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_ATEST
= 13,
46 BIFROST_CLAUSE_64BIT
= 15
49 struct bifrost_header
{
51 // If true, convert any infinite result of any floating-point operation to
52 // the biggest representable number.
53 unsigned suppress_inf
: 1;
54 // Convert any NaN results to 0.
55 unsigned suppress_nan
: 1;
57 // true if the execution mask of the next clause is the same as the mask of
58 // the current clause.
59 unsigned back_to_back
: 1;
60 unsigned no_end_of_shader
: 1;
62 // Set to true for fragment shaders, to implement this bit of spec text
63 // from section 7.1.5 of the GLSL ES spec:
65 // "Stores to image and buffer variables performed by helper invocations
66 // have no effect on the underlying image or buffer memory."
68 // Helper invocations are threads (invocations) corresponding to pixels in
69 // a quad that aren't actually part of the triangle, but are included to
70 // make derivatives work correctly. They're usually turned on, but they
71 // need to be masked off for GLSL-level stores. This bit seems to be the
72 // only bit that's actually different between fragment shaders and other
73 // shaders, so this is probably what it's doing.
74 unsigned elide_writes
: 1;
75 // If backToBack is off:
76 // - true for conditional branches and fallthrough
77 // - false for unconditional branches
78 // The blob seems to always set it to true if back-to-back is on.
79 unsigned branch_cond
: 1;
80 // This bit is set when the next clause writes to the data register of some
82 unsigned datareg_writebarrier
: 1;
84 unsigned scoreboard_deps
: 8;
85 unsigned scoreboard_index
: 3;
86 enum bifrost_clause_type clause_type
: 4;
87 unsigned unk3
: 1; // part of clauseType?
88 enum bifrost_clause_type next_clause_type
: 4;
89 unsigned unk4
: 1; // part of nextClauseType?
90 } __attribute__((packed
));
92 enum bifrost_packed_src
{
93 BIFROST_SRC_PORT0
= 0,
94 BIFROST_SRC_PORT1
= 1,
95 BIFROST_SRC_PORT3
= 2,
96 BIFROST_SRC_STAGE
= 3,
97 BIFROST_SRC_CONST_LO
= 4,
98 BIFROST_SRC_CONST_HI
= 5,
99 BIFROST_SRC_PASS_FMA
= 6,
100 BIFROST_SRC_PASS_ADD
= 7,
103 #define BIFROST_FMA_EXT (0xe0000)
104 #define BIFROST_FMA_OP_MOV BIFROST_FMA_EXT | (0x32d)
105 #define BIFROST_FMA_OP_FREXPE_LOG BIFROST_FMA_EXT | 0x3c5
106 #define BIFROST_FMA_OP_ADD_FREXPM ((BIFROST_FMA_EXT | 0x1e80) >> 3)
107 #define BIFROST_FMA_SEL_16(swiz) (((BIFROST_FMA_EXT | 0x1e00) >> 3) | (swiz))
109 #define BIFROST_FMA_ROUND_16(mode, swiz) (BIFROST_FMA_EXT | 0x1800 | (swiz) | ((mode) << 6))
110 #define BIFROST_FMA_ROUND_32(mode) (BIFROST_FMA_EXT | 0x1805 | ((mode) << 6))
112 struct bifrost_fma_inst
{
115 } __attribute__((packed
));
117 struct bifrost_fma_2src
{
121 } __attribute__((packed
));
123 #define BIFROST_FMA_OP_SEL8 (0x71)
125 struct bifrost_fma_sel8
{
130 unsigned swizzle
: 4;
132 } __attribute__((packed
