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
;
113 enum fma_src_type src_type
;
127 ADD_TEX_COMPACT
, // texture instruction with embedded sampler
128 ADD_TEX
, // texture instruction with sampler/etc. in uniform port
139 enum add_src_type src_type
;
143 struct bifrost_tex_ctrl
{
144 unsigned sampler_index
: 4; // also used to signal indirects
145 unsigned tex_index
: 7;
146 bool no_merge_index
: 1; // whether to merge (direct) sampler & texture indices
147 bool filter
: 1; // use the usual filtering pipeline (0 for texelFetch & textureGather)
149 bool texel_offset
: 1; // *Offset()
152 unsigned tex_type
: 2; // 2D, 3D, Cube, Buffer
153 bool compute_lod
: 1; // 0 for *Lod()
154 bool not_supply_lod
: 1; // 0 for *Lod() or when a bias is applied
155 bool calc_gradients
: 1; // 0 for *Grad()
157 unsigned result_type
: 4; // integer, unsigned, float TODO: why is this 4 bits?
161 struct bifrost_dual_tex_ctrl
{
162 unsigned sampler_index0
: 2;
164 unsigned tex_index0
: 2;
165 unsigned sampler_index1
: 2;
166 unsigned tex_index1
: 2;
170 enum branch_bit_size
{
174 // For the above combinations of bitsize and location, an extra bit is
175 // encoded via comparing the sources. The only possible source of ambiguity
176 // would be if the sources were the same, but then the branch condition
177 // would be always true or always false anyways, so we can ignore it. But
178 // this no longer works when comparing the y component to the x component,
179 // since it's valid to compare the y component of a source against its own
180 // x component. Instead, the extra bit is encoded via an extra bitsize.
183 BR_SIZE_32_AND_16X
= 5,
184 BR_SIZE_32_AND_16Y
= 6,
185 // Used for comparisons with zero and always-true, see below. I think this
186 // only works for integer comparisons.
190 void dump_header(struct bifrost_header header
, bool verbose
);
191 void dump_instr(const struct bifrost_alu_inst
*instr
, struct bifrost_regs next_regs
, uint64_t *consts
,
192 unsigned data_reg
, unsigned offset
, bool verbose
);
193 bool dump_clause(uint32_t *words
, unsigned *size
, unsigned offset
, bool verbose
);
195 void dump_header(struct bifrost_header header
, bool verbose
) {
196 if (header
.clause_type
!= 0) {
197 printf("id(%du) ", header
.scoreboard_index
);
200 if (header
.scoreboard_deps
!= 0) {
201 printf("next-wait(");
203 for (unsigned i
= 0; i
< 8; i
++) {
204 if (header
.scoreboard_deps
& (1 << i
)) {
215 if (header
.datareg_writebarrier
)
216 printf("data-reg-barrier ");
218 if (!header
.no_end_of_shader
)
221 if (!header
.back_to_back
) {
223 if (header
.branch_cond
)
224 printf("branch-cond ");
226 printf("branch-uncond ");
229 if (header
.elide_writes
)
232 if (header
.suppress_inf
)
233 printf("suppress-inf ");
234 if (header
.suppress_nan
)
235 printf("suppress-nan ");
251 printf("# clause type %d, next clause type %d\n",
252 header
.clause_type
, header
.next_clause_type
);
256 static struct bifrost_reg_ctrl
DecodeRegCtrl(struct bifrost_regs regs
)
258 struct bifrost_reg_ctrl decoded
= {};
260 if (regs
.ctrl
== 0) {
261 ctrl
= regs
.reg1
>> 2;
262 decoded
.read_reg0
= !(regs
.reg1
& 0x2);
263 decoded
.read_reg1
= false;
266 decoded
.read_reg0
= decoded
.read_reg1
= true;
270 decoded
.fma_write_unit
= REG_WRITE_TWO
;
274 decoded
.fma_write_unit
= REG_WRITE_TWO
;
275 decoded
.read_reg3
= true;
278 decoded
.read_reg3
= true;
281 decoded
.add_write_unit
= REG_WRITE_TWO
;
284 decoded
.add_write_unit
= REG_WRITE_TWO
;
285 decoded
.read_reg3
= true;
288 decoded
.clause_start
= true;
291 decoded
.fma_write_unit
= REG_WRITE_TWO
;
292 decoded
.clause_start
= true;
297 decoded
.read_reg3
= true;
298 decoded
.clause_start
= true;
301 decoded
.add_write_unit
= REG_WRITE_TWO
;
302 decoded
.clause_start
= true;
307 decoded
.fma_write_unit
= REG_WRITE_THREE
;
308 decoded
.add_write_unit
= REG_WRITE_TWO
;
311 printf("# unknown reg ctrl %d\n", ctrl
);
317 // Pass in the add_write_unit or fma_write_unit, and this returns which register
318 // the ADD/FMA units are writing to
319 static unsigned GetRegToWrite(enum bifrost_reg_write_unit unit
, struct bifrost_regs regs
)
324 case REG_WRITE_THREE
:
326 default: /* REG_WRITE_NONE */
332 static void dump_regs(struct bifrost_regs srcs
)
334 struct bifrost_reg_ctrl ctrl
= DecodeRegCtrl(srcs
);
337 printf("port 0: R%d ", get_reg0(srcs
));
339 printf("port 1: R%d ", get_reg1(srcs
));
341 if (ctrl
.fma_write_unit
== REG_WRITE_TWO
)
342 printf("port 2: R%d (write FMA) ", srcs
.reg2
);
343 else if (ctrl
.add_write_unit
== REG_WRITE_TWO
)
344 printf("port 2: R%d (write ADD) ", srcs
.reg2
);
346 if (ctrl
.fma_write_unit
== REG_WRITE_THREE
)
347 printf("port 3: R%d (write FMA) ", srcs
.reg3
);
348 else if (ctrl
.add_write_unit
== REG_WRITE_THREE
)
349 printf("port 3: R%d (write ADD) ", srcs
.reg3
);
350 else if (ctrl
.read_reg3
)
351 printf("port 3: R%d (read) ", srcs
.reg3
);
353 if (srcs
.uniform_const
) {
354 if (srcs
.uniform_const
& 0x80) {
355 printf("uniform: U%d", (srcs
.uniform_const
& 0x7f) * 2);
361 static void dump_const_imm(uint32_t imm
)
368 printf("0x%08x /* %f */", imm
, fi
.f
);
371 static uint64_t get_const(uint64_t *consts
, struct bifrost_regs srcs
)
373 unsigned low_bits
= srcs
.uniform_const
& 0xf;
375 switch (srcs
.uniform_const
>> 4) {
376 case 4: imm
= consts
[0]; break;
377 case 5: imm
= consts
[1]; break;
378 case 6: imm
= consts
[2]; break;
379 case 7: imm
= consts
[3]; break;
380 case 2: imm
= consts
[4]; break;
381 case 3: imm
= consts
[5]; break;
382 default: assert(0); break;
384 return imm
| low_bits
;
387 static void dump_uniform_const_src(struct bifrost_regs srcs
, uint64_t *consts
, bool high32
)
389 if (srcs
.uniform_const
& 0x80) {
390 unsigned uniform
= (srcs
.uniform_const
& 0x7f) * 2;
391 printf("U%d", uniform
+ (high32
? 1 : 0));
392 } else if (srcs
.uniform_const
>= 0x20) {
393 uint64_t imm
= get_const(consts
, srcs
);
395 dump_const_imm(imm
>> 32);
399 switch (srcs
.uniform_const
) {
400 case 0: printf("0"); break;
401 case 5: printf("atest-data"); break;
402 case 6: printf("sample-ptr"); break;
411 printf("blend-descriptor%u", (unsigned) srcs
.uniform_const
- 8);
414 printf("unkConst%u", (unsigned) srcs
.uniform_const
);
425 static void dump_src(unsigned src
, struct bifrost_regs srcs
, uint64_t *consts
, bool isFMA
)
428 case 0: printf("R%d", get_reg0(srcs
)); break;
429 case 1: printf("R%d", get_reg1(srcs
)); break;
430 case 2: printf("R%d", srcs
.reg3
); break;
435 printf("T"); // i.e. the output of FMA this cycle
438 dump_uniform_const_src(srcs
, consts
, false);
441 dump_uniform_const_src(srcs
, consts
, true);
443 case 6: printf("T0"); break;
444 case 7: printf("T1"); break;
448 static void dump_output_mod(unsigned mod
)
454 printf(".clamp_0_inf"); break; // max(out, 0)
456 printf(".clamp_m1_1"); break; // clamp(out, -1, 1)
458 printf(".clamp_0_1"); break; // clamp(out, 0, 1)
464 static void dump_minmax_mode(unsigned mod
)
468 /* Same as fmax() and fmin() -- return the other number if any
469 * number is NaN. Also always return +0 if one argument is +0 and
474 /* Instead of never returning a NaN, always return one. The
475 * "greater"/"lesser" NaN is always returned, first by checking the
476 * sign and then the mantissa bits.
