
, consts
, true);
985 dump_fma_expand_src0(fp
, (FMA
.op
>> 6) & 0x7);
993 dump_src(fp
, FMA
.op
& 0x7, regs
, consts
, true);
994 dump_fma_expand_src1(fp
, (FMA
.op
>> 6) & 0x7);
999 dump_src(fp
, FMA
.src0
, regs
, consts
, true);
1000 // Note: this is kinda a guess, I haven't seen the blob set this to
1001 // anything other than the identity, but it matches FMA_TWO_SRCFmod16
1002 dump_16swizzle(fp
, (FMA
.op
>> 6) & 0x3);
1004 dump_src(fp
, FMA
.op
& 0x7, regs
, consts
, true);
1005 dump_16swizzle(fp
, (FMA
.op
>> 8) & 0x3);
1007 case FMA_SHIFT_ADD64
:
1008 dump_src(fp
, FMA
.src0
, regs
, consts
, true);
1010 dump_src(fp
, FMA
.op
& 0x7, regs
, consts
, true);
1012 fprintf(fp
, "shift:%u", (FMA
.op
>> 3) & 0x7);
1015 dump_src(fp
, FMA
.src0
, regs
, consts
, true);
1017 dump_src(fp
, FMA
.op
& 0x7, regs
, consts
, true);
1019 dump_src(fp
, (FMA
.op
>> 3) & 0x7, regs
, consts
, true);
1022 struct bifrost_shift_fma shift
;
1023 memcpy(&shift
, &FMA
, sizeof(shift
));
1025 dump_src(fp
, shift
.src0
, regs
, consts
, true);
1027 dump_src(fp
, shift
.src1
, regs
, consts
, true);
1029 dump_src(fp
, shift
.src2
, regs
, consts
, true);
1033 if (FMA
.op
& (1 << 14))
1035 if (FMA
.op
& (1 << 9))
1036 fprintf(fp
, "abs(");
1037 dump_src(fp
, FMA
.src0
, regs
, consts
, true);
1038 dump_fma_expand_src0(fp
, (FMA
.op
>> 6) & 0x7);
1039 if (FMA
.op
& (1 << 9))
1042 if (FMA
.op
& (1 << 16))
1043 fprintf(fp
, "abs(");
1044 dump_src(fp
, FMA
.op
& 0x7, regs
, consts
, true);
1045 dump_fma_expand_src1(fp
, (FMA
.op
>> 6) & 0x7);
1046 if (FMA
.op
& (1 << 16))
1049 if (FMA
.op
& (1 << 15))
1051 if (FMA
.op
& (1 << 17))
1052 fprintf(fp
, "abs(");
1053 dump_src(fp
, (FMA
.op
>> 3) & 0x7, regs
, consts
, true);
1054 if (FMA
.op
& (1 << 17))
1058 if (FMA
.op
& (1 << 14))
1060 dump_src(fp
, FMA
.src0
, regs
, consts
, true);
1061 dump_16swizzle(fp
, (FMA
.op
>> 6) & 0x3);
1063 dump_src(fp
, FMA
.op
& 0x7, regs
, consts
, true);
1064 dump_16swizzle(fp
, (FMA
.op
>> 8) & 0x3);
1066 if (FMA
.op
& (1 << 15))
1068 dump_src(fp
, (FMA
.op
>> 3) & 0x7, regs
, consts
, true);
1069 dump_16swizzle(fp
, (FMA
.op
>> 16) & 0x3);
1072 struct bifrost_csel4 csel
;
1073 memcpy(&csel
, &FMA
, sizeof(csel
));
1074 fprintf(fp
, ".%s ", csel_cond_name(csel
.cond
));
1076 dump_src(fp
, csel
.src0
, regs
, consts
, true);
1078 dump_src(fp
, csel
.src1
, regs
, consts
, true);
1080 dump_src(fp
, csel
.src2
, regs
, consts
, true);
1082 dump_src(fp
, csel
.src3
, regs
, consts
, true);
1085 case FMA_FMA_MSCALE
:
1086 if (FMA
.op
& (1 << 12))
1087 fprintf(fp
, "abs(");
1088 dump_src(fp
, FMA
.src0
, regs
, consts
, true);
1089 if (FMA
.op
& (1 << 12))
1092 if (FMA
.op
& (1 << 13))
1094 dump_src(fp
, FMA
.op
& 0x7, regs
, consts
, true);
1096 if (FMA
.op
& (1 << 14))
1098 dump_src(fp
, (FMA
.op
>> 3) & 0x7, regs
, consts
, true);
1100 dump_src(fp
, (FMA
.op
>> 6) & 0x7, regs
, consts
, true);
1106 static const struct add_op_info add_op_infos
[] = {
1107 { 0x00000, "MAX.f32", ADD_FMINMAX
},
1108 { 0x02000, "MIN.f32", ADD_FMINMAX
},
1109 { 0x04000, "ADD.f32", ADD_FADD
},
1110 { 0x06000, "FCMP.GL", ADD_FCMP
},
1111 { 0x07000, "FCMP.D3D", ADD_FCMP
},
1112 { 0x07856, "F16_TO_I16", ADD_ONE_SRC
},
1113 { 0x07857, "F16_TO_U16", ADD_ONE_SRC
},
1114 { 0x078c0, "I16_TO_F16.XX", ADD_ONE_SRC
},
1115 { 0x078c1, "U16_TO_F16.XX", ADD_ONE_SRC
},
1116 { 0x078c8, "I16_TO_F16.YX", ADD_ONE_SRC
},
1117 { 0x078c9, "U16_TO_F16.YX", ADD_ONE_SRC
},
