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
32 enum bifrost_clause_type
{
33 BIFROST_CLAUSE_NONE
= 0,
34 BIFROST_CLAUSE_LOAD_VARY
= 1,
35 BIFROST_CLAUSE_UBO
= 2,
36 BIFROST_CLAUSE_TEX
= 3,
37 BIFROST_CLAUSE_SSBO_LOAD
= 5,
38 BIFROST_CLAUSE_SSBO_STORE
= 6,
39 BIFROST_CLAUSE_BLEND
= 9,
40 BIFROST_CLAUSE_ATEST
= 13,
41 BIFROST_CLAUSE_64BIT
= 15
44 struct bifrost_header
{
46 // If true, convert any infinite result of any floating-point operation to
47 // the biggest representable number.
48 unsigned suppress_inf
: 1;
49 // Convert any NaN results to 0.
50 unsigned suppress_nan
: 1;
52 // true if the execution mask of the next clause is the same as the mask of
53 // the current clause.
54 unsigned back_to_back
: 1;
55 unsigned no_end_of_shader
: 1;
57 // Set to true for fragment shaders, to implement this bit of spec text
58 // from section 7.1.5 of the GLSL ES spec:
60 // "Stores to image and buffer variables performed by helper invocations
61 // have no effect on the underlying image or buffer memory."
63 // Helper invocations are threads (invocations) corresponding to pixels in
64 // a quad that aren't actually part of the triangle, but are included to
65 // make derivatives work correctly. They're usually turned on, but they
66 // need to be masked off for GLSL-level stores. This bit seems to be the
67 // only bit that's actually different between fragment shaders and other
68 // shaders, so this is probably what it's doing.
69 unsigned elide_writes
: 1;
70 // If backToBack is off:
71 // - true for conditional branches and fallthrough
72 // - false for unconditional branches
73 // The blob seems to always set it to true if back-to-back is on.
74 unsigned branch_cond
: 1;
75 // This bit is set when the next clause writes to the data register of some
77 unsigned datareg_writebarrier
: 1;
79 unsigned scoreboard_deps
: 8;
80 unsigned scoreboard_index
: 3;
81 enum bifrost_clause_type clause_type
: 4;
82 unsigned unk3
: 1; // part of clauseType?
83 enum bifrost_clause_type next_clause_type
: 4;
84 unsigned unk4
: 1; // part of nextClauseType?
85 } __attribute__((packed
));
87 enum bifrost_packed_src
{
88 BIFROST_SRC_PORT0
= 0,
89 BIFROST_SRC_PORT1
= 1,
90 BIFROST_SRC_PORT3
= 2,
91 BIFROST_SRC_STAGE
= 3,
92 BIFROST_SRC_CONST_LO
= 4,
93 BIFROST_SRC_CONST_HI
= 5,
94 BIFROST_SRC_PASS_FMA
= 6,
95 BIFROST_SRC_PASS_ADD
= 7,
98 #define BIFROST_FMA_EXT (0xe0000)
99 #define BIFROST_FMA_OP_MOV BIFROST_FMA_EXT | (0x32d)
100 #define BIFROST_FMA_OP_FREXPE_LOG BIFROST_FMA_EXT | 0x3c5
101 #define BIFROST_FMA_OP_ADD_FREXPM ((BIFROST_FMA_EXT | 0x1e80) >> 3)
102 #define BIFROST_FMA_SEL_16(swiz) (((BIFROST_FMA_EXT | 0x1e00) >> 3) | (swiz))
104 #define BIFROST_FMA_ROUND_16(mode, swiz) (BIFROST_FMA_EXT | 0x1800 | (swiz) | ((mode) << 6))
105 #define BIFROST_FMA_ROUND_32(mode) (BIFROST_FMA_EXT | 0x1805 | ((mode) << 6))
107 struct bifrost_fma_inst