478 printf(".nan_wins"); break;
480 /* For max, implement src0 > src1 ? src0 : src1
481 * For min, implement src0 < src1 ? src0 : src1
483 * This includes handling NaN's and signedness of 0 differently
484 * from above, since +0 and -0 compare equal and comparisons always
485 * return false for NaN's. As a result, this mode is *not*
488 printf(".src1_wins"); break;
490 /* For max, implement src0 < src1 ? src1 : src0
491 * For min, implement src0 > src1 ? src1 : src0
493 printf(".src0_wins"); break;
499 static void dump_round_mode(unsigned mod
)
503 /* roundTiesToEven, the IEEE default. */
506 /* roundTowardPositive in the IEEE spec. */
507 printf(".round_pos"); break;
509 /* roundTowardNegative in the IEEE spec. */
510 printf(".round_neg"); break;
512 /* roundTowardZero in the IEEE spec. */
513 printf(".round_zero"); break;
519 static const struct fma_op_info FMAOpInfos
[] = {
520 { 0x00000, "FMA.f32", FMA_FMA
},
521 { 0x40000, "MAX.f32", FMA_FMINMAX
},
522 { 0x44000, "MIN.f32", FMA_FMINMAX
},
523 { 0x48000, "FCMP.GL", FMA_FCMP
},
524 { 0x4c000, "FCMP.D3D", FMA_FCMP
},
525 { 0x4ff98, "ADD.i32", FMA_TWO_SRC
},
526 { 0x4ffd8, "SUB.i32", FMA_TWO_SRC
},
527 { 0x4fff0, "SUBB.i32", FMA_TWO_SRC
},
528 { 0x50000, "FMA_MSCALE", FMA_FMA_MSCALE
},
529 { 0x58000, "ADD.f32", FMA_FADD
},
530 { 0x5c000, "CSEL.FEQ.f32", FMA_FOUR_SRC
},
531 { 0x5c200, "CSEL.FGT.f32", FMA_FOUR_SRC
},
532 { 0x5c400, "CSEL.FGE.f32", FMA_FOUR_SRC
},
533 { 0x5c600, "CSEL.IEQ.f32", FMA_FOUR_SRC
},
534 { 0x5c800, "CSEL.IGT.i32", FMA_FOUR_SRC
},
535 { 0x5ca00, "CSEL.IGE.i32", FMA_FOUR_SRC
},
536 { 0x5cc00, "CSEL.UGT.i32", FMA_FOUR_SRC
},
537 { 0x5ce00, "CSEL.UGE.i32", FMA_FOUR_SRC
},
538 { 0x5d8d0, "ICMP.D3D.GT.v2i16", FMA_TWO_SRC
},
539 { 0x5d9d0, "UCMP.D3D.GT.v2i16", FMA_TWO_SRC
},
540 { 0x5dad0, "ICMP.D3D.GE.v2i16", FMA_TWO_SRC
},
541 { 0x5dbd0, "UCMP.D3D.GE.v2i16", FMA_TWO_SRC
},
542 { 0x5dcd0, "ICMP.D3D.EQ.v2i16", FMA_TWO_SRC
},
543 { 0x5de40, "ICMP.GL.GT.i32", FMA_TWO_SRC
}, // src0 > src1 ? 1 : 0
544 { 0x5de48, "ICMP.GL.GE.i32", FMA_TWO_SRC
},
545 { 0x5de50, "UCMP.GL.GT.i32", FMA_TWO_SRC
},
546 { 0x5de58, "UCMP.GL.GE.i32", FMA_TWO_SRC
},
547 { 0x5de60, "ICMP.GL.EQ.i32", FMA_TWO_SRC
},
548 { 0x5dec0, "ICMP.D3D.GT.i32", FMA_TWO_SRC
}, // src0 > src1 ? ~0 : 0
549 { 0x5dec8, "ICMP.D3D.GE.i32", FMA_TWO_SRC
},
550 { 0x5ded0, "UCMP.D3D.GT.i32", FMA_TWO_SRC
},
551 { 0x5ded8, "UCMP.D3D.GE.i32", FMA_TWO_SRC
},
552 { 0x5dee0, "ICMP.D3D.EQ.i32", FMA_TWO_SRC
},
553 { 0x60200, "RSHIFT_NAND.i32", FMA_THREE_SRC
},
554 { 0x603c0, "RSHIFT_NAND.v2i16", FMA_THREE_SRC
},
555 { 0x60e00, "RSHIFT_OR.i32", FMA_THREE_SRC
},
556 { 0x60fc0, "RSHIFT_OR.v2i16", FMA_THREE_SRC
},
557 { 0x61200, "RSHIFT_AND.i32", FMA_THREE_SRC
},
558 { 0x613c0, "RSHIFT_AND.v2i16", FMA_THREE_SRC
},
559 { 0x61e00, "RSHIFT_NOR.i32", FMA_THREE_SRC
}, // ~((src0 << src2) | src1)
560 { 0x61fc0, "RSHIFT_NOR.v2i16", FMA_THREE_SRC
}, // ~((src0 << src2) | src1)
561 { 0x62200, "LSHIFT_NAND.i32", FMA_THREE_SRC
},
562 { 0x623c0, "LSHIFT_NAND.v2i16", FMA_THREE_SRC
},
563 { 0x62e00, "LSHIFT_OR.i32", FMA_THREE_SRC
}, // (src0 << src2) | src1
564 { 0x62fc0, "LSHIFT_OR.v2i16", FMA_THREE_SRC
}, // (src0 << src2) | src1
565 { 0x63200, "LSHIFT_AND.i32", FMA_THREE_SRC
}, // (src0 << src2) & src1
566 { 0x633c0, "LSHIFT_AND.v2i16", FMA_THREE_SRC
},
567 { 0x63e00, "LSHIFT_NOR.i32", FMA_THREE_SRC
},
568 { 0x63fc0, "LSHIFT_NOR.v2i16", FMA_THREE_SRC
},
569 { 0x64200, "RSHIFT_XOR.i32", FMA_THREE_SRC
},
570 { 0x643c0, "RSHIFT_XOR.v2i16", FMA_THREE_SRC
},
571 { 0x64600, "RSHIFT_XNOR.i32", FMA_THREE_SRC
}, // ~((src0 >> src2) ^ src1)
572 { 0x647c0, "RSHIFT_XNOR.v2i16", FMA_THREE_SRC
}, // ~((src0 >> src2) ^ src1)
573 { 0x64a00, "LSHIFT_XOR.i32", FMA_THREE_SRC
},
574 { 0x64bc0, "LSHIFT_XOR.v2i16", FMA_THREE_SRC
},
575 { 0x64e00, "LSHIFT_XNOR.i32", FMA_THREE_SRC
}, // ~((src0 >> src2) ^ src1)
576 { 0x64fc0, "LSHIFT_XNOR.v2i16", FMA_THREE_SRC
}, // ~((src0 >> src2) ^ src1)
577 { 0x65200, "LSHIFT_ADD.i32", FMA_THREE_SRC
},
578 { 0x65600, "LSHIFT_SUB.i32", FMA_THREE_SRC
}, // (src0 << src2) - src1
579 { 0x65a00, "LSHIFT_RSUB.i32", FMA_THREE_SRC
}, // src1 - (src0 << src2)
580 { 0x65e00, "RSHIFT_ADD.i32", FMA_THREE_SRC
},
581 { 0x66200, "RSHIFT_SUB.i32", FMA_THREE_SRC
},
582 { 0x66600, "RSHIFT_RSUB.i32", FMA_THREE_SRC
},
583 { 0x66a00, "ARSHIFT_ADD.i32", FMA_THREE_SRC
},
584 { 0x66e00, "ARSHIFT_SUB.i32", FMA_THREE_SRC
},
585 { 0x67200, "ARSHIFT_RSUB.i32", FMA_THREE_SRC
},
586 { 0x80000, "FMA.v2f16", FMA_FMA16
},
587 { 0xc0000, "MAX.v2f16", FMA_FMINMAX16
},
588 { 0xc4000, "MIN.v2f16", FMA_FMINMAX16
},
589 { 0xc8000, "FCMP.GL", FMA_FCMP16
},
590 { 0xcc000, "FCMP.D3D", FMA_FCMP16
},
591 { 0xcf900, "ADD.v2i16", FMA_TWO_SRC
},
592 { 0xcfc10, "ADDC.i32", FMA_TWO_SRC
},
593 { 0xcfd80, "ADD.i32.i16.X", FMA_TWO_SRC
},
594 { 0xcfd90, "ADD.i32.u16.X", FMA_TWO_SRC
},
595 { 0xcfdc0, "ADD.i32.i16.Y", FMA_TWO_SRC
},
596 { 0xcfdd0, "ADD.i32.u16.Y", FMA_TWO_SRC
},
597 { 0xd8000, "ADD.v2f16", FMA_FADD16
},
598 { 0xdc000, "CSEL.FEQ.v2f16", FMA_FOUR_SRC
},
599 { 0xdc200, "CSEL.FGT.v2f16", FMA_FOUR_SRC
},
600 { 0xdc400, "CSEL.FGE.v2f16", FMA_FOUR_SRC
},
601 { 0xdc600, "CSEL.IEQ.v2f16", FMA_FOUR_SRC
},
602 { 0xdc800, "CSEL.IGT.v2i16", FMA_FOUR_SRC
},
603 { 0xdca00, "CSEL.IGE.v2i16", FMA_FOUR_SRC
},
604 { 0xdcc00, "CSEL.UGT.v2i16", FMA_FOUR_SRC
},
605 { 0xdce00, "CSEL.UGE.v2i16", FMA_FOUR_SRC
},
606 { 0xdd000, "F32_TO_F16", FMA_TWO_SRC
},
607 { 0xe0046, "F16_TO_I16.XX", FMA_ONE_SRC
},
608 { 0xe0047, "F16_TO_U16.XX", FMA_ONE_SRC
},
609 { 0xe004e, "F16_TO_I16.YX", FMA_ONE_SRC
},
610 { 0xe004f, "F16_TO_U16.YX", FMA_ONE_SRC
},
611 { 0xe0056, "F16_TO_I16.XY", FMA_ONE_SRC
},
612 { 0xe0057, "F16_TO_U16.XY", FMA_ONE_SRC
},
613 { 0xe005e, "F16_TO_I16.YY", FMA_ONE_SRC
},
614 { 0xe005f, "F16_TO_U16.YY", FMA_ONE_SRC
},
615 { 0xe00c0, "I16_TO_F16.XX", FMA_ONE_SRC
},
616 { 0xe00c1, "U16_TO_F16.XX", FMA_ONE_SRC
},
617 { 0xe00c8, "I16_TO_F16.YX", FMA_ONE_SRC
},
618 { 0xe00c9, "U16_TO_F16.YX", FMA_ONE_SRC
},
619 { 0xe00d0, "I16_TO_F16.XY", FMA_ONE_SRC
},
620 { 0xe00d1, "U16_TO_F16.XY", FMA_ONE_SRC
},
621 { 0xe00d8, "I16_TO_F16.YY", FMA_ONE_SRC
},
622 { 0xe00d9, "U16_TO_F16.YY", FMA_ONE_SRC
},
623 { 0xe0136, "F32_TO_I32", FMA_ONE_SRC
},
624 { 0xe0137, "F32_TO_U32", FMA_ONE_SRC
},
625 { 0xe0178, "I32_TO_F32", FMA_ONE_SRC
},
626 { 0xe0179, "U32_TO_F32", FMA_ONE_SRC
},
627 { 0xe0198, "I16_TO_I32.X", FMA_ONE_SRC
},
628 { 0xe0199, "U16_TO_U32.X", FMA_ONE_SRC
},
629 { 0xe019a, "I16_TO_I32.Y", FMA_ONE_SRC
},
630 { 0xe019b, "U16_TO_U32.Y", FMA_ONE_SRC
},
631 { 0xe019c, "I16_TO_F32.X", FMA_ONE_SRC
},
632 { 0xe019d, "U16_TO_F32.X", FMA_ONE_SRC
},
633 { 0xe019e, "I16_TO_F32.Y", FMA_ONE_SRC
},
634 { 0xe019f, "U16_TO_F32.Y", FMA_ONE_SRC
},
635 { 0xe01a2, "F16_TO_F32.X", FMA_ONE_SRC
},
636 { 0xe01a3, "F16_TO_F32.Y", FMA_ONE_SRC
},
637 { 0xe032c, "NOP", FMA_ONE_SRC
},
638 { 0xe032d, "MOV", FMA_ONE_SRC
},
639 { 0xe032f, "SWZ.YY.v2i16", FMA_ONE_SRC
},
640 // From the ARM patent US20160364209A1:
641 // "Decompose v (the input) into numbers x1 and s such that v = x1 * 2^s,
642 // and x1 is a floating point value in a predetermined range where the
643 // value 1 is within the range and not at one extremity of the range (e.g.