1118 { 0x078d0, "I16_TO_F16.XY", ADD_ONE_SRC
},
1119 { 0x078d1, "U16_TO_F16.XY", ADD_ONE_SRC
},
1120 { 0x078d8, "I16_TO_F16.YY", ADD_ONE_SRC
},
1121 { 0x078d9, "U16_TO_F16.YY", ADD_ONE_SRC
},
1122 { 0x07936, "F32_TO_I32", ADD_ONE_SRC
},
1123 { 0x07937, "F32_TO_U32", ADD_ONE_SRC
},
1124 { 0x07978, "I32_TO_F32", ADD_ONE_SRC
},
1125 { 0x07979, "U32_TO_F32", ADD_ONE_SRC
},
1126 { 0x07998, "I16_TO_I32.X", ADD_ONE_SRC
},
1127 { 0x07999, "U16_TO_U32.X", ADD_ONE_SRC
},
1128 { 0x0799a, "I16_TO_I32.Y", ADD_ONE_SRC
},
1129 { 0x0799b, "U16_TO_U32.Y", ADD_ONE_SRC
},
1130 { 0x0799c, "I16_TO_F32.X", ADD_ONE_SRC
},
1131 { 0x0799d, "U16_TO_F32.X", ADD_ONE_SRC
},
1132 { 0x0799e, "I16_TO_F32.Y", ADD_ONE_SRC
},
1133 { 0x0799f, "U16_TO_F32.Y", ADD_ONE_SRC
},
1134 { 0x079a2, "F16_TO_F32.X", ADD_ONE_SRC
},
1135 { 0x079a3, "F16_TO_F32.Y", ADD_ONE_SRC
},
1136 { 0x07b2b, "SWZ.YX.v2i16", ADD_ONE_SRC
},
1137 { 0x07b2c, "NOP", ADD_ONE_SRC
},
1138 { 0x07b29, "SWZ.XX.v2i16", ADD_ONE_SRC
},
1139 { 0x07b2d, "MOV", ADD_ONE_SRC
},
1140 { 0x07b2f, "SWZ.YY.v2i16", ADD_ONE_SRC
},
1141 { 0x07b65, "FRCP_FREXPM", ADD_ONE_SRC
},
1142 { 0x07b75, "FSQRT_FREXPM", ADD_ONE_SRC
},
1143 { 0x07b8d, "FRCP_FREXPE", ADD_ONE_SRC
},
1144 { 0x07ba5, "FSQRT_FREXPE", ADD_ONE_SRC
},
1145 { 0x07bad, "FRSQ_FREXPE", ADD_ONE_SRC
},
1146 { 0x07bc5, "FLOG_FREXPE", ADD_ONE_SRC
},
1147 { 0x07d45, "CEIL", ADD_ONE_SRC
},
1148 { 0x07d85, "FLOOR", ADD_ONE_SRC
},
1149 { 0x07dc5, "TRUNC", ADD_ONE_SRC
},
1150 { 0x07f18, "LSHIFT_ADD_HIGH32.i32", ADD_TWO_SRC
},
1151 { 0x08000, "LD_ATTR.f16", ADD_LOAD_ATTR
, true },
1152 { 0x08100, "LD_ATTR.v2f16", ADD_LOAD_ATTR
, true },
1153 { 0x08200, "LD_ATTR.v3f16", ADD_LOAD_ATTR
, true },
1154 { 0x08300, "LD_ATTR.v4f16", ADD_LOAD_ATTR
, true },
1155 { 0x08400, "LD_ATTR.f32", ADD_LOAD_ATTR
, true },
1156 { 0x08500, "LD_ATTR.v3f32", ADD_LOAD_ATTR
, true },
1157 { 0x08600, "LD_ATTR.v3f32", ADD_LOAD_ATTR
, true },
1158 { 0x08700, "LD_ATTR.v4f32", ADD_LOAD_ATTR
, true },
1159 { 0x08800, "LD_ATTR.i32", ADD_LOAD_ATTR
, true },
1160 { 0x08900, "LD_ATTR.v3i32", ADD_LOAD_ATTR
, true },
1161 { 0x08a00, "LD_ATTR.v3i32", ADD_LOAD_ATTR
, true },
1162 { 0x08b00, "LD_ATTR.v4i32", ADD_LOAD_ATTR
, true },
1163 { 0x08c00, "LD_ATTR.u32", ADD_LOAD_ATTR
, true },
1164 { 0x08d00, "LD_ATTR.v3u32", ADD_LOAD_ATTR
, true },
1165 { 0x08e00, "LD_ATTR.v3u32", ADD_LOAD_ATTR
, true },
1166 { 0x08f00, "LD_ATTR.v4u32", ADD_LOAD_ATTR
, true },
1167 { 0x0a000, "LD_VAR.32", ADD_VARYING_INTERP
, true },
1168 { 0x0b000, "TEX", ADD_TEX_COMPACT
, true },
1169 { 0x0c188, "LOAD.i32", ADD_TWO_SRC
, true },
1170 { 0x0c1a0, "LD_UBO.i32", ADD_TWO_SRC
, true },
1171 { 0x0c1b8, "LD_SCRATCH.v2i32", ADD_TWO_SRC
, true },
1172 { 0x0c1c8, "LOAD.v2i32", ADD_TWO_SRC
, true },
1173 { 0x0c1e0, "LD_UBO.v2i32", ADD_TWO_SRC
, true },
1174 { 0x0c1f8, "LD_SCRATCH.v2i32", ADD_TWO_SRC
, true },
1175 { 0x0c208, "LOAD.v4i32", ADD_TWO_SRC
, true },
1176 { 0x0c220, "LD_UBO.v4i32", ADD_TWO_SRC
, true },
1177 { 0x0c238, "LD_SCRATCH.v4i32", ADD_TWO_SRC
, true },
1178 { 0x0c248, "STORE.v4i32", ADD_TWO_SRC
, true },
1179 { 0x0c278, "ST_SCRATCH.v4i32", ADD_TWO_SRC
, true },
1180 { 0x0c588, "STORE.i32", ADD_TWO_SRC
, true },
1181 { 0x0c5b8, "ST_SCRATCH.i32", ADD_TWO_SRC
, true },
1182 { 0x0c5c8, "STORE.v2i32", ADD_TWO_SRC
, true },
1183 { 0x0c5f8, "ST_SCRATCH.v2i32", ADD_TWO_SRC
, true },
1184 { 0x0c648, "LOAD.u16", ADD_TWO_SRC
, true }, // zero-extends
1185 { 0x0ca88, "LOAD.v3i32", ADD_TWO_SRC
