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)
103 struct bifrost_fma_inst
{
106 } __attribute__((packed
));
108 struct bifrost_fma_2src
{
112 } __attribute__((packed
));
114 #define BIFROST_FMA_OP_MSCALE (0x50 >> 3)
116 struct bifrost_fma_mscale
{
122 /* If mscale_mode is set - an MSCALE specific mode. If it is not set, a
125 unsigned mscale_mode
: 1;
127 unsigned src0_abs
: 1;
128 unsigned src1_neg
: 1;
129 unsigned src2_neg
: 1;
131 } __attribute__((packed
));
133 #define BIFROST_ADD_OP_BLEND (0x1952c)
134 #define BIFROST_ADD_OP_FRCP_FAST_F32 (0x0cc00)
135 #define BIFROST_ADD_OP_FRCP_FAST_F16_X (0x0ce10)
136 #define BIFROST_ADD_OP_FRCP_FAST_F16_Y (0x0ce30)
137 #define BIFROST_ADD_OP_FRSQ_FAST_F32 (0x0cc20)
138 #define BIFROST_ADD_OP_FRSQ_FAST_F16_X (0x0ce50)
139 #define BIFROST_ADD_OP_FRSQ_FAST_F16_Y (0x0ce70)
140 #define BIFROST_ADD_OP_LOG2_HELP (0x0cc68)
141 #define BIFROST_ADD_OP_FEXP2_FAST (0x0cd58)
143 struct bifrost_add_inst
{
146 } __attribute__((packed
));
148 #define BIFROST_ADD_OP_LD_UBO_1 (0x0c1a0 >> 3)
149 #define BIFROST_ADD_OP_LD_UBO_2 (0x0c1e0 >> 3)
150 #define BIFROST_ADD_OP_LD_UBO_3 (0x0caa0 >> 3)
151 #define BIFROST_ADD_OP_LD_UBO_4 (0x0c220 >> 3)
153 struct bifrost_add_2src
{
157 } __attribute__((packed
));
159 #define BIFROST_ADD_OP_FMAX32 (0x00)
160 #define BIFROST_ADD_OP_FMIN32 (0x01)
161 #define BIFROST_ADD_OP_FADD32 (0x02)
163 #define BIFROST_ADD_OP_FADD16 (0x0A)
165 struct bifrost_add_faddmin
{
168 unsigned src1_abs
: 1;
169 unsigned src0_neg
: 1;
170 unsigned src1_neg
: 1;
171 unsigned select
: 2; /* swizzle_0 for fp16 */
172 unsigned outmod
: 2; /* swizzle_1 for fp16 */
174 unsigned src0_abs
: 1;
176 } __attribute__((packed
));
178 #define BIFROST_ADD_OP_FMAX16 (0x10)
179 #define BIFROST_ADD_OP_FMIN16 (0x12)
181 struct bifrost_add_fmin16
{
184 /* abs2 inferred as with FMA */
186 unsigned src0_neg
: 1;
187 unsigned src1_neg
: 1;
188 unsigned src0_swizzle
: 2;
189 unsigned src1_swizzle
: 2;
192 } __attribute__((packed
));
194 #define BIFROST_ADD_OP_ST_VAR (0x19300 >> 8)
196 struct bifrost_st_vary
{
200 unsigned channels
: 2;
202 } __attribute__((packed
));
204 #define BIFROST_ADD_OP_ATEST (0xc8f)
206 struct bifrost_add_atest
{
207 /* gl_SampleMask (R60) */
213 /* If half, X/Y select. If !half, always set */
214 unsigned component
: 1;
218 } __attribute__((packed
));
220 enum bifrost_outmod
{
223 BIFROST_SAT_SIGNED
= 0x2,
227 enum bifrost_roundmode
{
228 BIFROST_RTE
= 0x0, /* round to even */
229 BIFROST_RTP
= 0x1, /* round to positive */
230 BIFROST_RTN
= 0x2, /* round to negative */
231 BIFROST_RTZ
= 0x3 /* round to zero */
234 /* NONE: Same as fmax() and fmin() -- return the other
235 * number if any number is NaN. Also always return +0 if
236 * one argument is +0 and the other is -0.
