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 struct bifrost_fma_inst
{
107 } __attribute__((packed
));
109 struct bifrost_fma_2src
{
113 } __attribute__((packed
));
115 #define BIFROST_FMA_OP_SEL8 (0x71)
117 struct bifrost_fma_sel8
{
122 unsigned swizzle
: 4;
124 } __attribute__((packed
));
126 #define BIFROST_FMA_OP_MSCALE (0x50 >> 3)
128 struct bifrost_fma_mscale
{
134 /* If mscale_mode is set - an MSCALE specific mode. If it is not set, a
137 unsigned mscale_mode
: 1;
139 unsigned src0_abs
: 1;
140 unsigned src1_neg
: 1;
141 unsigned src2_neg
: 1;
143 } __attribute__((packed
));
145 #define BIFROST_ADD_OP_BLEND (0x1952c)
146 #define BIFROST_ADD_OP_FRCP_FAST_F32 (0x0cc00)
147 #define BIFROST_ADD_OP_FRCP_FAST_F16_X (0x0ce10)
148 #define BIFROST_ADD_OP_FRCP_FAST_F16_Y (0x0ce30)
149 #define BIFROST_ADD_OP_FRSQ_FAST_F32 (0x0cc20)
150 #define BIFROST_ADD_OP_FRSQ_FAST_F16_X (0x0ce50)
151 #define BIFROST_ADD_OP_FRSQ_FAST_F16_Y (0x0ce70)
152 #define BIFROST_ADD_OP_LOG2_HELP (0x0cc68)
153 #define BIFROST_ADD_OP_FEXP2_FAST (0x0cd58)
155 struct bifrost_add_inst
{
158 } __attribute__((packed
));
160 #define BIFROST_ADD_OP_LD_UBO_1 (0x0c1a0 >> 3)
161 #define BIFROST_ADD_OP_LD_UBO_2 (0x0c1e0 >> 3)
162 #define BIFROST_ADD_OP_LD_UBO_3 (0x0caa0 >> 3)
163 #define BIFROST_ADD_OP_LD_UBO_4 (0x0c220 >> 3)
164 #define BIFROST_ADD_SEL_16(swiz) ((0xea60 >> 3) | (swiz))
166 struct bifrost_add_2src
{
170 } __attribute__((packed
));
172 #define BIFROST_ADD_OP_FMAX32 (0x00)
173 #define BIFROST_ADD_OP_FMIN32 (0x01)
174 #define BIFROST_ADD_OP_FADD32 (0x02)
176 #define BIFROST_ADD_OP_FADD16 (0x0A)
178 struct bifrost_add_faddmin
{
181 unsigned src1_abs
: 1;
182 unsigned src0_neg
: 1;
183 unsigned src1_neg
: 1;
184 unsigned select
: 2; /* swizzle_0 for fp16 */
185 unsigned outmod
: 2; /* swizzle_1 for fp16 */
187 unsigned src0_abs
: 1;
189 } __attribute__((packed
));
191 #define BIFROST_ADD_OP_FMAX16 (0x10)
192 #define BIFROST_ADD_OP_FMIN16 (0x12)
194 struct bifrost_add_fmin16
{
197 /* abs2 inferred as with FMA */
199 unsigned src0_neg
: 1;
200 unsigned src1_neg
: 1;
201 unsigned src0_swizzle
: 2;
202 unsigned src1_swizzle
: 2;
205 } __attribute__((packed
));
207 #define BIFROST_ADD_OP_ST_VAR (0x19300 >> 8)
209 struct bifrost_st_vary
{
213 unsigned channels
: 2;
215 } __attribute__((packed
));
217 #define BIFROST_ADD_OP_ATEST (0xc8f)
219 struct bifrost_add_atest
{
220 /* gl_SampleMask (R60) */
226 /* If half, X/Y select. If !half, always set */
227 unsigned component
: 1;
231 } __attribute__((packed
));
233 enum bifrost_outmod
{
236 BIFROST_SAT_SIGNED
= 0x2,
240 enum bifrost_roundmode
{
241 BIFROST_RTE
= 0x0, /* round to even */
242 BIFROST_RTP
= 0x1, /* round to positive */
243 BIFROST_RTN
= 0x2, /* round to negative */
244 BIFROST_RTZ
= 0x3 /* round to zero */
247 /* NONE: Same as fmax() and fmin() -- return the other
248 * number if any number is NaN. Also always return +0 if
249 * one argument is +0 and the other is -0.
251 * NAN_WINS: Instead of never returning a NaN, always return
252 * one. The "greater"/"lesser" NaN is always returned, first
253 * by checking the sign and then the mantissa bits.