, true },
1186 { 0x0caa0, "LD_UBO.v3i32", ADD_TWO_SRC
, true },
1187 { 0x0cab8, "LD_SCRATCH.v3i32", ADD_TWO_SRC
, true },
1188 { 0x0cb88, "STORE.v3i32", ADD_TWO_SRC
, true },
1189 { 0x0cbb8, "ST_SCRATCH.v3i32", ADD_TWO_SRC
, true },
1190 { 0x0cc00, "FRCP_FAST.f32", ADD_ONE_SRC
},
1191 { 0x0cc20, "FRSQ_FAST.f32", ADD_ONE_SRC
},
1192 { 0x0ce00, "FRCP_TABLE", ADD_ONE_SRC
},
1193 { 0x0ce10, "FRCP_FAST.f16.X", ADD_ONE_SRC
},
1194 { 0x0ce20, "FRSQ_TABLE", ADD_ONE_SRC
},
1195 { 0x0ce30, "FRCP_FAST.f16.Y", ADD_ONE_SRC
},
1196 { 0x0ce50, "FRSQ_FAST.f16.X", ADD_ONE_SRC
},
1197 { 0x0ce60, "FRCP_APPROX", ADD_ONE_SRC
},
1198 { 0x0ce70, "FRSQ_FAST.f16.Y", ADD_ONE_SRC
},
1199 { 0x0cf40, "ATAN_ASSIST", ADD_TWO_SRC
},
1200 { 0x0cf48, "ATAN_TABLE", ADD_TWO_SRC
},
1201 { 0x0cf50, "SIN_TABLE", ADD_ONE_SRC
},
1202 { 0x0cf51, "COS_TABLE", ADD_ONE_SRC
},
1203 { 0x0cf58, "EXP_TABLE", ADD_ONE_SRC
},
1204 { 0x0cf60, "FLOG2_TABLE", ADD_ONE_SRC
},
1205 { 0x0cf64, "FLOGE_TABLE", ADD_ONE_SRC
},
1206 { 0x0d000, "BRANCH", ADD_BRANCH
},
1207 { 0x0e8c0, "MUX", ADD_THREE_SRC
},
1208 { 0x0e9b0, "ATAN_LDEXP.Y.f32", ADD_TWO_SRC
},
1209 { 0x0e9b8, "ATAN_LDEXP.X.f32", ADD_TWO_SRC
},
1210 { 0x0ea60, "SEL.XX.i16", ADD_TWO_SRC
},
1211 { 0x0ea70, "SEL.XY.i16", ADD_TWO_SRC
},
1212 { 0x0ea68, "SEL.YX.i16", ADD_TWO_SRC
},
1213 { 0x0ea78, "SEL.YY.i16", ADD_TWO_SRC
},
1214 { 0x0ec00, "F32_TO_F16", ADD_TWO_SRC
},
1215 { 0x0f640, "ICMP.GL.GT", ADD_TWO_SRC
}, // src0 > src1 ? 1 : 0
1216 { 0x0f648, "ICMP.GL.GE", ADD_TWO_SRC
},
1217 { 0x0f650, "UCMP.GL.GT", ADD_TWO_SRC
},
1218 { 0x0f658, "UCMP.GL.GE", ADD_TWO_SRC
},
1219 { 0x0f660, "ICMP.GL.EQ", ADD_TWO_SRC
},
1220 { 0x0f669, "ICMP.GL.NEQ", ADD_TWO_SRC
},
1221 { 0x0f6c0, "ICMP.D3D.GT", ADD_TWO_SRC
}, // src0 > src1 ? ~0 : 0
1222 { 0x0f6c8, "ICMP.D3D.GE", ADD_TWO_SRC
},
1223 { 0x0f6d0, "UCMP.D3D.GT", ADD_TWO_SRC
},
1224 { 0x0f6d8, "UCMP.D3D.GE", ADD_TWO_SRC
},
1225 { 0x0f6e0, "ICMP.D3D.EQ", ADD_TWO_SRC
},
1226 { 0x10000, "MAX.v2f16", ADD_FMINMAX16
},
1227 { 0x11000, "ADD_MSCALE.f32", ADD_FADDMscale
},
1228 { 0x12000, "MIN.v2f16", ADD_FMINMAX16
},
1229 { 0x14000, "ADD.v2f16", ADD_FADD16
},
1230 { 0x17000, "FCMP.D3D", ADD_FCMP16
},
1231 { 0x178c0, "ADD.i32", ADD_TWO_SRC
},
1232 { 0x17900, "ADD.v2i16", ADD_TWO_SRC
},
1233 { 0x17ac0, "SUB.i32", ADD_TWO_SRC
},
1234 { 0x17c10, "ADDC.i32", ADD_TWO_SRC
}, // adds src0 to the bottom bit of src1
1235 { 0x17d80, "ADD.i32.i16.X", ADD_TWO_SRC
},
1236 { 0x17d90, "ADD.i32.u16.X", ADD_TWO_SRC
},
1237 { 0x17dc0, "ADD.i32.i16.Y", ADD_TWO_SRC
},
1238 { 0x17dd0, "ADD.i32.u16.Y", ADD_TWO_SRC
},
1239 { 0x18000, "LD_VAR_ADDR.f16", ADD_VARYING_ADDRESS
, true },
1240 { 0x18100, "LD_VAR_ADDR.f32", ADD_VARYING_ADDRESS
, true },
1241 { 0x18200, "LD_VAR_ADDR.i32", ADD_VARYING_ADDRESS
, true },
1242 { 0x18300, "LD_VAR_ADDR.u32", ADD_VARYING_ADDRESS
, true },
1243 { 0x19181, "DISCARD.FEQ.f32", ADD_TWO_SRC
, true },
1244 { 0x19189, "DISCARD.FNE.f32", ADD_TWO_SRC
, true },
1245 { 0x1918C, "DISCARD.GL.f32", ADD_TWO_SRC
, true }, /* Consumes ICMP.GL/etc with fixed 0 argument */
1246 { 0x19190, "DISCARD.FLE.f32", ADD_TWO_SRC
, true },
1247 { 0x19198, "DISCARD.FLT.f32", ADD_TWO_SRC
, true },
1248 { 0x191e8, "ATEST.f32", ADD_TWO_SRC
, true },
1249 { 0x191f0, "ATEST.X.f16", ADD_TWO_SRC
, true },
1250 { 0x191f8, "ATEST.Y.f16", ADD_TWO_SRC
, true },
1251 { 0x19300, "ST_VAR.v1", ADD_THREE_SRC
, true },
1252 { 0x19340, "ST_VAR.v2", ADD_THREE_SRC
, true },
1253 { 0x19380, "ST_VAR.v3", ADD_THREE_SRC