644 // choose a range where 1 is towards middle of range)."
647 { 0xe0345, "LOG_FREXPM", FMA_ONE_SRC
},
648 // Given a floating point number m * 2^e, returns m * 2^{-1}. This is
649 // exactly the same as the mantissa part of frexp().
650 { 0xe0365, "FRCP_FREXPM", FMA_ONE_SRC
},
651 // Given a floating point number m * 2^e, returns m * 2^{-2} if e is even,
652 // and m * 2^{-1} if e is odd. In other words, scales by powers of 4 until
653 // within the range [0.25, 1). Used for square-root and reciprocal
655 { 0xe0375, "FSQRT_FREXPM", FMA_ONE_SRC
},
656 // Given a floating point number m * 2^e, computes -e - 1 as an integer.
657 // Zero and infinity/NaN return 0.
658 { 0xe038d, "FRCP_FREXPE", FMA_ONE_SRC
},
659 // Computes floor(e/2) + 1.
660 { 0xe03a5, "FSQRT_FREXPE", FMA_ONE_SRC
},
661 // Given a floating point number m * 2^e, computes -floor(e/2) - 1 as an
663 { 0xe03ad, "FRSQ_FREXPE", FMA_ONE_SRC
},
664 { 0xe03c5, "LOG_FREXPE", FMA_ONE_SRC
},
665 { 0xe03fa, "CLZ", FMA_ONE_SRC
},
666 { 0xe0b80, "IMAX3", FMA_THREE_SRC
},
667 { 0xe0bc0, "UMAX3", FMA_THREE_SRC
},
668 { 0xe0c00, "IMIN3", FMA_THREE_SRC
},
669 { 0xe0c40, "UMIN3", FMA_THREE_SRC
},
670 { 0xe0ec5, "ROUND", FMA_ONE_SRC
},
671 { 0xe0f40, "CSEL", FMA_THREE_SRC
}, // src2 != 0 ? src1 : src0
672 { 0xe0fc0, "MUX.i32", FMA_THREE_SRC
}, // see ADD comment
673 { 0xe1805, "ROUNDEVEN", FMA_ONE_SRC
},
674 { 0xe1845, "CEIL", FMA_ONE_SRC
},
675 { 0xe1885, "FLOOR", FMA_ONE_SRC
},
676 { 0xe18c5, "TRUNC", FMA_ONE_SRC
},
677 { 0xe19b0, "ATAN_LDEXP.Y.f32", FMA_TWO_SRC
},
678 { 0xe19b8, "ATAN_LDEXP.X.f32", FMA_TWO_SRC
},
679 // These instructions in the FMA slot, together with LSHIFT_ADD_HIGH32.i32
680 // in the ADD slot, allow one to do a 64-bit addition with an extra small
681 // shift on one of the sources. There are three possible scenarios:
683 // 1) Full 64-bit addition. Do:
684 // out.x = LSHIFT_ADD_LOW32.i64 src1.x, src2.x, shift
685 // out.y = LSHIFT_ADD_HIGH32.i32 src1.y, src2.y
687 // The shift amount is applied to src2 before adding. The shift amount, and
688 // any extra bits from src2 plus the overflow bit, are sent directly from
689 // FMA to ADD instead of being passed explicitly. Hence, these two must be
690 // bundled together into the same instruction.
692 // 2) Add a 64-bit value src1 to a zero-extended 32-bit value src2. Do:
693 // out.x = LSHIFT_ADD_LOW32.u32 src1.x, src2, shift
694 // out.y = LSHIFT_ADD_HIGH32.i32 src1.x, 0
696 // Note that in this case, the second argument to LSHIFT_ADD_HIGH32 is
697 // ignored, so it can actually be anything. As before, the shift is applied
698 // to src2 before adding.
700 // 3) Add a 64-bit value to a sign-extended 32-bit value src2. Do:
701 // out.x = LSHIFT_ADD_LOW32.i32 src1.x, src2, shift
702 // out.y = LSHIFT_ADD_HIGH32.i32 src1.x, 0
704 // The only difference is the .i32 instead of .u32. Otherwise, this is
705 // exactly the same as before.
707 // In all these instructions, the shift amount is stored where the third
708 // source would be, so the shift has to be a small immediate from 0 to 7.
709 // This is fine for the expected use-case of these instructions, which is
710 // manipulating 64-bit pointers.
712 // These instructions can also be combined with various load/store
713 // instructions which normally take a 64-bit pointer in order to add a
714 // 32-bit or 64-bit offset to the pointer before doing the operation,
715 // optionally shifting the offset. The load/store op implicity does
716 // LSHIFT_ADD_HIGH32.i32 internally. Letting ptr be the pointer, and offset
717 // the desired offset, the cases go as follows:
719 // 1) Add a 64-bit offset:
720 // LSHIFT_ADD_LOW32.i64 ptr.x, offset.x, shift
721 // ld_st_op ptr.y, offset.y, ...
723 // Note that the output of LSHIFT_ADD_LOW32.i64 is not used, instead being
724 // implicitly sent to the load/store op to serve as the low 32 bits of the
727 // 2) Add a 32-bit unsigned offset:
728 // temp = LSHIFT_ADD_LOW32.u32 ptr.x, offset, shift
729 // ld_st_op temp, ptr.y, ...
731 // Now, the low 32 bits of offset << shift + ptr are passed explicitly to
732 // the ld_st_op, to match the case where there is no offset and ld_st_op is
735 // 3) Add a 32-bit signed offset:
736 // temp = LSHIFT_ADD_LOW32.i32 ptr.x, offset, shift
737 // ld_st_op temp, ptr.y, ...
739 // Again, the same as the unsigned case except for the offset.
740 { 0xe1c80, "LSHIFT_ADD_LOW32.u32", FMA_SHIFT_ADD64
},
741 { 0xe1cc0, "LSHIFT_ADD_LOW32.i64", FMA_SHIFT_ADD64
},
742 { 0xe1d80, "LSHIFT_ADD_LOW32.i32", FMA_SHIFT_ADD64
},
743 { 0xe1e00, "SEL.XX.i16", FMA_TWO_SRC
},
744 { 0xe1e08, "SEL.YX.i16", FMA_TWO_SRC
},
745 { 0xe1e10, "SEL.XY.i16", FMA_TWO_SRC
},
746 { 0xe1e18, "SEL.YY.i16", FMA_TWO_SRC
},
747 { 0xe7800, "IMAD", FMA_THREE_SRC
},
748 { 0xe78db, "POPCNT", FMA_ONE_SRC
},
751 static struct fma_op_info
find_fma_op_info(unsigned op
)
753 for (unsigned i
= 0; i
< ARRAY_SIZE(FMAOpInfos
); i
++) {
755 switch (FMAOpInfos
[i
].src_type
) {
764 opCmp
= op
& ~0x1fff;
767 case FMA_SHIFT_ADD64
:
774 opCmp
= op
& ~0x3fff;
778 opCmp
= op
& ~0x3ffff;
784 opCmp
= op
& ~0x7fff;
790 if (FMAOpInfos
[i
].op
== opCmp
)
791 return FMAOpInfos
[i
];
794 struct fma_op_info info
;
795 snprintf(info
.name
, sizeof(info
.name
), "op%04x", op
);
797 info
.src_type
= FMA_THREE_SRC
;
801 static void dump_fcmp(unsigned op
)
823 printf(".unk%d", op
);
828 static void dump_16swizzle(unsigned swiz
)
832 printf(".%c%c", "xy"[swiz
& 1], "xy"[(swiz
>> 1) & 1]);
835 static void dump_fma_expand_src0(unsigned ctrl
)
857 static void dump_fma_expand_src1(unsigned ctrl
)
879 static void dump_fma(uint64_t word
, struct bifrost_regs regs
, struct bifrost_regs next_regs
, uint64_t *consts
, bool verbose
)
882 printf("# FMA: %016" PRIx64
"\n", word
);
884 struct bifrost_fma_inst FMA
;
885 memcpy((char *) &FMA
, (char *) &word
, sizeof(struct bifrost_fma_inst
));
886 struct fma_op_info info
= find_fma_op_info(FMA
.op
);
888 printf("%s", info
.name
);
889 if (info
.src_type
== FMA_FADD
||
890 info
.src_type
== FMA_FMINMAX
||
891 info
.src_type
== FMA_FMA
||
892 info
.src_type
== FMA_FADD16
||
893 info
.src_type
== FMA_FMINMAX16
||
894 info
.src_type
== FMA_FMA16
) {
895 dump_output_mod(bits(FMA
.op
, 12, 14));
896 switch (info
.src_type
) {
901 dump_round_mode(bits(FMA
.op
, 10, 12));
905 dump_minmax_mode(bits(FMA
.op
, 10, 12));
910 } else if (info
.src_type
== FMA_FCMP
|| info
.src_type
== FMA_FCMP16
) {
911 dump_fcmp(bits(FMA
.op
, 10, 13));
912 if (info
.src_type
== FMA_FCMP
)
916 } else if (info
.src_type
== FMA_FMA_MSCALE
) {
917 if (FMA
.op
& (1 << 11)) {
918 switch ((FMA
.op
>> 9) & 0x3) {
920 /* This mode seems to do a few things:
921 * - Makes 0 * infinity (and incidentally 0 * nan) return 0,
922 * since generating a nan would poison the result of
923 * 1/infinity and 1/0.
