pan/bi: Structify FMA_MSCALE
[mesa.git] / src / panfrost / bifrost / bifrost.h
1 /*
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>
5 *
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:
12 *
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
15 * Software.
16 *
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
23 * SOFTWARE.
24 */
25
26 #ifndef __bifrost_h__
27 #define __bifrost_h__
28
29 #include <stdint.h>
30 #include <stdbool.h>
31
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
42 };
43
44 struct bifrost_header {
45 unsigned unk0 : 7;
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;
51 unsigned unk1 : 2;
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;
56 unsigned unk2 : 2;
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:
59 //
60 // "Stores to image and buffer variables performed by helper invocations
61 // have no effect on the underlying image or buffer memory."
62 //
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
76 // previous clause.
77 unsigned datareg_writebarrier: 1;
78 unsigned datareg : 6;
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));
86
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,
96 };
97
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
103 struct bifrost_fma_inst {
104 unsigned src0 : 3;
105 unsigned op : 20;
106 } __attribute__((packed));
107
108 struct bifrost_fma_2src {
109 unsigned src0 : 3;
110 unsigned src1 : 3;
111 unsigned op : 17;
112 } __attribute__((packed));
113
114 #define BIFROST_FMA_OP_MSCALE (0x50 >> 3)
115
116 struct bifrost_fma_mscale {
117 unsigned src0 : 3;
118 unsigned src1 : 3;
119 unsigned src2 : 3;
120 unsigned src3 : 3;
121
122 /* If mscale_mode is set - an MSCALE specific mode. If it is not set, a
123 * regular outmod */
124 unsigned mode : 2;
125 unsigned mscale_mode : 1;
126
127 unsigned src0_abs : 1;
128 unsigned src1_neg : 1;
129 unsigned src2_neg : 1;
130 unsigned op : 5;
131 } __attribute__((packed));
132
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
142 struct bifrost_add_inst {
143 unsigned src0 : 3;
144 unsigned op : 17;
145 } __attribute__((packed));
146
147 #define BIFROST_ADD_OP_LD_UBO_1 (0x0c1a0 >> 3)
148 #define BIFROST_ADD_OP_LD_UBO_2 (0x0c1e0 >> 3)
149 #define BIFROST_ADD_OP_LD_UBO_3 (0x0caa0 >> 3)
150 #define BIFROST_ADD_OP_LD_UBO_4 (0x0c220 >> 3)
151
152 struct bifrost_add_2src {
153 unsigned src0 : 3;
154 unsigned src1 : 3;
155 unsigned op : 14;
156 } __attribute__((packed));
157
158 #define BIFROST_ADD_OP_FMAX32 (0x00)
159 #define BIFROST_ADD_OP_FMIN32 (0x01)
160 #define BIFROST_ADD_OP_FADD32 (0x02)
161
162 struct bifrost_add_faddmin {
163 unsigned src0 : 3;
164 unsigned src1 : 3;
165 unsigned src1_abs : 1;
166 unsigned src0_neg : 1;
167 unsigned src1_neg : 1;
168 unsigned select : 2;
169 unsigned outmod : 2;
170 unsigned mode : 2;
171 unsigned src0_abs : 1;
172 unsigned op : 4;
173 } __attribute__((packed));
174
175 #define BIFROST_ADD_OP_ST_VAR (0x19300 >> 8)
176
177 struct bifrost_st_vary {
178 unsigned src0 : 3;
179 unsigned src1 : 3;
180 unsigned src2 : 3;
181 unsigned channels : 2;
182 unsigned op : 9;
183 } __attribute__((packed));
184
185 #define BIFROST_ADD_OP_ATEST (0xc8f)
186
187 struct bifrost_add_atest {
188 /* gl_SampleMask (R60) */
189 unsigned src0 : 3;
190
191 /* Alpha value */
192 unsigned src1 : 3;
193
194 /* If half, X/Y select. If !half, always set */
195 unsigned component : 1;
196 unsigned half : 1;
197
198 unsigned op : 12;
199 } __attribute__((packed));
200
201 enum bifrost_outmod {
202 BIFROST_NONE = 0x0,
203 BIFROST_POS = 0x1,
204 BIFROST_SAT_SIGNED = 0x2,
205 BIFROST_SAT = 0x3,
206 };
207
208 enum bifrost_roundmode {
209 BIFROST_RTE = 0x0, /* round to even */
210 BIFROST_RTP = 0x1, /* round to positive */
211 BIFROST_RTN = 0x2, /* round to negative */
212 BIFROST_RTZ = 0x3 /* round to zero */
213 };
214
215 /* NONE: Same as fmax() and fmin() -- return the other
216 * number if any number is NaN. Also always return +0 if
217 * one argument is +0 and the other is -0.
218 *
219 * NAN_WINS: Instead of never returning a NaN, always return
220 * one. The "greater"/"lesser" NaN is always returned, first
221 * by checking the sign and then the mantissa bits.
222 *
223 * SRC1_WINS: For max, implement src0 > src1 ? src0 : src1.
224 * For min, implement src0 < src1 ? src0 : src1. This
225 * includes handling NaN's and signedness of 0 differently
226 * from above, since +0 and -0 compare equal and comparisons
227 * always return false for NaN's. As a result, this mode is
228 * *not* commutative.