, true },
1254 { 0x193c0, "ST_VAR.v4", ADD_THREE_SRC
, true },
1255 { 0x1952c, "BLEND", ADD_BLENDING
, true },
1256 { 0x1a000, "LD_VAR.16", ADD_VARYING_INTERP
, true },
1257 { 0x1ae60, "TEX", ADD_TEX
, true },
1258 { 0x1c000, "RSHIFT_NAND.i32", ADD_THREE_SRC
},
1259 { 0x1c300, "RSHIFT_OR.i32", ADD_THREE_SRC
},
1260 { 0x1c400, "RSHIFT_AND.i32", ADD_THREE_SRC
},
1261 { 0x1c700, "RSHIFT_NOR.i32", ADD_THREE_SRC
},
1262 { 0x1c800, "LSHIFT_NAND.i32", ADD_THREE_SRC
},
1263 { 0x1cb00, "LSHIFT_OR.i32", ADD_THREE_SRC
},
1264 { 0x1cc00, "LSHIFT_AND.i32", ADD_THREE_SRC
},
1265 { 0x1cf00, "LSHIFT_NOR.i32", ADD_THREE_SRC
},
1266 { 0x1d000, "RSHIFT_XOR.i32", ADD_THREE_SRC
},
1267 { 0x1d100, "RSHIFT_XNOR.i32", ADD_THREE_SRC
},
1268 { 0x1d200, "LSHIFT_XOR.i32", ADD_THREE_SRC
},
1269 { 0x1d300, "LSHIFT_XNOR.i32", ADD_THREE_SRC
},
1270 { 0x1d400, "LSHIFT_ADD.i32", ADD_THREE_SRC
},
1271 { 0x1d500, "LSHIFT_SUB.i32", ADD_THREE_SRC
},
1272 { 0x1d500, "LSHIFT_RSUB.i32", ADD_THREE_SRC
},
1273 { 0x1d700, "RSHIFT_ADD.i32", ADD_THREE_SRC
},
1274 { 0x1d800, "RSHIFT_SUB.i32", ADD_THREE_SRC
},
1275 { 0x1d900, "RSHIFT_RSUB.i32", ADD_THREE_SRC
},
1276 { 0x1da00, "ARSHIFT_ADD.i32", ADD_THREE_SRC
},
1277 { 0x1db00, "ARSHIFT_SUB.i32", ADD_THREE_SRC
},
1278 { 0x1dc00, "ARSHIFT_RSUB.i32", ADD_THREE_SRC
},
1279 { 0x1dd18, "OR.i32", ADD_TWO_SRC
},
1280 { 0x1dd20, "AND.i32", ADD_TWO_SRC
},
1281 { 0x1dd60, "LSHIFT.i32", ADD_TWO_SRC
},
1282 { 0x1dd50, "XOR.i32", ADD_TWO_SRC
},
1283 { 0x1dd80, "RSHIFT.i32", ADD_TWO_SRC
},
1284 { 0x1dda0, "ARSHIFT.i32", ADD_TWO_SRC
},
1287 static struct add_op_info
find_add_op_info(unsigned op
)
1289 for (unsigned i
= 0; i
< ARRAY_SIZE(add_op_infos
); i
++) {
1290 unsigned opCmp
= ~0;
1291 switch (add_op_infos
[i
].src_type
) {
1308 opCmp
= op
& ~0x1fff;
1311 case ADD_FADDMscale
:
1312 opCmp
= op
& ~0xfff;
1316 opCmp
= op
& ~0x7ff;
1318 case ADD_TEX_COMPACT
:
1319 opCmp
= op
& ~0x3ff;
1321 case ADD_VARYING_INTERP
:
1322 opCmp
= op
& ~0x7ff;
1324 case ADD_VARYING_ADDRESS
:
1331 opCmp
= op
& ~0xfff;
1337 if (add_op_infos
[i
].op
== opCmp
)
1338 return add_op_infos
[i
];
1341 struct add_op_info info
;
1342 snprintf(info
.name
, sizeof(info
.name
), "op%04x", op
);
1344 info
.src_type
= ADD_TWO_SRC
;
1345 info
.has_data_reg
= true;
1349 static void dump_add(FILE *fp
, uint64_t word
, struct bifrost_regs regs
,
1350 struct bifrost_regs next_regs
, uint64_t *consts
,
1351 unsigned data_reg
, unsigned offset
, bool verbose
)
1354 fprintf(fp
, "# ADD: %016" PRIx64
"\n", word
);
1356 struct bifrost_add_inst ADD
;
1357 memcpy((char *) &ADD
, (char *) &word
, sizeof(ADD
));
1358 struct add_op_info info
= find_add_op_info(ADD
.op
);
1360 fprintf(fp
, "%s", info
.name
);
1362 // float16 seems like it doesn't support output modifiers
1363 if (info
.src_type
== ADD_FADD
|| info
.src_type
== ADD_FMINMAX
) {
1365 dump_output_mod(fp
, bits(ADD
.op
, 8, 10));
1366 if (info
.src_type
== ADD_FADD
)
1367 dump_round_mode(fp
, bits(ADD
.op
, 10, 12));
1369 dump_minmax_mode(fp
, bits(ADD
.op
, 10, 12));
1370 } else if (info
.src_type
== ADD_FCMP
|| info
.src_type
== ADD_FCMP16
) {
1371 dump_fcmp(fp
, bits(ADD
.op
, 3, 6));
1372 if (info
.src_type
== ADD_FCMP
)
1373 fprintf(fp
, ".f32");
1375 fprintf(fp
, ".v2f16");
1376 } else if (info
.src_type
== ADD_FADDMscale
) {
1377 switch ((ADD
.op
>> 6) & 0x7) {
1380 // causes GPU hangs on G71
1382 fprintf(fp
, ".invalid");
1384 // Same as usual outmod value.