924 * - Fiddles with which nan is returned in nan * nan,
925 * presumably to make sure that the same exact nan is
926 * returned for 1/nan.
931 /* Similar to the above, but src0 always wins when multiplying
934 printf(".sqrt_mode");
937 printf(".unk%d_mode", (int) (FMA
.op
>> 9) & 0x3);
940 dump_output_mod(bits(FMA
.op
, 9, 11));
946 struct bifrost_reg_ctrl next_ctrl
= DecodeRegCtrl(next_regs
);
947 if (next_ctrl
.fma_write_unit
!= REG_WRITE_NONE
) {
948 printf("{R%d, T0}, ", GetRegToWrite(next_ctrl
.fma_write_unit
, next_regs
));
953 switch (info
.src_type
) {
955 dump_src(FMA
.src0
, regs
, consts
, true);
958 dump_src(FMA
.src0
, regs
, consts
, true);
960 dump_src(FMA
.op
& 0x7, regs
, consts
, true);
968 dump_src(FMA
.src0
, regs
, consts
, true);
969 dump_fma_expand_src0((FMA
.op
>> 6) & 0x7);
977 dump_src(FMA
.op
& 0x7, regs
, consts
, true);
978 dump_fma_expand_src1((FMA
.op
>> 6) & 0x7);
983 case FMA_FMINMAX16
: {
984 bool abs1
= FMA
.op
& 0x8;
985 bool abs2
= (FMA
.op
& 0x7) < FMA
.src0
;
990 dump_src(FMA
.src0
, regs
, consts
, true);
991 dump_16swizzle((FMA
.op
>> 6) & 0x3);
999 dump_src(FMA
.op
& 0x7, regs
, consts
, true);
1000 dump_16swizzle((FMA
.op
>> 8) & 0x3);
1008 dump_src(FMA
.src0
, regs
, consts
, true);
1009 dump_fma_expand_src0((FMA
.op
>> 6) & 0x7);
1017 dump_src(FMA
.op
& 0x7, regs
, consts
, true);
1018 dump_fma_expand_src1((FMA
.op
>> 6) & 0x7);
1023 dump_src(FMA
.src0
, regs
, consts
, true);
1024 // Note: this is kinda a guess, I haven't seen the blob set this to
1025 // anything other than the identity, but it matches FMA_TWO_SRCFmod16
1026 dump_16swizzle((FMA
.op
>> 6) & 0x3);
1028 dump_src(FMA
.op
& 0x7, regs
, consts
, true);
1029 dump_16swizzle((FMA
.op
>> 8) & 0x3);
1031 case FMA_SHIFT_ADD64
:
1032 dump_src(FMA
.src0
, regs
, consts
, true);
1034 dump_src(FMA
.op
& 0x7, regs
, consts
, true);
1036 printf("shift:%u", (FMA
.op
>> 3) & 0x7);
1039 dump_src(FMA
.src0
, regs
, consts
, true);
1041 dump_src(FMA
.op
& 0x7, regs
, consts
, true);
1043 dump_src((FMA
.op
>> 3) & 0x7, regs
, consts
, true);
1046 if (FMA
.op
& (1 << 14))
1048 if (FMA
.op
& (1 << 9))
1050 dump_src(FMA
.src0
, regs
, consts
, true);
1051 dump_fma_expand_src0((FMA
.op
>> 6) & 0x7);
1052 if (FMA
.op
& (1 << 9))
1055 if (FMA
.op
& (1 << 16))
1057 dump_src(FMA
.op
& 0x7, regs
, consts
, true);
1058 dump_fma_expand_src1((FMA
.op
>> 6) & 0x7);
1059 if (FMA
.op
& (1 << 16))
1062 if (FMA
.op
& (1 << 15))
1064 if (FMA
.op
& (1 << 17))
1066 dump_src((FMA
.op
>> 3) & 0x7, regs
, consts
, true);
1067 if (FMA
.op
& (1 << 17))
1071 if (FMA
.op
& (1 << 14))
1073 dump_src(FMA
.src0
, regs
, consts
, true);
1074 dump_16swizzle((FMA
.op
>> 6) & 0x3);
1076 dump_src(FMA
.op
& 0x7, regs
, consts
, true);
1077 dump_16swizzle((FMA
.op
>> 8) & 0x3);
1079 if (FMA
.op
& (1 << 15))
1081 dump_src((FMA
.op
>> 3) & 0x7, regs
, consts
, true);
1082 dump_16swizzle((FMA
.op
>> 16) & 0x3);
1085 dump_src(FMA
.src0
, regs
, consts
, true);
1087 dump_src(FMA
.op
& 0x7, regs
, consts
, true);
1089 dump_src((FMA
.op
>> 3) & 0x7, regs
, consts
, true);
1091 dump_src((FMA
.op
>> 6) & 0x7, regs
, consts
, true);
1093 case FMA_FMA_MSCALE
:
1094 if (FMA
.op
& (1 << 12))
1096 dump_src(FMA
.src0
, regs
, consts
, true);
1097 if (FMA
.op
& (1 << 12))
1100 if (FMA
.op
& (1 << 13))
1102 dump_src(FMA
.op
& 0x7, regs
, consts
, true);
1104 if (FMA
.op
& (1 << 14))
1106 dump_src((FMA
.op
>> 3) & 0x7, regs
, consts
, true);
1108 dump_src((FMA
.op
>> 6) & 0x7, regs
, consts
, true);
1114 static const struct add_op_info add_op_infos
[] = {
1115 { 0x00000, "MAX.f32", ADD_FMINMAX
},
1116 { 0x02000, "MIN.f32", ADD_FMINMAX
},
1117 { 0x04000, "ADD.f32", ADD_FADD
},
1118 { 0x06000, "FCMP.GL", ADD_FCMP
},
1119 { 0x07000, "FCMP.D3D", ADD_FCMP
},
1120 { 0x07856, "F16_TO_I16", ADD_ONE_SRC
},
1121 { 0x07857, "F16_TO_U16", ADD_ONE_SRC
},
1122 { 0x078c0, "I16_TO_F16.XX", ADD_ONE_SRC
},
1123 { 0x078c1, "U16_TO_F16.XX", ADD_ONE_SRC
},
1124 { 0x078c8, "I16_TO_F16.YX", ADD_ONE_SRC
},
1125 { 0x078c9, "U16_TO_F16.YX", ADD_ONE_SRC
},
1126 { 0x078d0, "I16_TO_F16.XY", ADD_ONE_SRC
},
1127 { 0x078d1, "U16_TO_F16.XY", ADD_ONE_SRC
},
1128 { 0x078d8, "I16_TO_F16.YY", ADD_ONE_SRC
},
1129 { 0x078d9, "U16_TO_F16.YY", ADD_ONE_SRC
},
1130 { 0x07936, "F32_TO_I32", ADD_ONE_SRC
},
1131 { 0x07937, "F32_TO_U32", ADD_ONE_SRC
},
1132 { 0x07978, "I32_TO_F32", ADD_ONE_SRC
},
1133 { 0x07979, "U32_TO_F32", ADD_ONE_SRC
},
1134 { 0x07998, "I16_TO_I32.X", ADD_ONE_SRC
},
1135 { 0x07999, "U16_TO_U32.X", ADD_ONE_SRC
},
1136 { 0x0799a, "I16_TO_I32.Y", ADD_ONE_SRC
},
1137 { 0x0799b, "U16_TO_U32.Y", ADD_ONE_SRC
},
1138 { 0x0799c, "I16_TO_F32.X", ADD_ONE_SRC
},
1139 { 0x0799d, "U16_TO_F32.X", ADD_ONE_SRC
},
1140 { 0x0799e, "I16_TO_F32.Y", ADD_ONE_SRC
},
1141 { 0x0799f, "U16_TO_F32.Y", ADD_ONE_SRC
},
1142 // take the low 16 bits, and expand it to a 32-bit float
1143 { 0x079a2, "F16_TO_F32.X", ADD_ONE_SRC
},
1144 // take the high 16 bits, ...