{
110 } __attribute__((packed
));
112 struct bifrost_fma_2src
{
116 } __attribute__((packed
));
118 #define BIFROST_FMA_OP_SEL8 (0x71)
120 struct bifrost_fma_sel8
{
125 unsigned swizzle
: 4;
127 } __attribute__((packed
));
129 #define BIFROST_FMA_OP_MSCALE (0x50 >> 3)
131 struct bifrost_fma_mscale
{
137 /* If mscale_mode is set - an MSCALE specific mode. If it is not set, a
140 unsigned mscale_mode
: 1;
142 unsigned src0_abs
: 1;
143 unsigned src1_neg
: 1;
144 unsigned src2_neg
: 1;
146 } __attribute__((packed
));
148 #define BIFROST_ADD_OP_BLEND (0x1952c)
149 #define BIFROST_ADD_OP_FRCP_FAST_F32 (0x0cc00)
150 #define BIFROST_ADD_OP_FRCP_FAST_F16_X (0x0ce10)
151 #define BIFROST_ADD_OP_FRCP_FAST_F16_Y (0x0ce30)
152 #define BIFROST_ADD_OP_FRSQ_FAST_F32 (0x0cc20)
153 #define BIFROST_ADD_OP_FRSQ_FAST_F16_X (0x0ce50)
154 #define BIFROST_ADD_OP_FRSQ_FAST_F16_Y (0x0ce70)
155 #define BIFROST_ADD_OP_LOG2_HELP (0x0cc68)
156 #define BIFROST_ADD_OP_FEXP2_FAST (0x0cd58)
158 struct bifrost_add_inst
{
161 } __attribute__((packed
));
163 #define BIFROST_ADD_OP_LD_UBO_1 (0x0c1a0 >> 3)
164 #define BIFROST_ADD_OP_LD_UBO_2 (0x0c1e0 >> 3)
165 #define BIFROST_ADD_OP_LD_UBO_3 (0x0caa0 >> 3)
166 #define BIFROST_ADD_OP_LD_UBO_4 (0x0c220 >> 3)
167 #define BIFROST_ADD_SEL_16(swiz) ((0xea60 >> 3) | (swiz))
169 struct bifrost_add_2src
{
173 } __attribute__((packed
));
175 #define BIFROST_ADD_OP_FMAX32 (0x00)
176 #define BIFROST_ADD_OP_FMIN32 (0x01)
177 #define BIFROST_ADD_OP_FADD32 (0x02)
179 #define BIFROST_ADD_OP_FADD16 (0x0A)
181 struct bifrost_add_faddmin
{
184 unsigned src1_abs
: 1;
185 unsigned src0_neg
: 1;
186 unsigned src1_neg
: 1;
187 unsigned select
: 2; /* swizzle_0 for fp16 */
188 unsigned outmod
: 2; /* swizzle_1 for fp16 */
190 unsigned src0_abs
: 1;
192 } __attribute__((packed
));
194 #define BIFROST_ADD_OP_FMAX16 (0x10)
195 #define BIFROST_ADD_OP_FMIN16 (0x12)
197 struct bifrost_add_fmin16
{
200 /* abs2 inferred as with FMA */
202 unsigned src0_neg
: 1;
203 unsigned src1_neg
: 1;
204 unsigned src0_swizzle
: 2;
205 unsigned src1_swizzle
: 2;
208 } __attribute__((packed
));
210 #define BIFROST_ADD_OP_ST_VAR (0x19300 >> 8)
212 struct bifrost_st_vary
{
216 unsigned channels
: 2;
218 } __attribute__((packed
));
220 #define BIFROST_ADD_OP_ATEST (0xc8f)
222 struct bifrost_add_atest
{
223 /* gl_SampleMask (R60) */
229 /* If half, X/Y select. If !half, always set */
230 unsigned component
: 1;
234 } __attribute__((packed
));
236 enum bifrost_outmod
{
239 BIFROST_SAT_SIGNED
= 0x2,
243 enum bifrost_roundmode
{
244 BIFROST_RTE
= 0x0, /* round to even */
245 BIFROST_RTP
= 0x1, /* round to positive */
246 BIFROST_RTN
= 0x2, /* round to negative */
247 BIFROST_RTZ
= 0x3 /* round to zero */
250 /* NONE: Same as fmax() and fmin() -- return the other
251 * number if any number is NaN. Also always return +0 if
252 * one argument is +0 and the other is -0.
