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 #define BIFROST_DBG_MSGS 0x0001
33 #define BIFROST_DBG_SHADERS 0x0002
35 extern int bifrost_debug
;
37 enum bifrost_clause_type
{
38 BIFROST_CLAUSE_NONE
= 0,
39 BIFROST_CLAUSE_LOAD_VARY
= 1,
40 BIFROST_CLAUSE_UBO
= 2,
41 BIFROST_CLAUSE_TEX
= 3,
42 BIFROST_CLAUSE_SSBO_LOAD
= 5,
43 BIFROST_CLAUSE_SSBO_STORE
= 6,
44 BIFROST_CLAUSE_BLEND
= 9,
45 BIFROST_CLAUSE_FRAGZ
= 12,
46 BIFROST_CLAUSE_ATEST
= 13,
47 BIFROST_CLAUSE_64BIT
= 15
50 struct bifrost_header
{
52 // If true, convert any infinite result of any floating-point operation to
53 // the biggest representable number.
54 unsigned suppress_inf
: 1;
55 // Convert any NaN results to 0.
56 unsigned suppress_nan
: 1;
58 // true if the execution mask of the next clause is the same as the mask of
59 // the current clause.
60 unsigned back_to_back
: 1;
61 unsigned no_end_of_shader
: 1;
63 // Set to true for fragment shaders, to implement this bit of spec text
64 // from section 7.1.5 of the GLSL ES spec:
66 // "Stores to image and buffer variables performed by helper invocations
67 // have no effect on the underlying image or buffer memory."
69 // Helper invocations are threads (invocations) corresponding to pixels in
70 // a quad that aren't actually part of the triangle, but are included to
71 // make derivatives work correctly. They're usually turned on, but they
72 // need to be masked off for GLSL-level stores. This bit seems to be the
73 // only bit that's actually different between fragment shaders and other
74 // shaders, so this is probably what it's doing.
75 unsigned elide_writes
: 1;
76 // If backToBack is off:
77 // - true for conditional branches and fallthrough
78 // - false for unconditional branches
79 // The blob seems to always set it to true if back-to-back is on.
80 unsigned branch_cond
: 1;
81 // This bit is set when the next clause writes to the data register of some
83 unsigned datareg_writebarrier
: 1;
85 unsigned scoreboard_deps
: 8;
86 unsigned scoreboard_index
: 3;
87 enum bifrost_clause_type clause_type
: 4;
88 unsigned unk3
: 1; // part of clauseType?
89 enum bifrost_clause_type next_clause_type
: 4;
90 unsigned unk4
: 1; // part of nextClauseType?
91 } __attribute__((packed
));
93 enum bifrost_packed_src
{
94 BIFROST_SRC_PORT0
= 0,
95 BIFROST_SRC_PORT1
= 1,
96 BIFROST_SRC_PORT3
= 2,
97 BIFROST_SRC_STAGE
= 3,
98 BIFROST_SRC_CONST_LO
= 4,
99 BIFROST_SRC_CONST_HI
= 5,
100 BIFROST_SRC_PASS_FMA
= 6,
101 BIFROST_SRC_PASS_ADD
= 7,
104 #define BIFROST_FMA_EXT (0xe0000)
105 #define BIFROST_FMA_OP_MOV BIFROST_FMA_EXT | (0x32d)
106 #define BIFROST_FMA_OP_FREXPE_LOG BIFROST_FMA_EXT | 0x3c5
107 #define BIFROST_FMA_OP_ADD_FREXPM ((BIFROST_FMA_EXT | 0x1e80) >> 3)
108 #define BIFROST_FMA_SEL_16(swiz) (((BIFROST_FMA_EXT | 0x1e00) >> 3) | (swiz))
110 #define BIFROST_FMA_ROUND_16(mode, swiz) (BIFROST_FMA_EXT | 0x1800 | (swiz) | ((mode) << 6))
111 #define BIFROST_FMA_ROUND_32(mode) (BIFROST_FMA_EXT | 0x1805 | ((mode) << 6))