1145 { 0x079a3, "F16_TO_F32.Y", ADD_ONE_SRC
},
1146 { 0x07b2b, "SWZ.YX.v2i16", ADD_ONE_SRC
},
1147 { 0x07b2c, "NOP", ADD_ONE_SRC
},
1148 { 0x07b29, "SWZ.XX.v2i16", ADD_ONE_SRC
},
1149 // Logically, this should be SWZ.XY, but that's equivalent to a move, and
1150 // this seems to be the canonical way the blob generates a MOV.
1151 { 0x07b2d, "MOV", ADD_ONE_SRC
},
1152 { 0x07b2f, "SWZ.YY.v2i16", ADD_ONE_SRC
},
1153 // Given a floating point number m * 2^e, returns m ^ 2^{-1}.
1154 { 0x07b65, "FRCP_FREXPM", ADD_ONE_SRC
},
1155 { 0x07b75, "FSQRT_FREXPM", ADD_ONE_SRC
},
1156 { 0x07b8d, "FRCP_FREXPE", ADD_ONE_SRC
},
1157 { 0x07ba5, "FSQRT_FREXPE", ADD_ONE_SRC
},
1158 { 0x07bad, "FRSQ_FREXPE", ADD_ONE_SRC
},
1159 // From the ARM patent US20160364209A1:
1160 // "Decompose v (the input) into numbers x1 and s such that v = x1 * 2^s,
1161 // and x1 is a floating point value in a predetermined range where the
1162 // value 1 is within the range and not at one extremity of the range (e.g.
1163 // choose a range where 1 is towards middle of range)."
1166 { 0x07bc5, "FLOG_FREXPE", ADD_ONE_SRC
},
1167 { 0x07d45, "CEIL", ADD_ONE_SRC
},
1168 { 0x07d85, "FLOOR", ADD_ONE_SRC
},
1169 { 0x07dc5, "TRUNC", ADD_ONE_SRC
},
1170 { 0x07f18, "LSHIFT_ADD_HIGH32.i32", ADD_TWO_SRC
},
1171 { 0x08000, "LD_ATTR.f16", ADD_LOAD_ATTR
, true },
1172 { 0x08100, "LD_ATTR.v2f16", ADD_LOAD_ATTR
, true },
1173 { 0x08200, "LD_ATTR.v3f16", ADD_LOAD_ATTR
, true },
1174 { 0x08300, "LD_ATTR.v4f16", ADD_LOAD_ATTR
, true },
1175 { 0x08400, "LD_ATTR.f32", ADD_LOAD_ATTR
, true },
1176 { 0x08500, "LD_ATTR.v3f32", ADD_LOAD_ATTR
, true },
1177 { 0x08600, "LD_ATTR.v3f32", ADD_LOAD_ATTR
, true },
1178 { 0x08700, "LD_ATTR.v4f32", ADD_LOAD_ATTR
, true },
1179 { 0x08800, "LD_ATTR.i32", ADD_LOAD_ATTR
, true },
1180 { 0x08900, "LD_ATTR.v3i32", ADD_LOAD_ATTR
, true },
1181 { 0x08a00, "LD_ATTR.v3i32", ADD_LOAD_ATTR
, true },
1182 { 0x08b00, "LD_ATTR.v4i32", ADD_LOAD_ATTR
, true },
1183 { 0x08c00, "LD_ATTR.u32", ADD_LOAD_ATTR
, true },
1184 { 0x08d00, "LD_ATTR.v3u32", ADD_LOAD_ATTR
, true },
1185 { 0x08e00, "LD_ATTR.v3u32", ADD_LOAD_ATTR
, true },
1186 { 0x08f00, "LD_ATTR.v4u32", ADD_LOAD_ATTR
, true },
1187 { 0x0a000, "LD_VAR.32", ADD_VARYING_INTERP
, true },
1188 { 0x0b000, "TEX", ADD_TEX_COMPACT
, true },
1189 { 0x0c188, "LOAD.i32", ADD_TWO_SRC
, true },
1190 { 0x0c1a0, "LD_UBO.i32", ADD_TWO_SRC
, true },
1191 { 0x0c1b8, "LD_SCRATCH.v2i32", ADD_TWO_SRC
, true },
1192 { 0x0c1c8, "LOAD.v2i32", ADD_TWO_SRC
, true },
1193 { 0x0c1e0, "LD_UBO.v2i32", ADD_TWO_SRC
, true },
1194 { 0x0c1f8, "LD_SCRATCH.v2i32", ADD_TWO_SRC
, true },
1195 { 0x0c208, "LOAD.v4i32", ADD_TWO_SRC
, true },
1196 // src0 = offset, src1 = binding
1197 { 0x0c220, "LD_UBO.v4i32", ADD_TWO_SRC
, true },
1198 { 0x0c238, "LD_SCRATCH.v4i32", ADD_TWO_SRC
, true },
1199 { 0x0c248, "STORE.v4i32", ADD_TWO_SRC
, true },
1200 { 0x0c278, "ST_SCRATCH.v4i32", ADD_TWO_SRC
, true },
1201 { 0x0c588, "STORE.i32", ADD_TWO_SRC
, true },
1202 { 0x0c5b8, "ST_SCRATCH.i32", ADD_TWO_SRC
, true },
1203 { 0x0c5c8, "STORE.v2i32", ADD_TWO_SRC
, true },
1204 { 0x0c5f8, "ST_SCRATCH.v2i32", ADD_TWO_SRC
, true },
1205 { 0x0c648, "LOAD.u16", ADD_TWO_SRC
, true }, // zero-extends
1206 { 0x0ca88, "LOAD.v3i32", ADD_TWO_SRC
, true },
1207 { 0x0caa0, "LD_UBO.v3i32", ADD_TWO_SRC
, true },
1208 { 0x0cab8, "LD_SCRATCH.v3i32", ADD_TWO_SRC
, true },
1209 { 0x0cb88, "STORE.v3i32", ADD_TWO_SRC
, true },
1210 { 0x0cbb8, "ST_SCRATCH.v3i32", ADD_TWO_SRC
, true },
1211 // *_FAST does not exist on G71 (added to G51, G72, and everything after)
1212 { 0x0cc00, "FRCP_FAST.f32", ADD_ONE_SRC
},
1213 { 0x0cc20, "FRSQ_FAST.f32", ADD_ONE_SRC
},
1214 // Given a floating point number m * 2^e, produces a table-based
1215 // approximation of 2/m using the top 17 bits. Includes special cases for
1216 // infinity, NaN, and zero, and copies the sign bit.
1217 { 0x0ce00, "FRCP_TABLE", ADD_ONE_SRC
},
1219 { 0x0ce10, "FRCP_FAST.f16.X", ADD_ONE_SRC
},
1220 // A similar table for inverse square root, using the high 17 bits of the
1221 // mantissa as well as the low bit of the exponent.
1222 { 0x0ce20, "FRSQ_TABLE", ADD_ONE_SRC
},
1223 { 0x0ce30, "FRCP_FAST.f16.Y", ADD_ONE_SRC
},
1224 { 0x0ce50, "FRSQ_FAST.f16.X", ADD_ONE_SRC
},
1225 // Used in the argument reduction for log. Given a floating-point number
1226 // m * 2^e, uses the top 4 bits of m to produce an approximation to 1/m
1227 // with the exponent forced to 0 and only the top 5 bits are nonzero. 0,
1228 // infinity, and NaN all return 1.0.
1229 // See the ARM patent for more information.
1230 { 0x0ce60, "FRCP_APPROX", ADD_ONE_SRC
},
1231 { 0x0ce70, "FRSQ_FAST.f16.Y", ADD_ONE_SRC
},
1232 { 0x0cf40, "ATAN_ASSIST", ADD_TWO_SRC
},
1233 { 0x0cf48, "ATAN_TABLE", ADD_TWO_SRC
},
1234 { 0x0cf50, "SIN_TABLE", ADD_ONE_SRC
},
1235 { 0x0cf51, "COS_TABLE", ADD_ONE_SRC
},
1236 { 0x0cf58, "EXP_TABLE", ADD_ONE_SRC
},
1237 { 0x0cf60, "FLOG2_TABLE", ADD_ONE_SRC
},
1238 { 0x0cf64, "FLOGE_TABLE", ADD_ONE_SRC
},
1239 { 0x0d000, "BRANCH", ADD_BRANCH
},
1240 // For each bit i, return src2[i] ? src0[i] : src1[i]. In other words, this
1241 // is the same as (src2 & src0) | (~src2 & src1).