238 * NAN_WINS: Instead of never returning a NaN, always return
239 * one. The "greater"/"lesser" NaN is always returned, first
240 * by checking the sign and then the mantissa bits.
242 * SRC1_WINS: For max, implement src0 > src1 ? src0 : src1.
243 * For min, implement src0 < src1 ? src0 : src1. This
244 * includes handling NaN's and signedness of 0 differently
245 * from above, since +0 and -0 compare equal and comparisons
246 * always return false for NaN's. As a result, this mode is
249 * SRC0_WINS: For max, implement src0 < src1 ? src1 : src0
250 * For min, implement src0 > src1 ? src1 : src0
254 enum bifrost_minmax_mode
{
255 BIFROST_MINMAX_NONE
= 0x0,
256 BIFROST_NAN_WINS
= 0x1,
257 BIFROST_SRC1_WINS
= 0x2,
258 BIFROST_SRC0_WINS
= 0x3,
261 #define BIFROST_FMA_OP_FADD32 (0x58 >> 2)
262 #define BIFROST_FMA_OP_FMAX32 (0x40 >> 2)
263 #define BIFROST_FMA_OP_FMIN32 (0x44 >> 2)
265 struct bifrost_fma_add
{
268 unsigned src1_abs
: 1;
269 unsigned src0_neg
: 1;
270 unsigned src1_neg
: 1;
272 unsigned src0_abs
: 1;
273 enum bifrost_roundmode roundmode
: 2;
274 enum bifrost_outmod outmod
: 2;
276 } __attribute__((packed
));
278 #define BIFROST_FMA_OP_FMAX16 (0xC0 >> 2)
279 #define BIFROST_FMA_OP_FMIN16 (0xCC >> 2)
280 #define BIFROST_FMA_OP_FADD16 (0xD8 >> 2)
282 struct bifrost_fma_add_minmax16
{
285 /* abs2 inferred as (src1 < src0) */
287 unsigned src0_neg
: 1;
288 unsigned src1_neg
: 1;
289 unsigned src0_swizzle
: 2;
290 unsigned src1_swizzle
: 2;
292 enum bifrost_outmod outmod
: 2;
293 /* roundmode for add, min/max mode for min/max */
295 } __attribute__((packed
));
297 #define BIFROST_FMA_OP_FMA (0x00)
299 struct bifrost_fma_fma
{
303 unsigned src_expand
: 3;
304 unsigned src0_abs
: 1;
305 enum bifrost_roundmode roundmode
: 2;
306 enum bifrost_outmod outmod
: 2;
307 unsigned src0_neg
: 1; /* 14 */
308 unsigned src2_neg
: 1;
309 unsigned src1_abs
: 1;
310 unsigned src2_abs
: 1; /* 17 */
312 } __attribute__((packed
));
314 #define BIFROST_FMA_OP_FMA16 (0x2)
316 struct bifrost_fma_fma16
{
320 unsigned swizzle_0
: 2;
321 unsigned swizzle_1
: 2;
322 enum bifrost_roundmode roundmode
: 2;
323 enum bifrost_outmod outmod
: 2;
324 unsigned src0_neg
: 1;
325 unsigned src2_neg
: 1;
326 unsigned swizzle_2
: 2;
328 } __attribute__((packed
));
330 enum bifrost_csel_cond
{
341 #define BIFROST_FMA_OP_CSEL4 (0x5c)
342 #define BIFROST_FMA_OP_CSEL4_V16 (0xdc)
344 struct bifrost_csel4
{
349 enum bifrost_csel_cond cond
: 3;
351 } __attribute__((packed
));
353 struct bifrost_shift_fma
{
357 unsigned half
: 3; /* 000 for i32, 100 for i8, 111 for v2i16 */
358 unsigned unk
: 1; /* always set? */
359 unsigned invert_1
: 1; /* Inverts sources to combining op */
360 /* For XOR, switches RSHIFT to LSHIFT since only one invert needed */
361 unsigned invert_2