113 struct bifrost_fma_inst
{
116 } __attribute__((packed
));
118 struct bifrost_fma_2src
{
122 } __attribute__((packed
));
124 #define BIFROST_FMA_OP_SEL8 (0x71)
126 struct bifrost_fma_sel8
{
131 unsigned swizzle
: 4;
133 } __attribute__((packed
));
135 #define BIFROST_FMA_OP_MSCALE (0x50 >> 3)
137 struct bifrost_fma_mscale
{
143 /* If mscale_mode is set - an MSCALE specific mode. If it is not set, a
146 unsigned mscale_mode
: 1;
148 unsigned src0_abs
: 1;
149 unsigned src1_neg
: 1;
150 unsigned src2_neg
: 1;
152 } __attribute__((packed
));
154 #define BIFROST_ADD_OP_BLEND (0x1952c)
155 #define BIFROST_ADD_OP_FRCP_FAST_F32 (0x0cc00)
156 #define BIFROST_ADD_OP_FRCP_FAST_F16_X (0x0ce10)
157 #define BIFROST_ADD_OP_FRCP_FAST_F16_Y (0x0ce30)
158 #define BIFROST_ADD_OP_FRSQ_FAST_F32 (0x0cc20)
159 #define BIFROST_ADD_OP_FRSQ_FAST_F16_X (0x0ce50)
160 #define BIFROST_ADD_OP_FRSQ_FAST_F16_Y (0x0ce70)
161 #define BIFROST_ADD_OP_LOG2_HELP (0x0cc68)
162 #define BIFROST_ADD_OP_FEXP2_FAST (0x0cd58)
164 struct bifrost_add_inst
{
167 } __attribute__((packed
));
169 #define BIFROST_ADD_OP_LD_UBO_1 (0x0c1a0 >> 3)
170 #define BIFROST_ADD_OP_LD_UBO_2 (0x0c1e0 >> 3)
171 #define BIFROST_ADD_OP_LD_UBO_3 (0x0caa0 >> 3)
172 #define BIFROST_ADD_OP_LD_UBO_4 (0x0c220 >> 3)
173 #define BIFROST_ADD_SEL_16(swiz) ((0xea60 >> 3) | (swiz))
175 struct bifrost_add_2src
{
179 } __attribute__((packed
));
181 #define BIFROST_ADD_OP_FMAX32 (0x00)
182 #define BIFROST_ADD_OP_FMIN32 (0x01)
183 #define BIFROST_ADD_OP_FADD32 (0x02)
185 #define BIFROST_ADD_OP_FADD16 (0x0A)
187 struct bifrost_add_faddmin
{
190 unsigned src1_abs
: 1;
191 unsigned src0_neg
: 1;
192 unsigned src1_neg
: 1;
193 unsigned select
: 2; /* swizzle_0 for fp16 */
194 unsigned outmod
: 2; /* swizzle_1 for fp16 */
196 unsigned src0_abs
: 1;
198 } __attribute__((packed
));
200 #define BIFROST_ADD_OP_FMAX16 (0x10)
201 #define BIFROST_ADD_OP_FMIN16 (0x12)
203 struct bifrost_add_fmin16
{
206 /* abs2 inferred as with FMA */
208 unsigned src0_neg
: 1;
209 unsigned src1_neg
: 1;
210 unsigned src0_swizzle
: 2;
211 unsigned src1_swizzle
: 2;
214 } __attribute__((packed
));
216 #define BIFROST_ADD_OP_ST_VAR (0x19300 >> 8)
218 struct bifrost_st_vary
{
222 unsigned channels
: 2;
224 } __attribute__((packed
));
226 #define BIFROST_ADD_OP_ATEST (0xc8f)
228 struct bifrost_add_atest
{
229 /* gl_SampleMask (R60) */
235 /* If half, X/Y select. If !half, always set */
236 unsigned component
: 1;
240 } __attribute__((packed
));
242 enum bifrost_outmod
{
245 BIFROST_SAT_SIGNED
= 0x2,
249 enum bifrost_roundmode
{
250 BIFROST_RTE
= 0x0, /* round to even */
251 BIFROST_RTP
= 0x1, /* round to positive */
252 BIFROST_RTN
= 0x2, /* round to negative */
253 BIFROST_RTZ
= 0x3 /* round to zero */
256 /* NONE: Same as fmax() and fmin() -- return the other
257 * number if any number is NaN. Also always return +0 if
258 * one argument is +0 and the other is -0.