1242 { 0x0e8c0, "MUX", ADD_THREE_SRC
},
1243 { 0x0e9b0, "ATAN_LDEXP.Y.f32", ADD_TWO_SRC
},
1244 { 0x0e9b8, "ATAN_LDEXP.X.f32", ADD_TWO_SRC
},
1245 { 0x0ea60, "SEL.XX.i16", ADD_TWO_SRC
},
1246 { 0x0ea70, "SEL.XY.i16", ADD_TWO_SRC
},
1247 { 0x0ea68, "SEL.YX.i16", ADD_TWO_SRC
},
1248 { 0x0ea78, "SEL.YY.i16", ADD_TWO_SRC
},
1249 { 0x0ec00, "F32_TO_F16", ADD_TWO_SRC
},
1250 { 0x0f640, "ICMP.GL.GT", ADD_TWO_SRC
}, // src0 > src1 ? 1 : 0
1251 { 0x0f648, "ICMP.GL.GE", ADD_TWO_SRC
},
1252 { 0x0f650, "UCMP.GL.GT", ADD_TWO_SRC
},
1253 { 0x0f658, "UCMP.GL.GE", ADD_TWO_SRC
},
1254 { 0x0f660, "ICMP.GL.EQ", ADD_TWO_SRC
},
1255 { 0x0f6c0, "ICMP.D3D.GT", ADD_TWO_SRC
}, // src0 > src1 ? ~0 : 0
1256 { 0x0f6c8, "ICMP.D3D.GE", ADD_TWO_SRC
},
1257 { 0x0f6d0, "UCMP.D3D.GT", ADD_TWO_SRC
},
1258 { 0x0f6d8, "UCMP.D3D.GE", ADD_TWO_SRC
},
1259 { 0x0f6e0, "ICMP.D3D.EQ", ADD_TWO_SRC
},
1260 { 0x10000, "MAX.v2f16", ADD_FMINMAX16
},
1261 { 0x11000, "ADD_MSCALE.f32", ADD_FADDMscale
},
1262 { 0x12000, "MIN.v2f16", ADD_FMINMAX16
},
1263 { 0x14000, "ADD.v2f16", ADD_FADD16
},
1264 { 0x17000, "FCMP.D3D", ADD_FCMP16
},
1265 { 0x178c0, "ADD.i32", ADD_TWO_SRC
},
1266 { 0x17900, "ADD.v2i16", ADD_TWO_SRC
},
1267 { 0x17ac0, "SUB.i32", ADD_TWO_SRC
},
1268 { 0x17c10, "ADDC.i32", ADD_TWO_SRC
}, // adds src0 to the bottom bit of src1
1269 { 0x17d80, "ADD.i32.i16.X", ADD_TWO_SRC
},
1270 { 0x17d90, "ADD.i32.u16.X", ADD_TWO_SRC
},
1271 { 0x17dc0, "ADD.i32.i16.Y", ADD_TWO_SRC
},
1272 { 0x17dd0, "ADD.i32.u16.Y", ADD_TWO_SRC
},
1273 // Compute varying address and datatype (for storing in the vertex shader),
1274 // and store the vec3 result in the data register. The result is passed as
1275 // the 3 normal arguments to ST_VAR.
1276 { 0x18000, "LD_VAR_ADDR.f16", ADD_VARYING_ADDRESS
, true },
1277 { 0x18100, "LD_VAR_ADDR.f32", ADD_VARYING_ADDRESS
, true },
1278 { 0x18200, "LD_VAR_ADDR.i32", ADD_VARYING_ADDRESS
, true },
1279 { 0x18300, "LD_VAR_ADDR.u32", ADD_VARYING_ADDRESS
, true },
1280 // Implements alpha-to-coverage, as well as possibly the late depth and
1281 // stencil tests. The first source is the existing sample mask in R60
1282 // (possibly modified by gl_SampleMask), and the second source is the alpha
1283 // value. The sample mask is written right away based on the
1284 // alpha-to-coverage result using the normal register write mechanism,
1285 // since that doesn't need to read from any memory, and then written again
1286 // later based on the result of the stencil and depth tests using the
1287 // special register.
1288 { 0x191e8, "ATEST.f32", ADD_TWO_SRC
, true },
1289 { 0x191f0, "ATEST.X.f16", ADD_TWO_SRC
, true },
1290 { 0x191f8, "ATEST.Y.f16", ADD_TWO_SRC
, true },
1291 // store a varying given the address and datatype from LD_VAR_ADDR
1292 { 0x19300, "ST_VAR.v1", ADD_THREE_SRC
, true },
1293 { 0x19340, "ST_VAR.v2", ADD_THREE_SRC
, true },
1294 { 0x19380, "ST_VAR.v3", ADD_THREE_SRC
, true },
1295 { 0x193c0, "ST_VAR.v4", ADD_THREE_SRC
, true },
1296 // This takes the sample coverage mask (computed by ATEST above) as a
1297 // regular argument, in addition to the vec4 color in the special register.
1298 { 0x1952c, "BLEND", ADD_BLENDING
, true },
1299 { 0x1a000, "LD_VAR.16", ADD_VARYING_INTERP
, true },
1300 { 0x1ae60, "TEX", ADD_TEX
, true },
1301 { 0x1c000, "RSHIFT_NAND.i32", ADD_THREE_SRC
},
1302 { 0x1c300, "RSHIFT_OR.i32", ADD_THREE_SRC
},
1303 { 0x1c400, "RSHIFT_AND.i32", ADD_THREE_SRC
},
1304 { 0x1c700, "RSHIFT_NOR.i32", ADD_THREE_SRC
},
1305 { 0x1c800, "LSHIFT_NAND.i32", ADD_THREE_SRC
},
1306 { 0x1cb00, "LSHIFT_OR.i32", ADD_THREE_SRC
},
1307 { 0x1cc00, "LSHIFT_AND.i32", ADD_THREE_SRC
},
1308 { 0x1cf00, "LSHIFT_NOR.i32", ADD_THREE_SRC
},
1309 { 0x1d000, "RSHIFT_XOR.i32", ADD_THREE_SRC
},
1310 { 0x1d100, "RSHIFT_XNOR.i32", ADD_THREE_SRC
},
1311 { 0x1d200, "LSHIFT_XOR.i32", ADD_THREE_SRC
},
1312 { 0x1d300, "LSHIFT_XNOR.i32", ADD_THREE_SRC
},
1313 { 0x1d400, "LSHIFT_ADD.i32", ADD_THREE_SRC
},
1314 { 0x1d500, "LSHIFT_SUB.i32", ADD_THREE_SRC
},
1315 { 0x1d500, "LSHIFT_RSUB.i32", ADD_THREE_SRC
},
1316 { 0x1d700, "RSHIFT_ADD.i32", ADD_THREE_SRC
},
1317 { 0x1d800, "RSHIFT_SUB.i32", ADD_THREE_SRC
},
1318 { 0x1d900, "RSHIFT_RSUB.i32", ADD_THREE_SRC
},
1319 { 0x1da00, "ARSHIFT_ADD.i32", ADD_THREE_SRC
},
1320 { 0x1db00, "ARSHIFT_SUB.i32", ADD_THREE_SRC
},
1321 { 0x1dc00, "ARSHIFT_RSUB.i32", ADD_THREE_SRC
},
1322 { 0x1dd18, "OR.i32", ADD_TWO_SRC
},
1323 { 0x1dd20, "AND.i32", ADD_TWO_SRC
},
1324 { 0x1dd60, "LSHIFT.i32", ADD_TWO_SRC
},
1325 { 0x1dd50, "XOR.i32", ADD_TWO_SRC
},
1326 { 0x1dd80, "RSHIFT.i32", ADD_TWO_SRC
},
1327 { 0x1dda0, "ARSHIFT.i32", ADD_TWO_SRC
},
1330 static struct add_op_info
find_add_op_info(unsigned op
)
1332 for (unsigned i
= 0; i
< ARRAY_SIZE(add_op_infos
); i
++) {
1333 unsigned opCmp
= ~0;
1334 switch (add_op_infos
[i
].src_type
) {
1351 opCmp
= op
& ~0x1fff;
1354 case ADD_FADDMscale
:
1355 opCmp
= op
& ~0xfff;
1359 opCmp
= op
& ~0x7ff;
1361 case ADD_TEX_COMPACT
:
1362 opCmp
= op
& ~0x3ff;
1364 case ADD_VARYING_INTERP
:
1365 opCmp
= op
& ~0x7ff;
1367 case ADD_VARYING_ADDRESS
:
1374 opCmp
= op
& ~0xfff;
1380 if (add_op_infos
[i
].op
== opCmp
)
1381 return add_op_infos
[i
];
1384 struct add_op_info info
;
1385 snprintf(info
.name
, sizeof(info
.name
), "op%04x", op
);
1387 info
.src_type
= ADD_TWO_SRC
;
1388 info
.has_data_reg
= true;
1392 static void dump_add(uint64_t word
, struct bifrost_regs regs
, struct bifrost_regs next_regs
, uint64_t *consts
,
1393 unsigned data_reg
, unsigned offset
, bool verbose
)
1396 printf("# ADD: %016" PRIx64
"\n", word
);
1398 struct bifrost_add_inst ADD
;
1399 memcpy((char *) &ADD
, (char *) &word
, sizeof(ADD
));
1400 struct add_op_info info
= find_add_op_info(ADD
.op
);
1402 printf("%s", info
.name
);
1404 // float16 seems like it doesn't support output modifiers
1405 if (info
.src_type
== ADD_FADD
|| info
.src_type
== ADD_FMINMAX
) {
1407 dump_output_mod(bits(ADD
.op
, 8, 10));
1408 if (info
.src_type
== ADD_FADD
)
1409 dump_round_mode(bits(ADD
.op
, 10, 12));
1411 dump_minmax_mode(bits(ADD
.op
, 10, 12));
1412 } else if (info
.src_type
== ADD_FCMP
|| info
.src_type
== ADD_FCMP16
) {
1413 dump_fcmp(bits(ADD
.op
, 3, 6));
1414 if (info
.src_type
== ADD_FCMP
)
1418 } else if (info
.src_type
== ADD_FADDMscale
) {
1419 switch ((ADD
.op
>> 6) & 0x7) {
1421 // causes GPU hangs on G71
1422 case 1: printf(".invalid"); break;
1423 // Same as usual outmod value.
1424 case 2: printf(".clamp_0_1"); break;
1425 // If src0 is infinite or NaN, flush it to zero so that the other
1426 // source is passed through unmodified.
1427 case 3: printf(".flush_src0_inf_nan"); break;
1429 case 4: printf(".flush_src1_inf_nan"); break;
1430 // Every other case seems to behave the same as the above?