255 * SRC1_WINS: For max, implement src0 > src1 ? src0 : src1.
256 * For min, implement src0 < src1 ? src0 : src1. This
257 * includes handling NaN's and signedness of 0 differently
258 * from above, since +0 and -0 compare equal and comparisons
259 * always return false for NaN's. As a result, this mode is
262 * SRC0_WINS: For max, implement src0 < src1 ? src1 : src0
263 * For min, implement src0 > src1 ? src1 : src0
267 enum bifrost_minmax_mode
{
268 BIFROST_MINMAX_NONE
= 0x0,
269 BIFROST_NAN_WINS
= 0x1,
270 BIFROST_SRC1_WINS
= 0x2,
271 BIFROST_SRC0_WINS
= 0x3,
274 #define BIFROST_FMA_OP_FADD32 (0x58 >> 2)
275 #define BIFROST_FMA_OP_FMAX32 (0x40 >> 2)
276 #define BIFROST_FMA_OP_FMIN32 (0x44 >> 2)
278 struct bifrost_fma_add
{
281 unsigned src1_abs
: 1;
282 unsigned src0_neg
: 1;
283 unsigned src1_neg
: 1;
285 unsigned src0_abs
: 1;
286 enum bifrost_roundmode roundmode
: 2;
287 enum bifrost_outmod outmod
: 2;
289 } __attribute__((packed
));
291 #define BIFROST_FMA_OP_FMAX16 (0xC0 >> 2)
292 #define BIFROST_FMA_OP_FMIN16 (0xCC >> 2)
293 #define BIFROST_FMA_OP_FADD16 (0xD8 >> 2)
295 struct bifrost_fma_add_minmax16
{
298 /* abs2 inferred as (src1 < src0) */
300 unsigned src0_neg
: 1;
301 unsigned src1_neg
: 1;
302 unsigned src0_swizzle
: 2;
303 unsigned src1_swizzle
: 2;
305 enum bifrost_outmod outmod
: 2;
306 /* roundmode for add, min/max mode for min/max */
308 } __attribute__((packed
));
310 #define BIFROST_FMA_OP_FMA (0x00)
312 struct bifrost_fma_fma
{
316 unsigned src_expand
: 3;
317 unsigned src0_abs
: 1;
318 enum bifrost_roundmode roundmode
: 2;
319 enum bifrost_outmod outmod
: 2;
320 unsigned src0_neg
: 1; /* 14 */
321 unsigned src2_neg
: 1;
322 unsigned src1_abs
: 1;
323 unsigned src2_abs
: 1; /* 17 */
325 } __attribute__((packed
));
327 #define BIFROST_FMA_OP_FMA16 (0x2)
329 struct bifrost_fma_fma16
{
333 unsigned swizzle_0
: 2;
334 unsigned swizzle_1
: 2;
335 enum bifrost_roundmode roundmode
: 2;
336 enum bifrost_outmod outmod
: 2;
337 unsigned src0_neg
: 1;
338 unsigned src2_neg
: 1;
339 unsigned swizzle_2
: 2;
341 } __attribute__((packed
));
343 enum bifrost_csel_cond
{
354 #define BIFROST_FMA_OP_CSEL4 (0x5c)
355 #define BIFROST_FMA_OP_CSEL4_V16 (0xdc)
357 struct bifrost_csel4
{
362 enum bifrost_csel_cond cond
: 3;
364 } __attribute__((packed
));
366 struct bifrost_shift_fma
{
370 unsigned half
: 3; /* 000 for i32, 100 for i8, 111 for v2i16 */
371 unsigned unk
: 1; /* always set? */
372 unsigned invert_1
: 1; /* Inverts sources to combining op */
373 /* For XOR, switches RSHIFT to LSHIFT since only one invert needed */
374 unsigned invert_2
: 1;
376 } __attribute__((packed
));
378 struct bifrost_shift_add
{
384 unsigned invert_1
: 1;
385 unsigned invert_2
: 1;
388 } __attribute__((packed
));
390 enum bifrost_fcmp_cond
{
399 struct bifrost_fma_fcmp
{
402 unsigned src1_abs
: 1;
404 unsigned src1_neg
: 1;
405 unsigned src_expand
: 3;
406 unsigned src0_abs
: 1;
407 enum bifrost_fcmp_cond cond
: 3;
409 } __attribute__((packed
));
411 struct bifrost_add_fcmp
{
414 enum bifrost_fcmp_cond cond
: 3;
415 unsigned src_expand
: 2;
416 unsigned src0_abs
: 1;
417 unsigned src1_abs
: 1;
418 unsigned src1_neg
: 1;
420 } __attribute__((packed
));
422 struct bifrost_fma_fcmp16