: 1;
363 } __attribute__((packed
));
365 struct bifrost_shift_add
{
371 unsigned invert_1
: 1;
372 unsigned invert_2
: 1;
375 } __attribute__((packed
));
377 /* Two sources for vectorization */
378 #define BIFROST_FMA_FLOAT32_TO_16 (0xdd000 >> 3)
379 #define BIFROST_ADD_FLOAT32_TO_16 (0x0EC00 >> 3)
381 enum bifrost_convert_mode
{
382 BIFROST_CONV_UNK0
= 0,
383 BIFROST_CONV_F32_TO_I32
= 1,
384 BIFROST_CONV_F16_TO_I16
= 2,
385 BIFROST_CONV_I32_TO_F32
= 3,
386 BIFROST_CONV_I16_TO_X32
= 4,
387 BIFROST_CONV_F16_TO_F32
= 5,
388 BIFROST_CONV_I16_TO_F16
= 6,
389 BIFROST_CONV_UNK7
= 7
393 #define BIFROST_CONVERT_4(is_unsigned, component, to_float) \
394 ((is_unsigned & 1) | ((component & 1) << 1) | ((to_float & 1) << 2) | \
395 ((0x3) << 3) | ((4) << 5) | 0x100)
398 #define BIFROST_CONVERT_5(component) \
399 ((component & 1) | ((1) << 1) | ((5) << 5) | 0x100)
401 /* Other conversions */
402 #define BIFROST_CONVERT(is_unsigned, roundmode, swizzle, mode) \
403 ((is_unsigned & 1) | ((roundmode & 3) << 1) | ((swizzle & 3) << 3) | ((mode & 7) << 5))
405 #define BIFROST_FMA_CONVERT (0xe0000)
406 #define BIFROST_ADD_CONVERT (0x07800)
408 enum bifrost_ldst_type
{
409 BIFROST_LDST_F16
= 0,
410 BIFROST_LDST_F32
= 1,
411 BIFROST_LDST_I32
= 2,
415 #define BIFROST_ADD_OP_LD_VAR_ADDR (0x18000 >> 10)
417 struct bifrost_ld_var_addr
{
420 unsigned location
: 5;
421 enum bifrost_ldst_type type
: 2;
423 } __attribute__((packed
));
425 #define BIFROST_ADD_OP_LD_ATTR (0x08000 >> 12)
427 struct bifrost_ld_attr
{
430 unsigned location
: 5;
431 unsigned channels
: 2; /* MALI_POSITIVE */
432 enum bifrost_ldst_type type
: 2;
434 } __attribute__((packed
));
436 enum bifrost_interp_mode
{
437 BIFROST_INTERP_PER_FRAG
= 0x0,
438 BIFROST_INTERP_CENTROID
= 0x1,
439 BIFROST_INTERP_DEFAULT
= 0x2,
440 BIFROST_INTERP_EXPLICIT
= 0x3
443 #define BIFROST_ADD_OP_LD_VAR_16 (0x1a << 1)
444 #define BIFROST_ADD_OP_LD_VAR_32 (0x0a << 1)
446 struct bifrost_ld_var
{
449 /* If top two bits set, indirect with src in bottom three */
452 unsigned channels
: 2; /* MALI_POSITIVE */
453 enum bifrost_interp_mode interp_mode
: 2;
457 } __attribute__((packed
));
459 struct bifrost_tex_ctrl
{
460 unsigned sampler_index
: 4; // also used to signal indirects
461 unsigned tex_index
: 7;
462 bool no_merge_index
: 1; // whether to merge (direct) sampler & texture indices
463 bool filter
: 1; // use the usual filtering pipeline (0 for texelFetch & textureGather)
465 bool texel_offset
: 1; // *Offset()
468 unsigned tex_type
: 2; // 2D, 3D, Cube, Buffer
469 bool compute_lod
: 1; // 0 for *Lod()
470 bool not_supply_lod
: 1; // 0 for *Lod() or when a bias is applied
471 bool calc_gradients
: 1; // 0 for *Grad()
473 unsigned result_type
: 4; // integer, unsigned, float TODO: why is this 4 bits?