1386 fprintf(fp
, ".clamp_0_1");
1388 // If src0 is infinite or NaN, flush it to zero so that the other
1389 // source is passed through unmodified.
1391 fprintf(fp
, ".flush_src0_inf_nan");
1395 fprintf(fp
, ".flush_src1_inf_nan");
1397 // Every other case seems to behave the same as the above?
1399 fprintf(fp
, ".unk%d", (ADD
.op
>> 6) & 0x7);
1402 } else if (info
.src_type
== ADD_VARYING_INTERP
) {
1404 fprintf(fp
, ".reuse");
1406 fprintf(fp
, ".flat");
1407 switch ((ADD
.op
>> 7) & 0x3) {
1409 fprintf(fp
, ".per_frag");
1412 fprintf(fp
, ".centroid");
1417 fprintf(fp
, ".explicit");
1420 fprintf(fp
, ".v%d", ((ADD
.op
>> 5) & 0x3) + 1);
1421 } else if (info
.src_type
== ADD_BRANCH
) {
1422 enum branch_code branchCode
= (enum branch_code
) ((ADD
.op
>> 6) & 0x3f);
1423 if (branchCode
== BR_ALWAYS
) {
1424 // unconditional branch
1426 enum branch_cond cond
= (enum branch_cond
) ((ADD
.op
>> 6) & 0x7);
1427 enum branch_bit_size size
= (enum branch_bit_size
) ((ADD
.op
>> 9) & 0x7);
1428 bool portSwapped
= (ADD
.op
& 0x7) < ADD
.src0
;
1429 // See the comment in branch_bit_size
1430 if (size
== BR_SIZE_16YX0
)
1432 if (size
== BR_SIZE_16YX1
)
1433 portSwapped
= false;
1434 // These sizes are only for floating point comparisons, so the
1435 // non-floating-point comparisons are reused to encode the flipped
1437 if (size
== BR_SIZE_32_AND_16X
|| size
== BR_SIZE_32_AND_16Y
)
1438 portSwapped
= false;
1439 // There's only one argument, so we reuse the extra argument to
1441 if (size
== BR_SIZE_ZERO
)
1442 portSwapped
= !(ADD
.op
& 1);
1447 fprintf(fp
, ".LT.u");
1449 fprintf(fp
, ".LT.i");
1452 if (size
== BR_SIZE_32_AND_16X
|| size
== BR_SIZE_32_AND_16Y
) {
1453 fprintf(fp
, ".UNE.f");
1456 fprintf(fp
, ".LE.u");
1458 fprintf(fp
, ".LE.i");
1463 fprintf(fp
, ".GT.u");
1465 fprintf(fp
, ".GT.i");
1469 fprintf(fp
, ".GE.u");
1471 fprintf(fp
, ".GE.i");
1475 fprintf(fp
, ".NE.i");
1477 fprintf(fp
, ".EQ.i");
1481 fprintf(fp
, ".UNE.f");
1483 fprintf(fp
, ".OEQ.f");
1487 fprintf(fp
, ".OGT.unk.f");
1489 fprintf(fp
, ".OGT.f");
1493 fprintf(fp
, ".OLT.unk.f");
1495 fprintf(fp
, ".OLT.f");
1500 case BR_SIZE_32_AND_16X
:
1501 case BR_SIZE_32_AND_16Y
:
1510 case BR_SIZE_ZERO
: {
1511 unsigned ctrl
= (ADD
.op
>> 1) & 0x3;
1513 fprintf(fp
, "32.Z");
1515 fprintf(fp
, "16.Z");
1523 struct bifrost_reg_ctrl next_ctrl
= DecodeRegCtrl(fp
, next_regs
);
1524 if (next_ctrl
.add_write_unit
!= REG_WRITE_NONE
) {
1525 fprintf(fp
, "{R%d, T1}, ", GetRegToWrite(next_ctrl
.add_write_unit
, next_regs
));
1527 fprintf(fp
, "T1, ");
1530 switch (info
.src_type
) {
1532 // Note: in this case, regs.uniform_const == location | 0x8
1533 // This probably means we can't load uniforms or immediates in the
1534 // same instruction. This re-uses the encoding that normally means
1535 // "disabled", where the low 4 bits are ignored. Perhaps the extra
1536 // 0x8 or'd in indicates this is happening.