260 * NAN_WINS: Instead of never returning a NaN, always return
261 * one. The "greater"/"lesser" NaN is always returned, first
262 * by checking the sign and then the mantissa bits.
264 * SRC1_WINS: For max, implement src0 > src1 ? src0 : src1.
265 * For min, implement src0 < src1 ? src0 : src1. This
266 * includes handling NaN's and signedness of 0 differently
267 * from above, since +0 and -0 compare equal and comparisons
268 * always return false for NaN's. As a result, this mode is
271 * SRC0_WINS: For max, implement src0 < src1 ? src1 : src0
272 * For min, implement src0 > src1 ? src1 : src0
276 enum bifrost_minmax_mode
{
277 BIFROST_MINMAX_NONE
= 0x0,
278 BIFROST_NAN_WINS
= 0x1,
279 BIFROST_SRC1_WINS
= 0x2,
280 BIFROST_SRC0_WINS
= 0x3,
283 #define BIFROST_FMA_OP_FADD32 (0x58 >> 2)
284 #define BIFROST_FMA_OP_FMAX32 (0x40 >> 2)
285 #define BIFROST_FMA_OP_FMIN32 (0x44 >> 2)
287 struct bifrost_fma_add
{
290 unsigned src1_abs
: 1;
291 unsigned src0_neg
: 1;
292 unsigned src1_neg
: 1;
294 unsigned src0_abs
: 1;
295 enum bifrost_roundmode roundmode
: 2;
296 enum bifrost_outmod outmod
: 2;
298 } __attribute__((packed
));
300 #define BIFROST_FMA_OP_FMAX16 (0xC0 >> 2)
301 #define BIFROST_FMA_OP_FMIN16 (0xCC >> 2)
302 #define BIFROST_FMA_OP_FADD16 (0xD8 >> 2)
304 struct bifrost_fma_add_minmax16
{
307 /* abs2 inferred as (src1 < src0) */
309 unsigned src0_neg
: 1;
310 unsigned src1_neg
: 1;
311 unsigned src0_swizzle
: 2;
312 unsigned src1_swizzle
: 2;
314 enum bifrost_outmod outmod
: 2;
315 /* roundmode for add, min/max mode for min/max */
317 } __attribute__((packed
));
319 #define BIFROST_FMA_OP_FMA (0x00)
321 struct bifrost_fma_fma
{
325 unsigned src_expand
: 3;
326 unsigned src0_abs
: 1;
327 enum bifrost_roundmode roundmode
: 2;
328 enum bifrost_outmod outmod
: 2;
329 unsigned src0_neg
: 1; /* 14 */
330 unsigned src2_neg
: 1;
331 unsigned src1_abs
: 1;
332 unsigned src2_abs
: 1; /* 17 */
334 } __attribute__((packed
));
336 #define BIFROST_FMA_OP_FMA16 (0x2)
338 struct bifrost_fma_fma16
{
342 unsigned swizzle_0
: 2;
343 unsigned swizzle_1
: 2;
344 enum bifrost_roundmode roundmode
: 2;
345 enum bifrost_outmod outmod
: 2;
346 unsigned src0_neg
: 1;
347 unsigned src2_neg
: 1;
348 unsigned swizzle_2
: 2;
350 } __attribute__((packed
));
352 enum bifrost_csel_cond
{
363 #define BIFROST_FMA_OP_CSEL4 (0x5c)
364 #define BIFROST_FMA_OP_CSEL4_V16 (0xdc)
366 struct bifrost_csel4
{
371 enum bifrost_csel_cond cond
: 3;
373 } __attribute__((packed
));
375 #define BIFROST_FMA_OP_RSHIFT_NAND (0x60000 >> 12)