));
134 #define BIFROST_FMA_OP_MSCALE (0x50 >> 3)
136 struct bifrost_fma_mscale
{
142 /* If mscale_mode is set - an MSCALE specific mode. If it is not set, a
145 unsigned mscale_mode
: 1;
147 unsigned src0_abs
: 1;
148 unsigned src1_neg
: 1;
149 unsigned src2_neg
: 1;
151 } __attribute__((packed
));
153 #define BIFROST_ADD_OP_BLEND (0x1952c)
154 #define BIFROST_ADD_OP_FRCP_FAST_F32 (0x0cc00)
155 #define BIFROST_ADD_OP_FRCP_FAST_F16_X (0x0ce10)
156 #define BIFROST_ADD_OP_FRCP_FAST_F16_Y (0x0ce30)
157 #define BIFROST_ADD_OP_FRSQ_FAST_F32 (0x0cc20)
158 #define BIFROST_ADD_OP_FRSQ_FAST_F16_X (0x0ce50)
159 #define BIFROST_ADD_OP_FRSQ_FAST_F16_Y (0x0ce70)
160 #define BIFROST_ADD_OP_LOG2_HELP (0x0cc68)
161 #define BIFROST_ADD_OP_FEXP2_FAST (0x0cd58)
163 struct bifrost_add_inst
{
166 } __attribute__((packed
));
168 #define BIFROST_ADD_OP_LD_UBO_1 (0x0c1a0 >> 3)
169 #define BIFROST_ADD_OP_LD_UBO_2 (0x0c1e0 >> 3)
170 #define BIFROST_ADD_OP_LD_UBO_3 (0x0caa0 >> 3)
171 #define BIFROST_ADD_OP_LD_UBO_4 (0x0c220 >> 3)
172 #define BIFROST_ADD_SEL_16(swiz) ((0xea60 >> 3) | (swiz))
174 struct bifrost_add_2src
{
178 } __attribute__((packed
));
180 #define BIFROST_ADD_OP_FMAX32 (0x00)
181 #define BIFROST_ADD_OP_FMIN32 (0x01)
182 #define BIFROST_ADD_OP_FADD32 (0x02)
184 #define BIFROST_ADD_OP_FADD16 (0x0A)
186 struct bifrost_add_faddmin
{
189 unsigned src1_abs
: 1;
190 unsigned src0_neg
: 1;
191 unsigned src1_neg
: 1;
192 unsigned select
: 2; /* swizzle_0 for fp16 */
193 unsigned outmod
: 2; /* swizzle_1 for fp16 */
195 unsigned src0_abs
: 1;
197 } __attribute__((packed
));
199 #define BIFROST_ADD_OP_FMAX16 (0x10)
200 #define BIFROST_ADD_OP_FMIN16 (0x12)
202 struct bifrost_add_fmin16
{
205 /* abs2 inferred as with FMA */
207 unsigned src0_neg
: 1;
208 unsigned src1_neg
: 1;
209 unsigned src0_swizzle
: 2;
210 unsigned src1_swizzle
: 2;
213 } __attribute__((packed
));
215 #define BIFROST_ADD_OP_ST_VAR (0x19300 >> 8)
217 struct bifrost_st_vary
{
221 unsigned channels
: 2;
223 } __attribute__((packed
));
225 #define BIFROST_ADD_OP_ATEST (0xc8f)
227 struct bifrost_add_atest
{
228 /* gl_SampleMask (R60) */
234 /* If half, X/Y select. If !half, always set */
235 unsigned component
: 1;
239 } __attribute__((packed
));
241 enum bifrost_outmod
{
244 BIFROST_SAT_SIGNED
= 0x2,
248 enum bifrost_roundmode
{
249 BIFROST_RTE
= 0x0, /* round to even */
250 BIFROST_RTP
= 0x1, /* round to positive */
251 BIFROST_RTN
= 0x2, /* round to negative */
252 BIFROST_RTZ
= 0x3 /* round to zero */
255 /* NONE: Same as fmax() and fmin() -- return the other
256 * number if any number is NaN. Also always return +0 if
257 * one argument is +0 and the other is -0.
259 * NAN_WINS: Instead of never returning a NaN, always return
260 * one. The "greater"/"lesser" NaN is always returned, first
261 * by checking the sign and then the mantissa bits.