254 * NAN_WINS: Instead of never returning a NaN, always return
255 * one. The "greater"/"lesser" NaN is always returned, first
256 * by checking the sign and then the mantissa bits.
258 * SRC1_WINS: For max, implement src0 > src1 ? src0 : src1.
259 * For min, implement src0 < src1 ? src0 : src1. This
260 * includes handling NaN's and signedness of 0 differently
261 * from above, since +0 and -0 compare equal and comparisons
262 * always return false for NaN's. As a result, this mode is
265 * SRC0_WINS: For max, implement src0 < src1 ? src1 : src0
266 * For min, implement src0 > src1 ? src1 : src0
270 enum bifrost_minmax_mode
{
271 BIFROST_MINMAX_NONE
= 0x0,
272 BIFROST_NAN_WINS
= 0x1,
273 BIFROST_SRC1_WINS
= 0x2,
274 BIFROST_SRC0_WINS
= 0x3,
277 #define BIFROST_FMA_OP_FADD32 (0x58 >> 2)
278 #define BIFROST_FMA_OP_FMAX32 (0x40 >> 2)
279 #define BIFROST_FMA_OP_FMIN32 (0x44 >> 2)
281 struct bifrost_fma_add
{
284 unsigned src1_abs
: 1;
285 unsigned src0_neg
: 1;
286 unsigned src1_neg
: 1;
288 unsigned src0_abs
: 1;
289 enum bifrost_roundmode roundmode
: 2;
290 enum bifrost_outmod outmod
: 2;
292 } __attribute__((packed
));
294 #define BIFROST_FMA_OP_FMAX16 (0xC0 >> 2)
295 #define BIFROST_FMA_OP_FMIN16 (0xCC >> 2)
296 #define BIFROST_FMA_OP_FADD16 (0xD8 >> 2)
298 struct bifrost_fma_add_minmax16
{
301 /* abs2 inferred as (src1 < src0) */
303 unsigned src0_neg
: 1;
304 unsigned src1_neg
: 1;
305 unsigned src0_swizzle
: 2;
306 unsigned src1_swizzle
: 2;
308 enum bifrost_outmod outmod
: 2;
309 /* roundmode for add, min/max mode for min/max */
311 } __attribute__((packed
));
313 #define BIFROST_FMA_OP_FMA (0x00)
315 struct bifrost_fma_fma
{
319 unsigned src_expand
: 3;
320 unsigned src0_abs
: 1;
321 enum bifrost_roundmode roundmode
: 2;
322 enum bifrost_outmod outmod
: 2;
323 unsigned src0_neg
: 1; /* 14 */
324 unsigned src2_neg
: 1;
325 unsigned src1_abs
: 1;
326 unsigned src2_abs
: 1; /* 17 */
328 } __attribute__((packed
));
330 #define BIFROST_FMA_OP_FMA16 (0x2)
332 struct bifrost_fma_fma16
{
336 unsigned swizzle_0
: 2;
337 unsigned swizzle_1
: 2;
338 enum bifrost_roundmode roundmode
: 2;
339 enum bifrost_outmod outmod
: 2;
340 unsigned src0_neg
: 1;
341 unsigned src2_neg
: 1;
342 unsigned swizzle_2
: 2;
344 } __attribute__((packed
));
346 enum bifrost_csel_cond
{
357 #define BIFROST_FMA_OP_CSEL4 (0x5c)
358 #define BIFROST_FMA_OP_CSEL4_V16 (0xdc)
360 struct bifrost_csel4
{
365 enum bifrost_csel_cond cond
: 3;
367 } __attribute__((packed
));
369 #define BIFROST_FMA_OP_RSHIFT_NAND (0x60000 >> 12)