1431 default: printf(".unk%d", (ADD
.op
>> 6) & 0x7); break;
1433 } else if (info
.src_type
== ADD_VARYING_INTERP
) {
1438 switch ((ADD
.op
>> 7) & 0x3) {
1439 case 0: printf(".per_frag"); break;
1440 case 1: printf(".centroid"); break;
1442 case 3: printf(".explicit"); break;
1444 printf(".v%d", ((ADD
.op
>> 5) & 0x3) + 1);
1445 } else if (info
.src_type
== ADD_BRANCH
) {
1446 enum branch_code branchCode
= (enum branch_code
) ((ADD
.op
>> 6) & 0x3f);
1447 if (branchCode
== BR_ALWAYS
) {
1448 // unconditional branch
1450 enum branch_cond cond
= (enum branch_cond
) ((ADD
.op
>> 6) & 0x7);
1451 enum branch_bit_size size
= (enum branch_bit_size
) ((ADD
.op
>> 9) & 0x7);
1452 bool portSwapped
= (ADD
.op
& 0x7) < ADD
.src0
;
1453 // See the comment in branch_bit_size
1454 if (size
== BR_SIZE_16YX0
)
1456 if (size
== BR_SIZE_16YX1
)
1457 portSwapped
= false;
1458 // These sizes are only for floating point comparisons, so the
1459 // non-floating-point comparisons are reused to encode the flipped
1461 if (size
== BR_SIZE_32_AND_16X
|| size
== BR_SIZE_32_AND_16Y
)
1462 portSwapped
= false;
1463 // There's only one argument, so we reuse the extra argument to
1465 if (size
== BR_SIZE_ZERO
)
1466 portSwapped
= !(ADD
.op
& 1);
1476 if (size
== BR_SIZE_32_AND_16X
|| size
== BR_SIZE_32_AND_16Y
) {
1511 printf(".OGT.unk.f");
1517 printf(".OLT.unk.f");
1524 case BR_SIZE_32_AND_16X
:
1525 case BR_SIZE_32_AND_16Y
:
1534 case BR_SIZE_ZERO
: {
1535 unsigned ctrl
= (ADD
.op
>> 1) & 0x3;
1547 struct bifrost_reg_ctrl next_ctrl
= DecodeRegCtrl(next_regs
);
1548 if (next_ctrl
.add_write_unit
!= REG_WRITE_NONE
) {
1549 printf("{R%d, T1}, ", GetRegToWrite(next_ctrl
.add_write_unit
, next_regs
));
1554 switch (info
.src_type
) {
1556 // Note: in this case, regs.uniform_const == location | 0x8
1557 // This probably means we can't load uniforms or immediates in the
1558 // same instruction. This re-uses the encoding that normally means
1559 // "disabled", where the low 4 bits are ignored. Perhaps the extra
1560 // 0x8 or'd in indicates this is happening.
1561 printf("location:%d, ", regs
.uniform_const
& 0x7);
1564 dump_src(ADD
.src0
, regs
, consts
, false);
1567 case ADD_TEX_COMPACT
: {
1570 bool dualTex
= false;
1571 if (info
.src_type
== ADD_TEX_COMPACT
) {
1572 tex_index
= (ADD
.op
>> 3) & 0x7;
1573 sampler_index
= (ADD
.op
>> 7) & 0x7;
1574 bool unknown
= (ADD
.op
& 0x40);
1575 // TODO: figure out if the unknown bit is ever 0
1579 uint64_t constVal
= get_const(consts
, regs
);
1580 uint32_t controlBits
= (ADD
.op
& 0x8) ? (constVal
>> 32) : constVal
;
1581 struct bifrost_tex_ctrl ctrl
;
1582 memcpy((char *) &ctrl
, (char *) &controlBits
, sizeof(ctrl
));
1584 // TODO: figure out what actually triggers dual-tex
1585 if (ctrl
.result_type
== 9) {
1586 struct bifrost_dual_tex_ctrl dualCtrl
;
1587 memcpy((char *) &dualCtrl
, (char *) &controlBits
, sizeof(ctrl
));
1588 printf("(dualtex) tex0:%d samp0:%d tex1:%d samp1:%d ",
1589 dualCtrl
.tex_index0
, dualCtrl
.sampler_index0
,
1590 dualCtrl
.tex_index1
, dualCtrl
.sampler_index1
);
1591 if (dualCtrl
.unk0
!= 3)
1592 printf("unk:%d ", dualCtrl
.unk0
);
1595 if (ctrl
.no_merge_index
) {
1596 tex_index
= ctrl
.tex_index
;
1597 sampler_index
= ctrl
.sampler_index
;
1599 tex_index
= sampler_index
= ctrl
.tex_index
;
1600 unsigned unk
= ctrl
.sampler_index
>> 2;
1602 printf("unk:%d ", unk
);
1603 if (ctrl
.sampler_index
& 1)
1605 if (ctrl
.sampler_index
& 2)
1610 printf("unk0:%d ", ctrl
.unk0
);
1613 if (ctrl
.unk2
!= 0xf)
1614 printf("unk2:%x ", ctrl
.unk2
);
1616 switch (ctrl
.result_type
) {
1618 printf("f32 "); break;
1620 printf("i32 "); break;
1622 printf("u32 "); break;
1624 printf("unktype(%x) ", ctrl
.result_type
);
1627 switch (ctrl
.tex_type
) {
1629 printf("cube "); break;
1631 printf("buffer "); break;
1633 printf("2D "); break;
1635 printf("3D "); break;
1644 if (ctrl
.calc_gradients
) {
1645 int comp
= (controlBits
>> 20) & 0x3;
1646 printf("txg comp:%d ", comp
);
1651 if (!ctrl
.not_supply_lod
) {
1652 if (ctrl
.compute_lod
)
1653 printf("lod_bias ");
1658 if (!ctrl
.calc_gradients
)
1662 if (ctrl
.texel_offset
)
1668 if (tex_index
== -1)
1669 printf("tex:indirect ");
1671 printf("tex:%d ", tex_index
);
1673 if (sampler_index
== -1)
1674 printf("samp:indirect ");
1676 printf("samp:%d ", sampler_index
);
1680 case ADD_VARYING_INTERP
: {
1681 unsigned addr
= ADD
.op
& 0x1f;
1682 if (addr
< 0b10100) {
1685 } else if (addr
< 0b11000) {
1688 else if (addr
== 23)
1691 printf("unk%d", addr
);
1693 dump_src(ADD
.op
& 0x7, regs
, consts
, false);
1696 dump_src(ADD
.src0
, regs
, consts
, false);
1699 case ADD_VARYING_ADDRESS
: {
1700 dump_src(ADD
.src0
, regs
, consts
, false);
1702 dump_src(ADD
.op
& 0x7, regs
, consts
, false);
1704 unsigned location
= (ADD
.op
>> 3) & 0x1f;
1705 if (location
< 16) {
1706 printf("location:%d", location
);
1707 } else if (location
== 20) {
1708 printf("location:%u", (uint32_t) get_const(consts
, regs
));
1709 } else if (location
== 21) {
1710 printf("location:%u", (uint32_t) (get_const(consts
, regs
) >> 32));
1712 printf("location:%d(unk)", location
);
1717 printf("location:%d, ", (ADD
.op
>> 3) & 0xf);
1719 dump_src(ADD
.src0
, regs
, consts
, false);
1721 dump_src(ADD
.op
& 0x7, regs
, consts
, false);
1724 dump_src(ADD
.src0
, regs
, consts
, false);
1726 dump_src(ADD
.op
& 0x7, regs
, consts
, false);
1728 dump_src((ADD
.op
>> 3) & 0x7, regs
, consts
, false);
1734 if (ADD
.op
& 0x1000)
1736 dump_src(ADD
.src0
, regs
, consts
, false);
1737 switch ((ADD
.op
>> 6) & 0x3) {
1744 if (ADD
.op
& 0x1000)
1751 dump_src(ADD
.op
& 0x7, regs
, consts
, false);
1752 switch ((ADD
.op
>> 6) & 0x3) {
1772 if (ADD
.op
& 0x1000)
1774 dump_src(ADD
.src0
, regs
, consts
, false);
1775 if (ADD
.op
& 0x1000)
1777 dump_16swizzle((ADD
.op
>> 6) & 0x3);
1783 dump_src(ADD
.op
& 0x7, regs
, consts
, false);
1784 dump_16swizzle((ADD
.op
>> 8) & 0x3);
1788 case ADD_FMINMAX16
: {
1789 bool abs1
= ADD
.op
& 0x8;
1790 bool abs2
= (ADD
.op
& 0x7) < ADD
.src0
;
1795 dump_src(ADD
.src0
, regs
, consts
, false);
1796 dump_16swizzle((ADD
.op
>> 6) & 0x3);
1804 dump_src(ADD
.op
& 0x7, regs
, consts
, false);
1805 dump_16swizzle((ADD
.op
>> 8) & 0x3);
1810 case ADD_FADDMscale
: {
1815 dump_src(ADD
.src0
, regs
, consts
, false);
1823 dump_src(ADD
.op
& 0x7, regs
, consts
, false);
1827 dump_src((ADD
.op
>> 3) & 0x7, regs
, consts
, false);
1831 if (ADD
.op
& 0x400) {
1834 if (ADD
.op
& 0x100) {
1837 dump_src(ADD
.src0
, regs
, consts
, false);
1838 switch ((ADD
.op
>> 6) & 0x3) {
1845 if (ADD
.op
& 0x100) {
1849 if (ADD
.op
& 0x200) {
1852 dump_src(ADD
.op
& 0x7, regs
, consts
, false);
1853 switch ((ADD
.op
>> 6) & 0x3) {
1867 if (ADD
.op
& 0x200) {
1872 dump_src(ADD
.src0
, regs
, consts
, false);
1873 dump_16swizzle((ADD
.op
>> 6) & 0x3);
1875 dump_src(ADD
.op
& 0x7, regs
, consts
, false);
1876 dump_16swizzle((ADD
.op
>> 8) & 0x3);
1879 enum branch_code code
= (enum branch_code
) ((ADD
.op
>> 6) & 0x3f);
1880 enum branch_bit_size size
= (enum branch_bit_size
) ((ADD
.op
>> 9) & 0x7);
1881 if (code
!= BR_ALWAYS
) {
1882 dump_src(ADD
.src0
, regs
, consts
, false);
1892 case BR_SIZE_ZERO
: {
1893 unsigned ctrl
= (ADD
.op
>> 1) & 0x3;
1910 if (code
!= BR_ALWAYS
&& size
!= BR_SIZE_ZERO
) {
1911 dump_src(ADD
.op
& 0x7, regs
, consts
, false);
1916 case BR_SIZE_32_AND_16X
:
1920 case BR_SIZE_32_AND_16Y
:
1928 // I haven't had the chance to test if this actually specifies the
1929 // branch offset, since I couldn't get it to produce values other
1930 // than 5 (uniform/const high), but these three bits are always
1931 // consistent across branch instructions, so it makes sense...