263 * SRC1_WINS: For max, implement src0 > src1 ? src0 : src1.
264 * For min, implement src0 < src1 ? src0 : src1. This
265 * includes handling NaN's and signedness of 0 differently
266 * from above, since +0 and -0 compare equal and comparisons
267 * always return false for NaN's. As a result, this mode is
270 * SRC0_WINS: For max, implement src0 < src1 ? src1 : src0
271 * For min, implement src0 > src1 ? src1 : src0
275 enum bifrost_minmax_mode
{
276 BIFROST_MINMAX_NONE
= 0x0,
277 BIFROST_NAN_WINS
= 0x1,
278 BIFROST_SRC1_WINS
= 0x2,
279 BIFROST_SRC0_WINS
= 0x3,
282 #define BIFROST_FMA_OP_FADD32 (0x58 >> 2)
283 #define BIFROST_FMA_OP_FMAX32 (0x40 >> 2)
284 #define BIFROST_FMA_OP_FMIN32 (0x44 >> 2)
286 struct bifrost_fma_add
{
289 unsigned src1_abs
: 1;
290 unsigned src0_neg
: 1;
291 unsigned src1_neg
: 1;
293 unsigned src0_abs
: 1;
294 enum bifrost_roundmode roundmode
: 2;
295 enum bifrost_outmod outmod
: 2;
297 } __attribute__((packed
));
299 #define BIFROST_FMA_OP_FMAX16 (0xC0 >> 2)
300 #define BIFROST_FMA_OP_FMIN16 (0xCC >> 2)
301 #define BIFROST_FMA_OP_FADD16 (0xD8 >> 2)
303 struct bifrost_fma_add_minmax16
{
306 /* abs2 inferred as (src1 < src0) */
308 unsigned src0_neg
: 1;
309 unsigned src1_neg
: 1;
310 unsigned src0_swizzle
: 2;
311 unsigned src1_swizzle
: 2;
313 enum bifrost_outmod outmod
: 2;
314 /* roundmode for add, min/max mode for min/max */
316 } __attribute__((packed
));
318 #define BIFROST_FMA_OP_FMA (0x00)
320 struct bifrost_fma_fma
{
324 unsigned src_expand
: 3;
325 unsigned src0_abs
: 1;
326 enum bifrost_roundmode roundmode
: 2;
327 enum bifrost_outmod outmod
: 2;
328 unsigned src0_neg
: 1; /* 14 */
329 unsigned src2_neg
: 1;
330 unsigned src1_abs
: 1;
331 unsigned src2_abs
: 1; /* 17 */
333 } __attribute__((packed
));
335 #define BIFROST_FMA_OP_FMA16 (0x2)
337 struct bifrost_fma_fma16
{
341 unsigned swizzle_0
: 2;
342 unsigned swizzle_1
: 2;
343 enum bifrost_roundmode roundmode
: 2;
344 enum bifrost_outmod outmod
: 2;
345 unsigned src0_neg
: 1;
346 unsigned src2_neg
: 1;
347 unsigned swizzle_2
: 2;
349 } __attribute__((packed
));
351 enum bifrost_csel_cond
{
362 #define BIFROST_FMA_OP_CSEL4 (0x5c)
363 #define BIFROST_FMA_OP_CSEL4_V16 (0xdc)
365 struct bifrost_csel4
{
370 enum bifrost_csel_cond cond
: 3;
372 } __attribute__((packed
));
374 #define BIFROST_FMA_OP_RSHIFT_NAND (0x60000 >> 12)
375 #define BIFROST_FMA_OP_RSHIFT_AND (0x61000 >> 12)
376 #define BIFROST_FMA_OP_LSHIFT_NAND (0x62000 >> 12)
377 #define BIFROST_FMA_OP_LSHIFT_AND (0x63000 >> 12)