1537 fprintf(fp
, "location:%d, ", regs
.uniform_const
& 0x7);
1540 dump_src(fp
, ADD
.src0
, regs
, consts
, false);
1543 case ADD_TEX_COMPACT
: {
1546 bool dualTex
= false;
1547 if (info
.src_type
== ADD_TEX_COMPACT
) {
1548 tex_index
= (ADD
.op
>> 3) & 0x7;
1549 sampler_index
= (ADD
.op
>> 7) & 0x7;
1550 bool unknown
= (ADD
.op
& 0x40);
1551 // TODO: figure out if the unknown bit is ever 0
1553 fprintf(fp
, "unknown ");
1555 uint64_t constVal
= get_const(consts
, regs
);
1556 uint32_t controlBits
= (ADD
.op
& 0x8) ? (constVal
>> 32) : constVal
;
1557 struct bifrost_tex_ctrl ctrl
;
1558 memcpy((char *) &ctrl
, (char *) &controlBits
, sizeof(ctrl
));
1560 // TODO: figure out what actually triggers dual-tex
1561 if (ctrl
.result_type
== 9) {
1562 struct bifrost_dual_tex_ctrl dualCtrl
;
1563 memcpy((char *) &dualCtrl
, (char *) &controlBits
, sizeof(ctrl
));
1564 fprintf(fp
, "(dualtex) tex0:%d samp0:%d tex1:%d samp1:%d ",
1565 dualCtrl
.tex_index0
, dualCtrl
.sampler_index0
,
1566 dualCtrl
.tex_index1
, dualCtrl
.sampler_index1
);
1567 if (dualCtrl
.unk0
!= 3)
1568 fprintf(fp
, "unk:%d ", dualCtrl
.unk0
);
1571 if (ctrl
.no_merge_index
) {
1572 tex_index
= ctrl
.tex_index
;
1573 sampler_index
= ctrl
.sampler_index
;
1575 tex_index
= sampler_index
= ctrl
.tex_index
;
1576 unsigned unk
= ctrl
.sampler_index
>> 2;
1578 fprintf(fp
, "unk:%d ", unk
);
1579 if (ctrl
.sampler_index
& 1)
1581 if (ctrl
.sampler_index
& 2)
1586 fprintf(fp
, "unk0:%d ", ctrl
.unk0
);
1588 fprintf(fp
, "unk1 ");
1589 if (ctrl
.unk2
!= 0xf)
1590 fprintf(fp
, "unk2:%x ", ctrl
.unk2
);
1592 switch (ctrl
.result_type
) {
1594 fprintf(fp
, "f32 ");
1597 fprintf(fp
, "i32 ");
1600 fprintf(fp
, "u32 ");
1603 fprintf(fp
, "unktype(%x) ", ctrl
.result_type
);
1606 switch (ctrl
.tex_type
) {
1608 fprintf(fp
, "cube ");
1611 fprintf(fp
, "buffer ");
1622 fprintf(fp
, "shadow ");
1624 fprintf(fp
, "array ");
1627 if (ctrl
.calc_gradients
) {
1628 int comp
= (controlBits
>> 20) & 0x3;
1629 fprintf(fp
, "txg comp:%d ", comp
);
1631 fprintf(fp
, "txf ");
1634 if (!ctrl
.not_supply_lod
) {
1635 if (ctrl
.compute_lod
)
1636 fprintf(fp
, "lod_bias ");
1638 fprintf(fp
, "lod ");
1641 if (!ctrl
.calc_gradients
)
1642 fprintf(fp
, "grad ");
1645 if (ctrl
.texel_offset
)
1646 fprintf(fp
, "offset ");
1651 if (tex_index
== -1)
1652 fprintf(fp
, "tex:indirect ");
1654 fprintf(fp
, "tex:%d ", tex_index
);
1656 if (sampler_index
== -1)
1657 fprintf(fp
, "samp:indirect ");
1659 fprintf(fp
, "samp:%d ", sampler_index
);
1663 case ADD_VARYING_INTERP
: {
1664 unsigned addr
= ADD
.op
& 0x1f;
1665 if (addr
< 0b10100) {
1667 fprintf(fp
, "%d", addr
);
1668 } else if (addr
< 0b11000) {
1670 fprintf(fp
, "fragw");
1671 else if (addr
== 23)
1672 fprintf(fp
, "fragz");
1674 fprintf(fp
, "unk%d", addr
);
1676 dump_src(fp
, ADD
.op
& 0x7, regs
, consts
, false);
1679 dump_src(fp
, ADD
.src0
, regs
, consts
, false);
1682 case ADD_VARYING_ADDRESS
: {
1683 dump_src(fp
, ADD
.src0
, regs
, consts
, false);
1685 dump_src(fp
, ADD
.op
& 0x7, regs
, consts
, false);
1687 unsigned location
= (ADD
.op
>> 3) & 0x1f;
1688 if (location
< 16) {
1689 fprintf(fp
, "location:%d", location
);
1690 } else if (location
== 20) {
1691 fprintf(fp
, "location:%u", (uint32_t) get_const(consts
, regs
));
1692 } else if (location
== 21) {
1693 fprintf(fp
, "location:%u", (uint32_t) (get_const(consts
, regs
) >> 32));
1695 fprintf(fp
, "location:%d(unk)", location
);
1700 fprintf(fp
, "location:%d, ", (ADD
.op
>> 3) & 0xf);
1702 dump_src(fp
, ADD
.src0
, regs
, consts
, false);
1704 dump_src(fp
, ADD
.op
& 0x7, regs
, consts
, false);
1707 dump_src(fp
, ADD
.src0
, regs
, consts
, false);
1709 dump_src(fp
, ADD
.op
& 0x7, regs
, consts
, false);
1711 dump_src(fp
, (ADD
.op
>> 3) & 0x7, regs
, consts
, false);
1717 if (ADD
.op
& 0x1000)
1718 fprintf(fp
, "abs(");
1719 dump_src(fp
, ADD
.src0
, regs
, consts
, false);
1720 switch ((ADD
.op
>> 6) & 0x3) {
1727 if (ADD
.op
& 0x1000)
1733 fprintf(fp
, "abs(");
1734 dump_src(fp
, ADD
.op
& 0x7, regs
, consts
, false);
1735 switch ((ADD
.op
>> 6) & 0x3) {
1746 fprintf(fp
, ".unk");
1755 if (ADD
.op
& 0x1000)
1756 fprintf(fp
, "abs(");
1757 dump_src(fp
, ADD
.src0
, regs
, consts
, false);
1758 if (ADD
.op
& 0x1000)
1760 dump_16swizzle(fp
, (ADD
.op
>> 6) & 0x3);
1765 fprintf(fp
, "abs(");
1766 dump_src(fp
, ADD
.op
& 0x7, regs
, consts
, false);
1767 dump_16swizzle(fp
, (ADD
.op
>> 8) & 0x3);
1771 case ADD_FMINMAX16
: {
1772 bool abs1
= ADD