370 #define BIFROST_FMA_OP_RSHIFT_AND (0x61000 >> 12)
371 #define BIFROST_FMA_OP_LSHIFT_NAND (0x62000 >> 12)
372 #define BIFROST_FMA_OP_LSHIFT_AND (0x63000 >> 12)
373 #define BIFROST_FMA_OP_RSHIFT_XOR (0x64000 >> 12)
374 #define BIFROST_FMA_OP_LSHIFT_ADD_32 (0x65200 >> 6)
375 #define BIFROST_FMA_OP_LSHIFT_SUB_32 (0x65600 >> 6)
376 #define BIFROST_FMA_OP_LSHIFT_RSUB_32 (0x65a00 >> 6)
377 #define BIFROST_FMA_OP_RSHIFT_ADD_32 (0x65e00 >> 6)
378 #define BIFROST_FMA_OP_RSHIFT_SUB_32 (0x66200 >> 6)
379 #define BIFROST_FMA_OP_RSHIFT_RSUB_32 (0x66600 >> 6)
381 struct bifrost_shift_fma
{
386 unsigned unk
: 1; /* always set? */
387 unsigned invert_1
: 1; /* Inverts sources to combining op */
388 /* For XOR, switches RSHIFT to LSHIFT since only one invert needed */
389 unsigned invert_2
: 1;
391 } __attribute__((packed
));
393 struct bifrost_shift_add
{
399 unsigned invert_1
: 1;
400 unsigned invert_2
: 1;
403 } __attribute__((packed
));
405 enum bifrost_fcmp_cond
{
414 #define BIFROST_FMA_OP_FCMP_GL (0x48000 >> 13)
415 #define BIFROST_FMA_OP_FCMP_D3D (0x4c000 >> 13)
417 struct bifrost_fma_fcmp
{
420 unsigned src1_abs
: 1;
422 unsigned src1_neg
: 1;
423 unsigned src_expand
: 3;
424 unsigned src0_abs
: 1;
425 enum bifrost_fcmp_cond cond
: 3;
427 } __attribute__((packed
));
429 struct bifrost_add_fcmp
{
432 enum bifrost_fcmp_cond cond
: 3;
433 unsigned src_expand
: 2;
434 unsigned src0_abs
: 1;
435 unsigned src1_abs
: 1;
436 unsigned src1_neg
: 1;
438 } __attribute__((packed
));
440 #define BIFROST_FMA_OP_FCMP_GL_16 (0xc8000 >> 13)
441 #define BIFROST_FMA_OP_FCMP_D3D_16 (0xcc000 >> 13)
443 struct bifrost_fma_fcmp16
{
451 unsigned src0_swizzle
: 2;
452 unsigned src1_swizzle
: 2;
454 enum bifrost_fcmp_cond cond
: 3;
456 } __attribute__((packed
));
458 struct bifrost_add_fcmp16
{
461 enum bifrost_fcmp_cond cond
: 3;
463 unsigned src0_swizzle
: 2;
464 unsigned src1_swizzle
: 2;
467 unsigned src0_neg
: 1;
470 } __attribute__((packed
));
472 enum bifrost_icmp_cond
{
473 BIFROST_ICMP_IGT
= 0,
474 BIFROST_ICMP_IGE
= 1,
475 BIFROST_ICMP_UGT
= 2,
476 BIFROST_ICMP_UGE
= 3,
478 BIFROST_ICMP_NEQ
= 5,
481 struct bifrost_fma_icmp32
{
484 enum bifrost_icmp_cond cond
: 3;
485 unsigned unk1
: 1; /* set */
488 } __attribute__((packed
));
490 struct bifrost_fma_icmp16
{
493 unsigned unk
: 5; /* 11010 */
494 enum bifrost_icmp_cond cond
: 3;
496 } __attribute__((packed
));
498 struct bifrost_add_icmp
{
501 enum bifrost_icmp_cond cond
: 3;
502 unsigned sz
: 1; /* 1 for 32, 0 for 8 */
505 } __attribute__((packed
));
507 /* Two sources for vectorization */