378 #define BIFROST_FMA_OP_RSHIFT_XOR (0x64000 >> 12)
379 #define BIFROST_FMA_OP_LSHIFT_ADD_32 (0x65200 >> 6)
380 #define BIFROST_FMA_OP_LSHIFT_SUB_32 (0x65600 >> 6)
381 #define BIFROST_FMA_OP_LSHIFT_RSUB_32 (0x65a00 >> 6)
382 #define BIFROST_FMA_OP_RSHIFT_ADD_32 (0x65e00 >> 6)
383 #define BIFROST_FMA_OP_RSHIFT_SUB_32 (0x66200 >> 6)
384 #define BIFROST_FMA_OP_RSHIFT_RSUB_32 (0x66600 >> 6)
386 struct bifrost_shift_fma
{
391 unsigned unk
: 1; /* always set? */
392 unsigned invert_1
: 1; /* Inverts sources to combining op */
393 /* For XOR, switches RSHIFT to LSHIFT since only one invert needed */
394 unsigned invert_2
: 1;
396 } __attribute__((packed
));
398 struct bifrost_shift_add
{
404 unsigned invert_1
: 1;
405 unsigned invert_2
: 1;
408 } __attribute__((packed
));
410 enum bifrost_fcmp_cond
{
419 #define BIFROST_FMA_OP_FCMP_GL (0x48000 >> 13)
420 #define BIFROST_FMA_OP_FCMP_D3D (0x4c000 >> 13)
422 struct bifrost_fma_fcmp
{
425 unsigned src1_abs
: 1;
427 unsigned src1_neg
: 1;
428 unsigned src_expand
: 3;
429 unsigned src0_abs
: 1;
430 enum bifrost_fcmp_cond cond
: 3;
432 } __attribute__((packed
));
434 struct bifrost_add_fcmp
{
437 enum bifrost_fcmp_cond cond
: 3;
438 unsigned src_expand
: 2;
439 unsigned src0_abs
: 1;
440 unsigned src1_abs
: 1;
441 unsigned src1_neg
: 1;
443 } __attribute__((packed
));
445 #define BIFROST_FMA_OP_FCMP_GL_16 (0xc8000 >> 13)
446 #define BIFROST_FMA_OP_FCMP_D3D_16 (0xcc000 >> 13)
448 struct bifrost_fma_fcmp16
{
456 unsigned src0_swizzle
: 2;
457 unsigned src1_swizzle
: 2;
459 enum bifrost_fcmp_cond cond
: 3;
461 } __attribute__((packed
));
463 struct bifrost_add_fcmp16
{
466 enum bifrost_fcmp_cond cond
: 3;
468 unsigned src0_swizzle
: 2;
469 unsigned src1_swizzle
: 2;
472 unsigned src0_neg
: 1;
475 } __attribute__((packed
));
477 enum bifrost_icmp_cond
{
478 BIFROST_ICMP_IGT
= 0,
479 BIFROST_ICMP_IGE
= 1,
480 BIFROST_ICMP_UGT
= 2,
481 BIFROST_ICMP_UGE
= 3,
483 BIFROST_ICMP_NEQ
= 5,
486 struct bifrost_fma_icmp32
{
489 enum bifrost_icmp_cond cond
: 3;
490 unsigned unk1
: 1; /* set */
493 } __attribute__((packed
));
495 struct bifrost_fma_icmp16
{
498 unsigned unk
: 5; /* 11010 */
499 enum bifrost_icmp_cond cond
: 3;
501 } __attribute__((packed
));
503 struct bifrost_add_icmp
{
506 enum bifrost_icmp_cond cond
: 3;
507 unsigned sz
: 1; /* 1 for 32, 0 for 8 */
510 } __attribute__((packed
));
512 /* Two sources for vectorization */
513 #define BIFROST_FMA_FLOAT32_TO_16 (0xdd000 >> 3)