.op
& 0x8;
1773 bool abs2
= (ADD
.op
& 0x7) < ADD
.src0
;
1777 fprintf(fp
, "abs(");
1778 dump_src(fp
, ADD
.src0
, regs
, consts
, false);
1779 dump_16swizzle(fp
, (ADD
.op
>> 6) & 0x3);
1786 fprintf(fp
, "abs(");
1787 dump_src(fp
, ADD
.op
& 0x7, regs
, consts
, false);
1788 dump_16swizzle(fp
, (ADD
.op
>> 8) & 0x3);
1793 case ADD_FADDMscale
: {
1797 fprintf(fp
, "abs(");
1798 dump_src(fp
, ADD
.src0
, regs
, consts
, false);
1806 dump_src(fp
, ADD
.op
& 0x7, regs
, consts
, false);
1810 dump_src(fp
, (ADD
.op
>> 3) & 0x7, regs
, consts
, false);
1814 if (ADD
.op
& 0x400) {
1817 if (ADD
.op
& 0x100) {
1818 fprintf(fp
, "abs(");
1820 dump_src(fp
, ADD
.src0
, regs
, consts
, false);
1821 switch ((ADD
.op
>> 6) & 0x3) {
1828 if (ADD
.op
& 0x100) {
1832 if (ADD
.op
& 0x200) {
1833 fprintf(fp
, "abs(");
1835 dump_src(fp
, ADD
.op
& 0x7, regs
, consts
, false);
1836 switch ((ADD
.op
>> 6) & 0x3) {
1847 fprintf(fp
, ".unk");
1850 if (ADD
.op
& 0x200) {
1855 dump_src(fp
, ADD
.src0
, regs
, consts
, false);
1856 dump_16swizzle(fp
, (ADD
.op
>> 6) & 0x3);
1858 dump_src(fp
, ADD
.op
& 0x7, regs
, consts
, false);
1859 dump_16swizzle(fp
, (ADD
.op
>> 8) & 0x3);
1862 enum branch_code code
= (enum branch_code
) ((ADD
.op
>> 6) & 0x3f);
1863 enum branch_bit_size size
= (enum branch_bit_size
) ((ADD
.op
>> 9) & 0x7);
1864 if (code
!= BR_ALWAYS
) {
1865 dump_src(fp
, ADD
.src0
, regs
, consts
, false);
1875 case BR_SIZE_ZERO
: {
1876 unsigned ctrl
= (ADD
.op
>> 1) & 0x3;
1893 if (code
!= BR_ALWAYS
&& size
!= BR_SIZE_ZERO
) {
1894 dump_src(fp
, ADD
.op
& 0x7, regs
, consts
, false);
1899 case BR_SIZE_32_AND_16X
:
1903 case BR_SIZE_32_AND_16Y
:
1911 // I haven't had the chance to test if this actually specifies the
1912 // branch offset, since I couldn't get it to produce values other
1913 // than 5 (uniform/const high), but these three bits are always
1914 // consistent across branch instructions, so it makes sense...
1915 int offsetSrc
= (ADD
.op
>> 3) & 0x7;
1916 if (offsetSrc
== 4 || offsetSrc
== 5) {
1917 // If the offset is known/constant, we can decode it
1918 uint32_t raw_offset
;
1920 raw_offset
= get_const(consts
, regs
);
1922 raw_offset
= get_const(consts
, regs
) >> 32;
1923 // The high 4 bits are flags, while the rest is the
1924 // twos-complement offset in bytes (here we convert to
1926 int32_t branch_offset
= ((int32_t) raw_offset
<< 4) >> 8;
1928 // If high4 is the high 4 bits of the last 64-bit constant,
1929 // this is calculated as (high4 + 4) & 0xf, or 0 if the branch
1930 // offset itself is the last constant. Not sure if this is
1931 // actually used, or just garbage in unused bits, but in any
1932 // case, we can just ignore it here since it's redundant. Note
1933 // that if there is any padding, this will be 4 since the
1934 // padding counts as the last constant.
1935 unsigned flags
= raw_offset
>> 28;
1938 // Note: the offset is in bytes, relative to the beginning of the
1939 // current clause, so a zero offset would be a loop back to the
1940 // same clause (annoyingly different from Midgard).
1941 fprintf(fp
, "clause_%d", offset
+ branch_offset
);
1943 dump_src(fp
, offsetSrc
, regs
, consts
, false);
1947 if (info
.has_data_reg
) {
1948 fprintf(fp
, ", R%d", data_reg
);
1953 void dump_instr(FILE *fp
, const struct bifrost_alu_inst
*instr
,
1954 struct bifrost_regs next_regs
, uint64_t *consts
,
1955 unsigned data_reg
, unsigned offset
, bool verbose
)
1957 struct bifrost_regs regs
;
1958 memcpy((char *) ®s
, (char *) &instr
->reg_bits
, sizeof(regs
));
1961 fprintf(fp
, "# regs: %016" PRIx64
"\n", instr
->reg_bits
);
1962 dump_regs(fp
, regs
);
1964 dump_fma(fp
, instr
->fma_bits
, regs
, next_regs
, consts
, verbose
);
1965 dump_add(fp
, instr
->add_bits
, regs
, next_regs
, consts
, data_reg
, offset
, verbose
);
1968 bool dump_clause(FILE *fp
, uint32_t *words
, unsigned *size
, unsigned offset
, bool verbose
)
1970 // State for a decoded clause
1971 struct bifrost_alu_inst instrs
[8] = {};
1972 uint64_t consts
[6] = {};
1973 unsigned num_instrs
= 0;
1974 unsigned num_consts
= 0;
1975 uint64_t header_bits
= 0;
1976 bool stopbit
= false;
1979 for (i
= 0; ; i
++, words
+= 4) {
1982 for (int j
= 0; j
< 4; j
++)
1983 fprintf(fp
, "%08x ", words
[3 - j
]); // low bit on the right
1986 unsigned tag
= bits(words
[0], 0, 8);
1988 // speculatively decode some things that are common between many formats, so we can share some code
1989 struct bifrost_alu_inst main_instr
= {};