376 #define BIFROST_FMA_OP_RSHIFT_AND (0x61000 >> 12)
377 #define BIFROST_FMA_OP_LSHIFT_NAND (0x62000 >> 12)
378 #define BIFROST_FMA_OP_LSHIFT_AND (0x63000 >> 12)
379 #define BIFROST_FMA_OP_RSHIFT_XOR (0x64000 >> 12)
380 #define BIFROST_FMA_OP_LSHIFT_ADD_32 (0x65200 >> 6)
381 #define BIFROST_FMA_OP_LSHIFT_SUB_32 (0x65600 >> 6)
382 #define BIFROST_FMA_OP_LSHIFT_RSUB_32 (0x65a00 >> 6)
383 #define BIFROST_FMA_OP_RSHIFT_ADD_32 (0x65e00 >> 6)
384 #define BIFROST_FMA_OP_RSHIFT_SUB_32 (0x66200 >> 6)
385 #define BIFROST_FMA_OP_RSHIFT_RSUB_32 (0x66600 >> 6)
387 struct bifrost_shift_fma
{
392 unsigned unk
: 1; /* always set? */
393 unsigned invert_1
: 1; /* Inverts sources to combining op */
394 /* For XOR, switches RSHIFT to LSHIFT since only one invert needed */
395 unsigned invert_2
: 1;
397 } __attribute__((packed
));
399 struct bifrost_shift_add
{
405 unsigned invert_1
: 1;
406 unsigned invert_2
: 1;
409 } __attribute__((packed
));
411 enum bifrost_fcmp_cond
{
420 #define BIFROST_FMA_OP_FCMP_GL (0x48000 >> 13)
421 #define BIFROST_FMA_OP_FCMP_D3D (0x4c000 >> 13)
423 struct bifrost_fma_fcmp
{
426 unsigned src1_abs
: 1;
428 unsigned src1_neg
: 1;
429 unsigned src_expand
: 3;
430 unsigned src0_abs
: 1;
431 enum bifrost_fcmp_cond cond
: 3;
433 } __attribute__((packed
));
435 struct bifrost_add_fcmp
{
438 enum bifrost_fcmp_cond cond
: 3;
439 unsigned src_expand
: 2;
440 unsigned src0_abs
: 1;
441 unsigned src1_abs
: 1;
442 unsigned src1_neg
: 1;
444 } __attribute__((packed
));
446 #define BIFROST_FMA_OP_FCMP_GL_16 (0xc8000 >> 13)
447 #define BIFROST_FMA_OP_FCMP_D3D_16 (0xcc000 >> 13)
449 struct bifrost_fma_fcmp16
{
457 unsigned src0_swizzle
: 2;
458 unsigned src1_swizzle
: 2;
460 enum bifrost_fcmp_cond cond
: 3;
462 } __attribute__((packed
));
464 struct bifrost_add_fcmp16
{
467 enum bifrost_fcmp_cond cond
: 3;
469 unsigned src0_swizzle
: 2;
470 unsigned src1_swizzle
: 2;
473 unsigned src0_neg
: 1;
476 } __attribute__((packed
));
478 enum bifrost_icmp_cond
{
479 BIFROST_ICMP_IGT
= 0,
480 BIFROST_ICMP_IGE
= 1,
481 BIFROST_ICMP_UGT
= 2,
482 BIFROST_ICMP_UGE
= 3,
484 BIFROST_ICMP_NEQ
= 5,
487 struct bifrost_fma_icmp32
{
490 enum bifrost_icmp_cond cond
: 3;
491 unsigned unk1
: 1; /* set */
494 } __attribute__((packed
));
496 struct bifrost_fma_icmp16
{
499 unsigned unk
: 5; /* 11010 */
500 enum bifrost_icmp_cond cond
: 3;
502 } __attribute__((packed
));
504 struct bifrost_add_icmp
{
507 enum bifrost_icmp_cond cond
: 3;
508 unsigned sz
: 1; /* 1 for 32, 0 for 8 */
511 } __attribute__((packed
));
513 /* Two sources for vectorization */