514 #define BIFROST_ADD_FLOAT32_TO_16 (0x0EC00 >> 3)
516 enum bifrost_convert_mode
{
517 BIFROST_CONV_UNK0
= 0,
518 BIFROST_CONV_F32_TO_I32
= 1,
519 BIFROST_CONV_F16_TO_I16
= 2,
520 BIFROST_CONV_I32_TO_F32
= 3,
521 BIFROST_CONV_I16_TO_X32
= 4,
522 BIFROST_CONV_F16_TO_F32
= 5,
523 BIFROST_CONV_I16_TO_F16
= 6,
524 BIFROST_CONV_UNK7
= 7
528 #define BIFROST_CONVERT_4(is_unsigned, component, to_float) \
529 ((is_unsigned & 1) | ((component & 1) << 1) | ((to_float & 1) << 2) | \
530 ((0x3) << 3) | ((4) << 5) | 0x100)
533 #define BIFROST_CONVERT_5(component) \
534 ((component & 1) | ((1) << 1) | ((5) << 5) | 0x100)
536 /* Other conversions */
537 #define BIFROST_CONVERT(is_unsigned, roundmode, swizzle, mode) \
538 ((is_unsigned & 1) | ((roundmode & 3) << 1) | ((swizzle & 3) << 3) | ((mode & 7) << 5))
540 #define BIFROST_FMA_CONVERT (0xe0000)
541 #define BIFROST_ADD_CONVERT (0x07800)
543 enum bifrost_ldst_type
{
544 BIFROST_LDST_F16
= 0,
545 BIFROST_LDST_F32
= 1,
546 BIFROST_LDST_I32
= 2,
550 #define BIFROST_ADD_OP_LD_VAR_ADDR (0x18000 >> 10)
552 struct bifrost_ld_var_addr
{
555 unsigned location
: 5;
556 enum bifrost_ldst_type type
: 2;
558 } __attribute__((packed
));
560 #define BIFROST_ADD_OP_LD_ATTR (0x08000 >> 12)
562 struct bifrost_ld_attr
{
565 unsigned location
: 5;
566 unsigned channels
: 2; /* MALI_POSITIVE */
567 enum bifrost_ldst_type type
: 2;
569 } __attribute__((packed
));
571 enum bifrost_interp_mode
{
572 BIFROST_INTERP_PER_FRAG
= 0x0,
573 BIFROST_INTERP_CENTROID
= 0x1,
574 BIFROST_INTERP_DEFAULT
= 0x2,
575 BIFROST_INTERP_EXPLICIT
= 0x3
578 #define BIFROST_ADD_OP_LD_VAR_16 (0x1a << 1)
579 #define BIFROST_ADD_OP_LD_VAR_32 (0x0a << 1)
581 struct bifrost_ld_var
{
584 /* If top two bits set, indirect with src in bottom three */
587 unsigned channels
: 2; /* MALI_POSITIVE */
588 enum bifrost_interp_mode interp_mode
: 2;
592 } __attribute__((packed
));
594 struct bifrost_tex_ctrl
{
595 unsigned sampler_index
: 4; // also used to signal indirects
596 unsigned tex_index
: 7;
597 bool no_merge_index
: 1; // whether to merge (direct) sampler & texture indices
598 bool filter
: 1; // use the usual filtering pipeline (0 for texelFetch & textureGather)
600 bool texel_offset
: 1; // *Offset()
603 unsigned tex_type
: 2; // 2D, 3D, Cube, Buffer
604 bool compute_lod
: 1; // 0 for *Lod()
605 bool not_supply_lod
: 1; // 0 for *Lod() or when a bias is applied
606 bool calc_gradients
: 1; // 0 for *Grad()
608 unsigned result_type
: 4; // integer, unsigned, float TODO: why is this 4 bits?