508 #define BIFROST_FMA_FLOAT32_TO_16 (0xdd000 >> 3)
509 #define BIFROST_ADD_FLOAT32_TO_16 (0x0EC00 >> 3)
511 enum bifrost_convert_mode
{
512 BIFROST_CONV_UNK0
= 0,
513 BIFROST_CONV_F32_TO_I32
= 1,
514 BIFROST_CONV_F16_TO_I16
= 2,
515 BIFROST_CONV_I32_TO_F32
= 3,
516 BIFROST_CONV_I16_TO_X32
= 4,
517 BIFROST_CONV_F16_TO_F32
= 5,
518 BIFROST_CONV_I16_TO_F16
= 6,
519 BIFROST_CONV_UNK7
= 7
523 #define BIFROST_CONVERT_4(is_unsigned, component, to_float) \
524 ((is_unsigned & 1) | ((component & 1) << 1) | ((to_float & 1) << 2) | \
525 ((0x3) << 3) | ((4) << 5) | 0x100)
528 #define BIFROST_CONVERT_5(component) \
529 ((component & 1) | ((1) << 1) | ((5) << 5) | 0x100)
531 /* Other conversions */
532 #define BIFROST_CONVERT(is_unsigned, roundmode, swizzle, mode) \
533 ((is_unsigned & 1) | ((roundmode & 3) << 1) | ((swizzle & 3) << 3) | ((mode & 7) << 5))
535 #define BIFROST_FMA_CONVERT (0xe0000)
536 #define BIFROST_ADD_CONVERT (0x07800)
538 enum bifrost_ldst_type
{
539 BIFROST_LDST_F16
= 0,
540 BIFROST_LDST_F32
= 1,
541 BIFROST_LDST_I32
= 2,
545 #define BIFROST_ADD_OP_LD_VAR_ADDR (0x18000 >> 10)
547 struct bifrost_ld_var_addr
{
550 unsigned location
: 5;
551 enum bifrost_ldst_type type
: 2;
553 } __attribute__((packed
));
555 #define BIFROST_ADD_OP_LD_ATTR (0x08000 >> 12)
557 struct bifrost_ld_attr
{
560 unsigned location
: 5;
561 unsigned channels
: 2; /* MALI_POSITIVE */
562 enum bifrost_ldst_type type
: 2;
564 } __attribute__((packed
));
566 enum bifrost_interp_mode
{
567 BIFROST_INTERP_PER_FRAG
= 0x0,
568 BIFROST_INTERP_CENTROID
= 0x1,
569 BIFROST_INTERP_DEFAULT
= 0x2,
570 BIFROST_INTERP_EXPLICIT
= 0x3
573 #define BIFROST_ADD_OP_LD_VAR_16 (0x1a << 1)
574 #define BIFROST_ADD_OP_LD_VAR_32 (0x0a << 1)
576 struct bifrost_ld_var
{
579 /* If top two bits set, indirect with src in bottom three */
582 unsigned channels
: 2; /* MALI_POSITIVE */
583 enum bifrost_interp_mode interp_mode
: 2;
587 } __attribute__((packed
));
589 struct bifrost_tex_ctrl
{
590 unsigned sampler_index
: 4; // also used to signal indirects
591 unsigned tex_index
: 7;
592 bool no_merge_index
: 1; // whether to merge (direct) sampler & texture indices
593 bool filter
: 1; // use the usual filtering pipeline (0 for texelFetch & textureGather)
595 bool texel_offset
: 1; // *Offset()
598 unsigned tex_type
: 2; // 2D, 3D, Cube, Buffer
599 bool compute_lod
: 1; // 0 for *Lod()
600 bool not_supply_lod
: 1; // 0 for *Lod() or when a bias is applied
601 bool calc_gradients
: 1; // 0 for *Grad()
603 unsigned result_type
: 4; // integer, unsigned, float TODO: why is this 4 bits?