514 #define BIFROST_FMA_FLOAT32_TO_16 (0xdd000 >> 3)
515 #define BIFROST_ADD_FLOAT32_TO_16 (0x0EC00 >> 3)
517 enum bifrost_convert_mode
{
518 BIFROST_CONV_UNK0
= 0,
519 BIFROST_CONV_F32_TO_I32
= 1,
520 BIFROST_CONV_F16_TO_I16
= 2,
521 BIFROST_CONV_I32_TO_F32
= 3,
522 BIFROST_CONV_I16_TO_X32
= 4,
523 BIFROST_CONV_F16_TO_F32
= 5,
524 BIFROST_CONV_I16_TO_F16
= 6,
525 BIFROST_CONV_UNK7
= 7
529 #define BIFROST_CONVERT_4(is_unsigned, component, to_float) \
530 ((is_unsigned & 1) | ((component & 1) << 1) | ((to_float & 1) << 2) | \
531 ((0x3) << 3) | ((4) << 5) | 0x100)
534 #define BIFROST_CONVERT_5(component) \
535 ((component & 1) | ((1) << 1) | ((5) << 5) | 0x100)
537 /* Other conversions */
538 #define BIFROST_CONVERT(is_unsigned, roundmode, swizzle, mode) \
539 ((is_unsigned & 1) | ((roundmode & 3) << 1) | ((swizzle & 3) << 3) | ((mode & 7) << 5))
541 #define BIFROST_FMA_CONVERT (0xe0000)
542 #define BIFROST_ADD_CONVERT (0x07800)
544 enum bifrost_ldst_type
{
545 BIFROST_LDST_F16
= 0,
546 BIFROST_LDST_F32
= 1,
547 BIFROST_LDST_I32
= 2,
551 #define BIFROST_ADD_OP_LD_VAR_ADDR (0x18000 >> 10)
553 struct bifrost_ld_var_addr
{
556 unsigned location
: 5;
557 enum bifrost_ldst_type type
: 2;
559 } __attribute__((packed
));
561 #define BIFROST_ADD_OP_LD_ATTR (0x08000 >> 12)
563 struct bifrost_ld_attr
{
566 unsigned location
: 5;
567 unsigned channels
: 2; /* MALI_POSITIVE */
568 enum bifrost_ldst_type type
: 2;
570 } __attribute__((packed
));
572 enum bifrost_interp_mode
{
573 BIFROST_INTERP_PER_FRAG
= 0x0,
574 BIFROST_INTERP_CENTROID
= 0x1,
575 BIFROST_INTERP_DEFAULT
= 0x2,
576 BIFROST_INTERP_EXPLICIT
= 0x3
579 #define BIFROST_ADD_OP_LD_VAR_16 (0x1a << 1)
580 #define BIFROST_ADD_OP_LD_VAR_32 (0x0a << 1)
582 struct bifrost_ld_var
{
585 /* If top two bits set, indirect with src in bottom three */
588 unsigned channels
: 2; /* MALI_POSITIVE */
589 enum bifrost_interp_mode interp_mode
: 2;
593 } __attribute__((packed
));
595 struct bifrost_tex_ctrl
{
596 unsigned sampler_index
: 4; // also used to signal indirects
597 unsigned tex_index
: 7;
598 bool no_merge_index
: 1; // whether to merge (direct) sampler & texture indices
599 bool filter
: 1; // use the usual filtering pipeline (0 for texelFetch & textureGather)
601 bool texel_offset
: 1; // *Offset()
604 unsigned tex_type
: 2; // 2D, 3D, Cube, Buffer
605 bool compute_lod
: 1; // 0 for *Lod()
606 bool not_supply_lod
: 1; // 0 for *Lod() or when a bias is applied
607 bool calc_gradients
: 1; // 0 for *Grad()
609 unsigned result_type
: 4; // integer, unsigned, float TODO: why is this 4 bits?