475 } __attribute__((packed
));
477 struct bifrost_dual_tex_ctrl
{
478 unsigned sampler_index0
: 2;
480 unsigned tex_index0
: 2;
481 unsigned sampler_index1
: 2;
482 unsigned tex_index1
: 2;
484 } __attribute__((packed
));
486 #define BIFROST_ADD_OP_TEX_COMPACT_F32 (0x0b000 >> 10)
487 #define BIFROST_ADD_OP_TEX_COMPACT_F16 (0x1b000 >> 10)
489 struct bifrost_tex_compact
{
492 unsigned tex_index
: 3;
493 unsigned unknown
: 1;
494 unsigned sampler_index
: 3;
496 } __attribute__((packed
));
498 enum branch_bit_size
{
502 // For the above combinations of bitsize and location, an extra bit is
503 // encoded via comparing the sources. The only possible source of ambiguity
504 // would be if the sources were the same, but then the branch condition
505 // would be always true or always false anyways, so we can ignore it. But
506 // this no longer works when comparing the y component to the x component,
507 // since it's valid to compare the y component of a source against its own
508 // x component. Instead, the extra bit is encoded via an extra bitsize.
511 BR_SIZE_32_AND_16X
= 5,
512 BR_SIZE_32_AND_16Y
= 6,
513 // Used for comparisons with zero and always-true, see below. I think this
514 // only works for integer comparisons.
518 enum bifrost_reg_write_unit
{
519 REG_WRITE_NONE
= 0, // don't write
520 REG_WRITE_TWO
, // write using reg2
521 REG_WRITE_THREE
, // write using reg3
524 struct bifrost_regs
{
525 unsigned uniform_const
: 8;
531 } __attribute__((packed
));
533 enum bifrost_branch_cond
{
538 // Equal vs. not-equal determined by src0/src1 comparison
540 // floating-point comparisons
541 // Becomes UNE when you flip the arguments
543 // TODO what happens when you flip the arguments?
548 enum bifrost_branch_code
{
552 struct bifrost_branch
{
555 /* For BR_SIZE_ZERO, upper two bits become ctrl */
558 /* Offset source -- always uniform/const but
559 * theoretically could support indirect jumps? */
562 enum bifrost_branch_cond cond
: 3;
563 enum branch_bit_size size
: 3;
570 #define BIFROST_FMA_NOP (0x701960 | BIFROST_SRC_STAGE)
571 #define BIFROST_ADD_NOP (0x3D960 | BIFROST_SRC_STAGE)
573 struct bifrost_fmt1
{
578 uint64_t header
: 45;
579 } __attribute__((packed
));
581 #define BIFROST_FMT1_INSTRUCTIONS 0b00101
582 #define BIFROST_FMT1_FINAL 0b01001
583 #define BIFROST_FMT1_CONSTANTS 0b00001
585 #define BIFROST_FMTC_CONSTANTS 0b0011
586 #define BIFROST_FMTC_FINAL 0b0111
588 struct bifrost_fmt_constant
{
593 } __attribute__((packed
));
595 enum bifrost_reg_control
{
596 BIFROST_WRITE_FMA_P2
= 1,
597 BIFROST_WRITE_FMA_P2_READ_P3
= 2,
598 BIFROST_FIRST_WRITE_FMA_P2_READ_P3
= 3,
600 BIFROST_WRITE_ADD_P2
= 5,
601 BIFROST_WRITE_ADD_P2_READ_P3
= 6,
602 BIFROST_WRITE_ADD_P2_FMA_P3
= 7,
604 BIFROST_FIRST_NONE
= 8,
605 BIFROST_FIRST_WRITE_FMA_P2
= 9,
606 /* INSTR_INVALID_ENC */
607 BIFROST_REG_NONE
= 11,
608 BIFROST_FIRST_READ_P3
= 12,
609 BIFROST_FIRST_WRITE_ADD_P2
= 13,
610 BIFROST_FIRST_WRITE_ADD_P2_READ_P3
= 14,
611 BIFROST_FIRST_WRITE_ADD_P2_FMA_P3
= 15