610 } __attribute__((packed
));
612 struct bifrost_dual_tex_ctrl
{
613 unsigned sampler_index0
: 2;
615 unsigned tex_index0
: 2;
616 unsigned sampler_index1
: 2;
617 unsigned tex_index1
: 2;
619 } __attribute__((packed
));
621 #define BIFROST_ADD_OP_TEX_COMPACT_F32 (0x0b000 >> 10)
622 #define BIFROST_ADD_OP_TEX_COMPACT_F16 (0x1b000 >> 10)
624 struct bifrost_tex_compact
{
627 unsigned tex_index
: 3;
628 unsigned unknown
: 1;
629 unsigned sampler_index
: 3;
631 } __attribute__((packed
));
633 enum branch_bit_size
{
637 // For the above combinations of bitsize and location, an extra bit is
638 // encoded via comparing the sources. The only possible source of ambiguity
639 // would be if the sources were the same, but then the branch condition
640 // would be always true or always false anyways, so we can ignore it. But
641 // this no longer works when comparing the y component to the x component,
642 // since it's valid to compare the y component of a source against its own
643 // x component. Instead, the extra bit is encoded via an extra bitsize.
646 BR_SIZE_32_AND_16X
= 5,
647 BR_SIZE_32_AND_16Y
= 6,
648 // Used for comparisons with zero and always-true, see below. I think this
649 // only works for integer comparisons.
653 enum bifrost_reg_write_unit
{
654 REG_WRITE_NONE
= 0, // don't write
655 REG_WRITE_TWO
, // write using reg2
656 REG_WRITE_THREE
, // write using reg3
659 struct bifrost_regs
{
660 unsigned uniform_const
: 8;
666 } __attribute__((packed
));
668 enum bifrost_branch_cond
{
673 // Equal vs. not-equal determined by src0/src1 comparison
675 // floating-point comparisons
676 // Becomes UNE when you flip the arguments
678 // TODO what happens when you flip the arguments?
683 enum bifrost_branch_code
{
687 struct bifrost_branch
{
690 /* For BR_SIZE_ZERO, upper two bits become ctrl */
693 /* Offset source -- always uniform/const but
694 * theoretically could support indirect jumps? */
697 enum bifrost_branch_cond cond
: 3;
698 enum branch_bit_size size
: 3;
705 #define BIFROST_FMA_NOP (0x701960 | BIFROST_SRC_STAGE)
706 #define BIFROST_ADD_NOP (0x3D960 | BIFROST_SRC_STAGE)
708 struct bifrost_fmt1
{
713 uint64_t header
: 45;
714 } __attribute__((packed
));
716 #define BIFROST_FMT1_INSTRUCTIONS 0b00101
717 #define BIFROST_FMT1_FINAL 0b01001
718 #define BIFROST_FMT1_CONSTANTS 0b00001
720 #define BIFROST_FMTC_CONSTANTS 0b0011
721 #define BIFROST_FMTC_FINAL 0b0111
723 struct bifrost_fmt_constant
{
728 } __attribute__((packed
));
730 enum bifrost_reg_control
{
731 BIFROST_WRITE_FMA_P2
= 1,
732 BIFROST_WRITE_FMA_P2_READ_P3
= 2,
733 BIFROST_FIRST_WRITE_FMA_P2_READ_P3
= 3,
735 BIFROST_WRITE_ADD_P2
= 5,
736 BIFROST_WRITE_ADD_P2_READ_P3
= 6,
737 BIFROST_WRITE_ADD_P2_FMA_P3
= 7,
739 BIFROST_FIRST_NONE
= 8,
740 BIFROST_FIRST_WRITE_FMA_P2
= 9,
741 /* INSTR_INVALID_ENC */
742 BIFROST_REG_NONE
= 11,
743 BIFROST_FIRST_READ_P3
= 12,
744 BIFROST_FIRST_WRITE_ADD_P2
= 13,
745 BIFROST_FIRST_WRITE_ADD_P2_READ_P3
= 14,
746 BIFROST_FIRST_WRITE_ADD_P2_FMA_P3
= 15