1991 main_instr
.add_bits
= bits(words
[2], 2, 32 - 13);
1993 main_instr
.fma_bits
= bits(words
[1], 11, 32) | bits(words
[2], 0, 2) << (32 - 11);
1995 main_instr
.reg_bits
= ((uint64_t) bits(words
[1], 0, 11)) << 24 | (uint64_t) bits(words
[0], 8, 32);
1997 uint64_t const0
= bits(words
[0], 8, 32) << 4 | (uint64_t) words
[1] << 28 | bits(words
[2], 0, 4) << 60;
1998 uint64_t const1
= bits(words
[2], 4, 32) << 4 | (uint64_t) words
[3] << 32;
2000 bool stop
= tag
& 0x40;
2003 fprintf(fp
, "# tag: 0x%02x\n", tag
);
2006 unsigned idx
= stop
? 5 : 2;
2007 main_instr
.add_bits
|= ((tag
>> 3) & 0x7) << 17;
2008 instrs
[idx
+ 1] = main_instr
;
2009 instrs
[idx
].add_bits
= bits(words
[3], 0, 17) | ((tag
& 0x7) << 17);
2010 instrs
[idx
].fma_bits
|= bits(words
[2], 19, 32) << 10;
2011 consts
[0] = bits(words
[3], 17, 32) << 4;
2014 switch ((tag
>> 3) & 0x7) {
2016 switch (tag
& 0x7) {
2018 main_instr
.add_bits
|= bits(words
[3], 29, 32) << 17;
2019 instrs
[1] = main_instr
;
2024 instrs
[2].add_bits
= bits(words
[3], 0, 17) | bits(words
[3], 29, 32) << 17;
2025 instrs
[2].fma_bits
|= bits(words
[2], 19, 32) << 10;
2033 instrs
[2].add_bits
= bits(words
[3], 0, 17) | bits(words
[3], 29, 32) << 17;
2034 instrs
[2].fma_bits
|= bits(words
[2], 19, 32) << 10;
2035 main_instr
.add_bits
|= bits(words
[3], 26, 29) << 17;
2036 instrs
[3] = main_instr
;
2037 if ((tag
& 0x7) == 0x5) {
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;
2051 instrs
[5].add_bits
= bits(words
[3], 0, 17) | bits(words
[3], 29, 32) << 17;
2052 instrs
[5].fma_bits
|= bits(words
[2], 19, 32) << 10;
2053 main_instr
.add_bits
|= bits(words
[3], 26, 29) << 17;
2054 instrs
[6] = main_instr
;
2059 fprintf(fp
, "unknown tag bits 0x%02x\n", tag
);
2064 unsigned idx
= ((tag
>> 3) & 0x7) == 2 ? 4 : 7;
2065 main_instr
.add_bits
|= (tag
& 0x7) << 17;
2066 instrs
[idx
] = main_instr
;
2067 consts
[0] |= (bits(words
[2], 19, 32) | ((uint64_t) words
[3] << 13)) << 19;
2069 num_instrs
= idx
+ 1;
2074 unsigned idx
= stop
? 4 : 1;
2075 main_instr
.add_bits
|= (tag
& 0x7) << 17;
2076 instrs
[idx
] = main_instr
;
2077 instrs
[idx
+ 1].fma_bits
|= bits(words
[3], 22, 32);
2078 instrs
[idx
+ 1].reg_bits
= bits(words
[2], 19, 32) | (bits(words
[3], 0, 22) << (32 - 19));
2082 // only constants can come after this
2086 header_bits
= bits(words
[2], 19, 32) | ((uint64_t) words
[3] << (32 - 19));
2087 main_instr
.add_bits
|= (tag
& 0x7) << 17;
2088 instrs
[0] = main_instr
;
2092 unsigned pos
= tag
& 0xf;
2093 // note that `pos' encodes both the total number of
2094 // instructions and the position in the constant stream,
2095 // presumably because decoded constants and instructions
2096 // share a buffer in the decoder, but we only care about
2097 // the position in the constant stream; the total number of
2098 // instructions is redundant.
2099 unsigned const_idx
= 0;
2126 fprintf(fp
, "# unknown pos 0x%x\n", pos
);
2130 if (num_consts
< const_idx
+ 2)
2131 num_consts
= const_idx
+ 2;
2133 consts
[const_idx
] = const0
;
2134 consts
[const_idx
+ 1] = const1
;
2150 fprintf(fp
, "# header: %012" PRIx64
"\n", header_bits
);
2153 struct bifrost_header header
;
2154 memcpy((char *) &header
, (char *) &header_bits
, sizeof(struct bifrost_header
));
2155 dump_header(fp
, header
, verbose
);
2156 if (!header
.no_end_of_shader
)
2160 for (i
= 0; i
< num_instrs
; i
++) {
2161 struct bifrost_regs next_regs
;
2162 if (i
+ 1 == num_instrs
) {
2163 memcpy((char *) &next_regs
, (char *) &instrs
[0].reg_bits
,
2166 memcpy((char *) &next_regs
, (char *) &instrs
[i
+ 1].reg_bits
,
2170 dump_instr(fp
, &instrs
[i
], next_regs
, consts
, header
.datareg
, offset
, verbose
);
2175 for (unsigned i
= 0; i
< num_consts
; i
++) {
2176 fprintf(fp
, "# const%d: %08" PRIx64
"\n", 2 * i
, consts
[i
] & 0xffffffff);
2177 fprintf(fp
, "# const%d: %08" PRIx64
"\n", 2 * i
+ 1, consts
[i
] >> 32);
2183 void disassemble_bifrost(FILE *fp
, uint8_t *code
, size_t size
, bool verbose
)
2185 uint32_t *words
= (uint32_t *) code
;
2186 uint32_t *words_end
= words
+ (size
/ 4);
2187 // used for displaying branch targets
2188 unsigned offset
= 0;
2189 while (words
!= words_end
) {
2190 // we don't know what the program-end bit is quite yet, so for now just
2191 // assume that an all-0 quadword is padding
2192 uint32_t zero
[4] = {};
2193 if (memcmp(words
, zero
, 4 * sizeof(uint32_t)) == 0)
2195 fprintf(fp
, "clause_%d:\n", offset
);
2197 if (dump_clause(fp
, words
, &size
, offset
, verbose
) == true) {