1932 int offsetSrc
= (ADD
.op
>> 3) & 0x7;
1933 if (offsetSrc
== 4 || offsetSrc
== 5) {
1934 // If the offset is known/constant, we can decode it
1935 uint32_t raw_offset
;
1937 raw_offset
= get_const(consts
, regs
);
1939 raw_offset
= get_const(consts
, regs
) >> 32;
1940 // The high 4 bits are flags, while the rest is the
1941 // twos-complement offset in bytes (here we convert to
1943 int32_t branch_offset
= ((int32_t) raw_offset
<< 4) >> 8;
1945 // If high4 is the high 4 bits of the last 64-bit constant,
1946 // this is calculated as (high4 + 4) & 0xf, or 0 if the branch
1947 // offset itself is the last constant. Not sure if this is
1948 // actually used, or just garbage in unused bits, but in any
1949 // case, we can just ignore it here since it's redundant. Note
1950 // that if there is any padding, this will be 4 since the
1951 // padding counts as the last constant.
1952 unsigned flags
= raw_offset
>> 28;
1955 // Note: the offset is in bytes, relative to the beginning of the
1956 // current clause, so a zero offset would be a loop back to the
1957 // same clause (annoyingly different from Midgard).
1958 printf("clause_%d", offset
+ branch_offset
);
1960 dump_src(offsetSrc
, regs
, consts
, false);
1964 if (info
.has_data_reg
) {
1965 printf(", R%d", data_reg
);
1970 void dump_instr(const struct bifrost_alu_inst
*instr
, struct bifrost_regs next_regs
, uint64_t *consts
,
1971 unsigned data_reg
, unsigned offset
, bool verbose
)
1973 struct bifrost_regs regs
;
1974 memcpy((char *) ®s
, (char *) &instr
->reg_bits
, sizeof(regs
));
1977 printf("# regs: %016" PRIx64
"\n", instr
->reg_bits
);
1980 dump_fma(instr
->fma_bits
, regs
, next_regs
, consts
, verbose
);
1981 dump_add(instr
->add_bits
, regs
, next_regs
, consts
, data_reg
, offset
, verbose
);
1984 bool dump_clause(uint32_t *words
, unsigned *size
, unsigned offset
, bool verbose
) {
1985 // State for a decoded clause
1986 struct bifrost_alu_inst instrs
[8] = {};
1987 uint64_t consts
[6] = {};
1988 unsigned num_instrs
= 0;
1989 unsigned num_consts
= 0;
1990 uint64_t header_bits
= 0;
1991 bool stopbit
= false;
1994 for (i
= 0; ; i
++, words
+= 4) {
1997 for (int j
= 0; j
< 4; j
++)
1998 printf("%08x ", words
[3 - j
]); // low bit on the right
2001 unsigned tag
= bits(words
[0], 0, 8);
2003 // speculatively decode some things that are common between many formats, so we can share some code
2004 struct bifrost_alu_inst main_instr
= {};
2006 main_instr
.add_bits
= bits(words
[2], 2, 32 - 13);
2008 main_instr
.fma_bits
= bits(words
[1], 11, 32) | bits(words
[2], 0, 2) << (32 - 11);
2010 main_instr
.reg_bits
= ((uint64_t) bits(words
[1], 0, 11)) << 24 | (uint64_t) bits(words
[0], 8, 32);
2012 uint64_t const0
= bits(words
[0], 8, 32) << 4 | (uint64_t) words
[1] << 28 | bits(words
[2], 0, 4) << 60;
2013 uint64_t const1
= bits(words
[2], 4, 32) << 4 | (uint64_t) words
[3] << 32;
2015 bool stop
= tag
& 0x40;
2018 printf("# tag: 0x%02x\n", tag
);
2021 unsigned idx
= stop
? 5 : 2;
2022 main_instr
.add_bits
|= ((tag
>> 3) & 0x7) << 17;
2023 instrs
[idx
+ 1] = main_instr
;
2024 instrs
[idx
].add_bits
= bits(words
[3], 0, 17) | ((tag
& 0x7) << 17);
2025 instrs
[idx
].fma_bits
|= bits(words
[2], 19, 32) << 10;
2026 consts
[0] = bits(words
[3], 17, 32) << 4;
2029 switch ((tag
>> 3) & 0x7) {
2031 switch (tag
& 0x7) {
2033 main_instr
.add_bits
|= bits(words
[3], 29, 32) << 17;
2034 instrs
[1] = main_instr
;
2039 instrs
[2].add_bits
= bits(words
[3], 0, 17) | bits(words
[3], 29, 32) << 17;
2040 instrs
[2].fma_bits
|= bits(words
[2], 19, 32) << 10;
2048 instrs
[2].add_bits
= bits(words
[3], 0, 17) | bits(words
[3], 29, 32) << 17;
2049 instrs
[2].fma_bits
|= bits(words
[2], 19, 32) << 10;
2050 main_instr
.add_bits
|= bits(words
[3], 26, 29) << 17;
2051 instrs
[3] = main_instr
;
2052 if ((tag
& 0x7) == 0x5) {
2058 instrs
[5].add_bits
= bits(words
[3], 0, 17) | bits(words
[3], 29, 32) << 17;
2059 instrs
[5].fma_bits
|= bits(words
[2], 19, 32) << 10;
2066 instrs
[5].add_bits
= bits(words
[3], 0, 17) | bits(words
[3], 29, 32) << 17;
2067 instrs
[5].fma_bits
|= bits(words
[2], 19, 32) << 10;
2068 main_instr
.add_bits
|= bits(words
[3], 26, 29) << 17;
2069 instrs
[6] = main_instr
;
2074 printf("unknown tag bits 0x%02x\n", tag
);
2079 unsigned idx
= ((tag
>> 3) & 0x7) == 2 ? 4 : 7;
2080 main_instr
.add_bits
|= (tag
& 0x7) << 17;
2081 instrs
[idx
] = main_instr
;
2082 consts
[0] |= (bits(words
[2], 19, 32) | ((uint64_t) words
[3] << 13)) << 19;
2084 num_instrs
= idx
+ 1;
2089 unsigned idx
= stop
? 4 : 1;
2090 main_instr
.add_bits
|= (tag
& 0x7) << 17;
2091 instrs
[idx
] = main_instr
;
2092 instrs
[idx
+ 1].fma_bits
|= bits(words
[3], 22, 32);
2093 instrs
[idx
+ 1].reg_bits
= bits(words
[2], 19, 32) | (bits(words
[3], 0, 22) << (32 - 19));
2097 // only constants can come after this
2101 header_bits
= bits(words
[2], 19, 32) | ((uint64_t) words
[3] << (32 - 19));
2102 main_instr
.add_bits
|= (tag
& 0x7) << 17;
2103 instrs
[0] = main_instr
;
2107 unsigned pos
= tag
& 0xf;
2108 // note that `pos' encodes both the total number of
2109 // instructions and the position in the constant stream,
2110 // presumably because decoded constants and instructions
2111 // share a buffer in the decoder, but we only care about
2112 // the position in the constant stream; the total number of
2113 // instructions is redundant.
2114 unsigned const_idx
= 7;
2141 printf("# unknown pos 0x%x\n", pos
);
2143 if (num_consts
< const_idx
+ 2)
2144 num_consts
= const_idx
+ 2;
2145 consts
[const_idx
] = const0
;
2146 consts
[const_idx
+ 1] = const1
;
2162 printf("# header: %012" PRIx64
"\n", header_bits
);
2165 struct bifrost_header header
;
2166 memcpy((char *) &header
, (char *) &header_bits
, sizeof(struct bifrost_header
));
2167 dump_header(header
, verbose
);
2168 if (!header
.no_end_of_shader
)
2172 for (i
= 0; i
< num_instrs
; i
++) {
2173 struct bifrost_regs next_regs
;
2174 if (i
+ 1 == num_instrs
) {
2175 memcpy((char *) &next_regs
, (char *) &instrs
[0].reg_bits
,
2178 memcpy((char *) &next_regs
, (char *) &instrs
[i
+ 1].reg_bits
,
2182 dump_instr(&instrs
[i
], next_regs
, consts
, header
.datareg
, offset
, verbose
);
2187 for (unsigned i
= 0; i
< num_consts
; i
++) {
2188 printf("# const%d: %08" PRIx64
"\n", 2 * i
, consts
[i
] & 0xffffffff);
2189 printf("# const%d: %08" PRIx64
"\n", 2 * i
+ 1, consts
[i
] >> 32);
2195 void disassemble_bifrost(uint8_t *code
, size_t size
, bool verbose
)
2197 uint32_t *words
= (uint32_t *) code
;
2198 uint32_t *words_end
= words
+ (size
/ 4);
2199 // used for displaying branch targets
2200 unsigned offset
= 0;
2201 while (words
!= words_end
)
2203 // we don't know what the program-end bit is quite yet, so for now just
2204 // assume that an all-0 quadword is padding
2205 uint32_t zero
[4] = {};
2206 if (memcmp(words
, zero
, 4 * sizeof(uint32_t)) == 0)
2208 printf("clause_%d:\n", offset
);
2210 if (dump_clause(words
, &size
, offset
, verbose
) == true) {