{
430 unsigned src0_swizzle
: 2;
431 unsigned src1_swizzle
: 2;
433 enum bifrost_fcmp_cond cond
: 3;
435 } __attribute__((packed
));
437 struct bifrost_add_fcmp16
{
440 enum bifrost_fcmp_cond cond
: 3;
442 unsigned src0_swizzle
: 2;
443 unsigned src1_swizzle
: 2;
446 unsigned src0_neg
: 1;
449 } __attribute__((packed
));
451 enum bifrost_icmp_cond
{
452 BIFROST_ICMP_IGT
= 0,
453 BIFROST_ICMP_IGE
= 1,
454 BIFROST_ICMP_UGT
= 2,
455 BIFROST_ICMP_UGE
= 3,
459 struct bifrost_fma_icmp32
{
462 enum bifrost_icmp_cond cond
: 3;
463 unsigned unk1
: 1; /* set */
466 } __attribute__((packed
));
468 /* Two sources for vectorization */
469 #define BIFROST_FMA_FLOAT32_TO_16 (0xdd000 >> 3)
470 #define BIFROST_ADD_FLOAT32_TO_16 (0x0EC00 >> 3)
472 enum bifrost_convert_mode
{
473 BIFROST_CONV_UNK0
= 0,
474 BIFROST_CONV_F32_TO_I32
= 1,
475 BIFROST_CONV_F16_TO_I16
= 2,
476 BIFROST_CONV_I32_TO_F32
= 3,
477 BIFROST_CONV_I16_TO_X32
= 4,
478 BIFROST_CONV_F16_TO_F32
= 5,
479 BIFROST_CONV_I16_TO_F16
= 6,
480 BIFROST_CONV_UNK7
= 7
484 #define BIFROST_CONVERT_4(is_unsigned, component, to_float) \
485 ((is_unsigned & 1) | ((component & 1) << 1) | ((to_float & 1) << 2) | \
486 ((0x3) << 3) | ((4) << 5) | 0x100)
489 #define BIFROST_CONVERT_5(component) \
490 ((component & 1) | ((1) << 1) | ((5) << 5) | 0x100)
492 /* Other conversions */
493 #define BIFROST_CONVERT(is_unsigned, roundmode, swizzle, mode) \
494 ((is_unsigned & 1) | ((roundmode & 3) << 1) | ((swizzle & 3) << 3) | ((mode & 7) << 5))
496 #define BIFROST_FMA_CONVERT (0xe0000)
497 #define BIFROST_ADD_CONVERT (0x07800)
499 enum bifrost_ldst_type
{
500 BIFROST_LDST_F16
= 0,
501 BIFROST_LDST_F32
= 1,
502 BIFROST_LDST_I32
= 2,
506 #define BIFROST_ADD_OP_LD_VAR_ADDR (0x18000 >> 10)
508 struct bifrost_ld_var_addr
{
511 unsigned location
: 5;
512 enum bifrost_ldst_type type
: 2;
514 } __attribute__((packed
));
516 #define BIFROST_ADD_OP_LD_ATTR (0x08000 >> 12)
518 struct bifrost_ld_attr
{
521 unsigned location
: 5;
522 unsigned channels
: 2; /* MALI_POSITIVE */
523 enum bifrost_ldst_type type
: 2;
525 } __attribute__((packed
));
527 enum bifrost_interp_mode
{
528 BIFROST_INTERP_PER_FRAG
= 0x0,
529 BIFROST_INTERP_CENTROID
= 0x1,
530 BIFROST_INTERP_DEFAULT
= 0x2,
531 BIFROST_INTERP_EXPLICIT
= 0x3
534 #define BIFROST_ADD_OP_LD_VAR_16 (0x1a << 1)
535 #define BIFROST_ADD_OP_LD_VAR_32 (0x0a << 1)
537 struct bifrost_ld_var
{
540 /* If top two bits set, indirect with src in bottom three */
543 unsigned channels
: 2; /* MALI_POSITIVE */
544 enum bifrost_interp_mode interp_mode
: 2;
548 } __attribute__((packed
));
550 struct bifrost_tex_ctrl
{
551 unsigned sampler_index
: 4; // also used to signal indirects
552 unsigned tex_index
: 7;
553 bool no_merge_index
: 1; // whether to merge (direct) sampler & texture indices
554 bool filter
: 1; // use the usual filtering pipeline (0 for texelFetch & textureGather)
556 bool texel_offset
: 1; // *Offset()
559 unsigned tex_type
: 2; // 2D, 3D, Cube, Buffer
560 bool compute_lod
: 1; // 0 for *Lod()
561 bool not_supply_lod
: 1; // 0 for *Lod() or when a bias is applied
562 bool calc_gradients
: 1; // 0 for *Grad()
564 unsigned result_type
: 4; // integer, unsigned, float TODO: why is this 4 bits?