611 } __attribute__((packed
));
613 struct bifrost_dual_tex_ctrl
{
614 unsigned sampler_index0
: 2;
616 unsigned tex_index0
: 2;
617 unsigned sampler_index1
: 2;
618 unsigned tex_index1
: 2;
620 } __attribute__((packed
));
622 #define BIFROST_ADD_OP_TEX_COMPACT_F32 (0x0b000 >> 10)
623 #define BIFROST_ADD_OP_TEX_COMPACT_F16 (0x1b000 >> 10)
625 struct bifrost_tex_compact
{
628 unsigned tex_index
: 3;
629 unsigned unknown
: 1;
630 unsigned sampler_index
: 3;
632 } __attribute__((packed
));
634 enum branch_bit_size
{
638 // For the above combinations of bitsize and location, an extra bit is
639 // encoded via comparing the sources. The only possible source of ambiguity
640 // would be if the sources were the same, but then the branch condition
641 // would be always true or always false anyways, so we can ignore it. But
642 // this no longer works when comparing the y component to the x component,
643 // since it's valid to compare the y component of a source against its own
644 // x component. Instead, the extra bit is encoded via an extra bitsize.
647 BR_SIZE_32_AND_16X
= 5,
648 BR_SIZE_32_AND_16Y
= 6,
649 // Used for comparisons with zero and always-true, see below. I think this
650 // only works for integer comparisons.
654 enum bifrost_reg_write_unit
{
655 REG_WRITE_NONE
= 0, // don't write
656 REG_WRITE_TWO
, // write using reg2
657 REG_WRITE_THREE
, // write using reg3
660 struct bifrost_regs
{
661 unsigned uniform_const
: 8;
667 } __attribute__((packed
));
669 enum bifrost_branch_cond
{
674 // Equal vs. not-equal determined by src0/src1 comparison
676 // floating-point comparisons
677 // Becomes UNE when you flip the arguments
679 // TODO what happens when you flip the arguments?
684 enum bifrost_branch_code
{
688 struct bifrost_branch
{
691 /* For BR_SIZE_ZERO, upper two bits become ctrl */
694 /* Offset source -- always uniform/const but
695 * theoretically could support indirect jumps? */
698 enum bifrost_branch_cond cond
: 3;
699 enum branch_bit_size size
: 3;
706 #define BIFROST_FMA_NOP (0x701960 | BIFROST_SRC_STAGE)
707 #define BIFROST_ADD_NOP (0x3D960 | BIFROST_SRC_STAGE)
709 struct bifrost_fmt1
{
714 uint64_t header
: 45;
715 } __attribute__((packed
));
717 #define BIFROST_FMT1_INSTRUCTIONS 0b00101
718 #define BIFROST_FMT1_FINAL 0b01001
719 #define BIFROST_FMT1_CONSTANTS 0b00001
721 #define BIFROST_FMTC_CONSTANTS 0b0011
722 #define BIFROST_FMTC_FINAL 0b0111
724 struct bifrost_fmt_constant
{
729 } __attribute__((packed
));
731 enum bifrost_reg_control
{
732 BIFROST_WRITE_FMA_P2
= 1,
733 BIFROST_WRITE_FMA_P2_READ_P3
= 2,
734 BIFROST_FIRST_WRITE_FMA_P2_READ_P3
= 3,
736 BIFROST_WRITE_ADD_P2
= 5,
737 BIFROST_WRITE_ADD_P2_READ_P3
= 6,
738 BIFROST_WRITE_ADD_P2_FMA_P3
= 7,
740 BIFROST_FIRST_NONE
= 8,
741 BIFROST_FIRST_WRITE_FMA_P2
= 9,
742 /* INSTR_INVALID_ENC */
743 BIFROST_REG_NONE
= 11,
744 BIFROST_FIRST_READ_P3
= 12,
745 BIFROST_FIRST_WRITE_ADD_P2
= 13,
746 BIFROST_FIRST_WRITE_ADD_P2_READ_P3
= 14,
747 BIFROST_FIRST_WRITE_ADD_P2_FMA_P3
= 15