605 } __attribute__((packed
));
607 struct bifrost_dual_tex_ctrl
{
608 unsigned sampler_index0
: 2;
610 unsigned tex_index0
: 2;
611 unsigned sampler_index1
: 2;
612 unsigned tex_index1
: 2;
614 } __attribute__((packed
));
616 #define BIFROST_ADD_OP_TEX_COMPACT_F32 (0x0b000 >> 10)
617 #define BIFROST_ADD_OP_TEX_COMPACT_F16 (0x1b000 >> 10)
619 struct bifrost_tex_compact
{
622 unsigned tex_index
: 3;
623 unsigned unknown
: 1;
624 unsigned sampler_index
: 3;
626 } __attribute__((packed
));
628 enum branch_bit_size
{
632 // For the above combinations of bitsize and location, an extra bit is
633 // encoded via comparing the sources. The only possible source of ambiguity
634 // would be if the sources were the same, but then the branch condition
635 // would be always true or always false anyways, so we can ignore it. But
636 // this no longer works when comparing the y component to the x component,
637 // since it's valid to compare the y component of a source against its own
638 // x component. Instead, the extra bit is encoded via an extra bitsize.
641 BR_SIZE_32_AND_16X
= 5,
642 BR_SIZE_32_AND_16Y
= 6,
643 // Used for comparisons with zero and always-true, see below. I think this
644 // only works for integer comparisons.
648 enum bifrost_reg_write_unit
{
649 REG_WRITE_NONE
= 0, // don't write
650 REG_WRITE_TWO
, // write using reg2
651 REG_WRITE_THREE
, // write using reg3
654 struct bifrost_regs
{
655 unsigned uniform_const
: 8;
661 } __attribute__((packed
));
663 enum bifrost_branch_cond
{
668 // Equal vs. not-equal determined by src0/src1 comparison
670 // floating-point comparisons
671 // Becomes UNE when you flip the arguments
673 // TODO what happens when you flip the arguments?
678 enum bifrost_branch_code
{
682 struct bifrost_branch
{
685 /* For BR_SIZE_ZERO, upper two bits become ctrl */
688 /* Offset source -- always uniform/const but
689 * theoretically could support indirect jumps? */
692 enum bifrost_branch_cond cond
: 3;
693 enum branch_bit_size size
: 3;
700 #define BIFROST_FMA_NOP (0x701960 | BIFROST_SRC_STAGE)
701 #define BIFROST_ADD_NOP (0x3D960 | BIFROST_SRC_STAGE)
703 struct bifrost_fmt1
{
708 uint64_t header
: 45;
709 } __attribute__((packed
));
711 #define BIFROST_FMT1_INSTRUCTIONS 0b00101
712 #define BIFROST_FMT1_FINAL 0b01001
713 #define BIFROST_FMT1_CONSTANTS 0b00001
715 #define BIFROST_FMTC_CONSTANTS 0b0011
716 #define BIFROST_FMTC_FINAL 0b0111
718 struct bifrost_fmt_constant
{
723 } __attribute__((packed
));
725 enum bifrost_reg_control
{
726 BIFROST_WRITE_FMA_P2
= 1,
727 BIFROST_WRITE_FMA_P2_READ_P3
= 2,
728 BIFROST_FIRST_WRITE_FMA_P2_READ_P3
= 3,
730 BIFROST_WRITE_ADD_P2
= 5,
731 BIFROST_WRITE_ADD_P2_READ_P3
= 6,
732 BIFROST_WRITE_ADD_P2_FMA_P3
= 7,
734 BIFROST_FIRST_NONE
= 8,
735 BIFROST_FIRST_WRITE_FMA_P2
= 9,
736 /* INSTR_INVALID_ENC */
737 BIFROST_REG_NONE
= 11,
738 BIFROST_FIRST_READ_P3
= 12,
739 BIFROST_FIRST_WRITE_ADD_P2
= 13,
740 BIFROST_FIRST_WRITE_ADD_P2_READ_P3
= 14,
741 BIFROST_FIRST_WRITE_ADD_P2_FMA_P3
= 15