566 } __attribute__((packed
));
568 struct bifrost_dual_tex_ctrl
{
569 unsigned sampler_index0
: 2;
571 unsigned tex_index0
: 2;
572 unsigned sampler_index1
: 2;
573 unsigned tex_index1
: 2;
575 } __attribute__((packed
));
577 #define BIFROST_ADD_OP_TEX_COMPACT_F32 (0x0b000 >> 10)
578 #define BIFROST_ADD_OP_TEX_COMPACT_F16 (0x1b000 >> 10)
580 struct bifrost_tex_compact
{
583 unsigned tex_index
: 3;
584 unsigned unknown
: 1;
585 unsigned sampler_index
: 3;
587 } __attribute__((packed
));
589 enum branch_bit_size
{
593 // For the above combinations of bitsize and location, an extra bit is
594 // encoded via comparing the sources. The only possible source of ambiguity
595 // would be if the sources were the same, but then the branch condition
596 // would be always true or always false anyways, so we can ignore it. But
597 // this no longer works when comparing the y component to the x component,
598 // since it's valid to compare the y component of a source against its own
599 // x component. Instead, the extra bit is encoded via an extra bitsize.
602 BR_SIZE_32_AND_16X
= 5,
603 BR_SIZE_32_AND_16Y
= 6,
604 // Used for comparisons with zero and always-true, see below. I think this
605 // only works for integer comparisons.
609 enum bifrost_reg_write_unit
{
610 REG_WRITE_NONE
= 0, // don't write
611 REG_WRITE_TWO
, // write using reg2
612 REG_WRITE_THREE
, // write using reg3
615 struct bifrost_regs
{
616 unsigned uniform_const
: 8;
622 } __attribute__((packed
));
624 enum bifrost_branch_cond
{
629 // Equal vs. not-equal determined by src0/src1 comparison
631 // floating-point comparisons
632 // Becomes UNE when you flip the arguments
634 // TODO what happens when you flip the arguments?
639 enum bifrost_branch_code
{
643 struct bifrost_branch
{
646 /* For BR_SIZE_ZERO, upper two bits become ctrl */
649 /* Offset source -- always uniform/const but
650 * theoretically could support indirect jumps? */
653 enum bifrost_branch_cond cond
: 3;
654 enum branch_bit_size size
: 3;
661 #define BIFROST_FMA_NOP (0x701960 | BIFROST_SRC_STAGE)
662 #define BIFROST_ADD_NOP (0x3D960 | BIFROST_SRC_STAGE)
664 struct bifrost_fmt1
{
669 uint64_t header
: 45;
670 } __attribute__((packed
));
672 #define BIFROST_FMT1_INSTRUCTIONS 0b00101
673 #define BIFROST_FMT1_FINAL 0b01001
674 #define BIFROST_FMT1_CONSTANTS 0b00001
676 #define BIFROST_FMTC_CONSTANTS 0b0011
677 #define BIFROST_FMTC_FINAL 0b0111
679 struct bifrost_fmt_constant
{
684 } __attribute__((packed
));
686 enum bifrost_reg_control
{
687 BIFROST_WRITE_FMA_P2
= 1,
688 BIFROST_WRITE_FMA_P2_READ_P3
= 2,
689 BIFROST_FIRST_WRITE_FMA_P2_READ_P3
= 3,
691 BIFROST_WRITE_ADD_P2
= 5,
692 BIFROST_WRITE_ADD_P2_READ_P3
= 6,
693 BIFROST_WRITE_ADD_P2_FMA_P3
= 7,
695 BIFROST_FIRST_NONE
= 8,
696 BIFROST_FIRST_WRITE_FMA_P2
= 9,
697 /* INSTR_INVALID_ENC */
698 BIFROST_REG_NONE
= 11,
699 BIFROST_FIRST_READ_P3
= 12,
700 BIFROST_FIRST_WRITE_ADD_P2
= 13,
701 BIFROST_FIRST_WRITE_ADD_P2_READ_P3
= 14,
702 BIFROST_FIRST_WRITE_ADD_P2_FMA_P3
= 15