229 *
230 * SRC0_WINS: For max, implement src0 < src1 ? src1 : src0
231 * For min, implement src0 > src1 ? src1 : src0
232 */
233
234
235 enum bifrost_minmax_mode {
236 BIFROST_MINMAX_NONE = 0x0,
237 BIFROST_NAN_WINS = 0x1,
238 BIFROST_SRC1_WINS = 0x2,
239 BIFROST_SRC0_WINS = 0x3,
240 };
241
242 #define BIFROST_FMA_OP_FADD32 (0x58 >> 2)
243 #define BIFROST_FMA_OP_FMAX32 (0x40 >> 2)
244 #define BIFROST_FMA_OP_FMIN32 (0x44 >> 2)
245
246 struct bifrost_fma_add {
247 unsigned src0 : 3;
248 unsigned src1 : 3;
249 unsigned src1_abs : 1;
250 unsigned src0_neg : 1;
251 unsigned src1_neg : 1;
252 unsigned unk : 3;
253 unsigned src0_abs : 1;
254 enum bifrost_roundmode roundmode : 2;
255 enum bifrost_outmod outmod : 2;
256 unsigned op : 6;
257 } __attribute__((packed));
258
259 #define BIFROST_FMA_OP_FMAX16 (0xC0 >> 2)
260 #define BIFROST_FMA_OP_FMIN16 (0xCC >> 2)
261 #define BIFROST_FMA_OP_FADD16 (0xD8 >> 2)
262
263 struct bifrost_fma_add_minmax16 {
264 unsigned src0 : 3;
265 unsigned src1 : 3;
266 /* abs2 inferred as (src1 < src0) */
267 unsigned abs1 : 1;
268 unsigned src0_neg : 1;
269 unsigned src1_neg : 1;
270 unsigned src0_swizzle : 2;
271 unsigned src1_swizzle : 2;
272 unsigned mode : 2;
273 enum bifrost_outmod outmod : 2;
274 /* roundmode for add, min/max mode for min/max */
275 unsigned op : 6;
276 } __attribute__((packed));
277
278 #define BIFROST_FMA_OP_FMA (0x00)
279
280 struct bifrost_fma_fma {
281 unsigned src0 : 3;
282 unsigned src1 : 3;
283 unsigned src2 : 3;
284 unsigned src_expand : 3;
285 unsigned src0_abs : 1;
286 enum bifrost_roundmode roundmode : 2;
287 enum bifrost_outmod outmod : 2;
288 unsigned src0_neg : 1; /* 14 */
289 unsigned src2_neg : 1;
290 unsigned src1_abs : 1;
291 unsigned src2_abs : 1; /* 17 */
292 unsigned op : 2;
293 } __attribute__((packed));
294
295 #define BIFROST_FMA_OP_FMA16 (0x2)
296
297 struct bifrost_fma_fma16 {
298 unsigned src0 : 3;
299 unsigned src1 : 3;
300 unsigned src2 : 3;
301 unsigned swizzle_0 : 2;
302 unsigned swizzle_1 : 2;
303 enum bifrost_roundmode roundmode : 2;
304 enum bifrost_outmod outmod : 2;
305 unsigned src0_neg : 1;
306 unsigned src2_neg : 1;
307 unsigned swizzle_2 : 2;
308 unsigned op : 2;
309 } __attribute__((packed));
310
311 enum bifrost_csel_cond {
312 BIFROST_FEQ_F = 0x0,
313 BIFROST_FGT_F = 0x1,
314 BIFROST_FGE_F = 0x2,
315 BIFROST_IEQ_F = 0x3,
316 BIFROST_IGT_I = 0x4,
317 BIFROST_IGE_I = 0x5,
318 BIFROST_UGT_I = 0x6,
319 BIFROST_UGE_I = 0x7
320 };
321
322 #define BIFROST_FMA_OP_CSEL4 (0x5c)
323 #define BIFROST_FMA_OP_CSEL4_V16 (0xdc)
324
325 struct bifrost_csel4 {
326 unsigned src0 : 3;
327 unsigned src1 : 3;
328 unsigned src2 : 3;
329 unsigned src3 : 3;
330 enum bifrost_csel_cond cond : 3;
331 unsigned op : 8;
332 } __attribute__((packed));
333
334 struct bifrost_shift_fma {
335 unsigned src0 : 3;
336 unsigned src1 : 3;
337 unsigned src2 : 3;
338 unsigned half : 3; /* 000 for i32, 100 for i8, 111 for v2i16 */
339 unsigned unk : 1; /* always set? */
340 unsigned invert_1 : 1; /* Inverts sources to combining op */
341 /* For XOR, switches RSHIFT to LSHIFT since only one invert needed */
342 unsigned invert_2 : 1;
343 unsigned op : 8;
344 } __attribute__((packed));
345
346 struct bifrost_shift_add {
347 unsigned src0 : 3;
348 unsigned src1 : 3;
349 unsigned src2 : 3;
350 unsigned zero : 2;
351
352 unsigned invert_1 : 1;
353 unsigned invert_2 : 1;
354
355 unsigned op : 7;
356 } __attribute__((packed));
357
358 /* Two sources for vectorization */
359 #define BIFROST_FMA_FLOAT32_TO_16 (0xdd000 >> 3)
360 #define BIFROST_ADD_FLOAT32_TO_16 (0x0EC00 >> 3)
361
362 enum bifrost_convert_mode {
363 BIFROST_CONV_UNK0 = 0,
364 BIFROST_CONV_F32_TO_I32 = 1,
365 BIFROST_CONV_F16_TO_I16 = 2,
366 BIFROST_CONV_I32_TO_F32 = 3,
367 BIFROST_CONV_I16_TO_X32 = 4,
368 BIFROST_CONV_F16_TO_F32 = 5,
369 BIFROST_CONV_I16_TO_F16 = 6,
370 BIFROST_CONV_UNK7 = 7
371 };
372
373 /* i16 to x32 */
374 #define BIFROST_CONVERT_4(is_unsigned, component, to_float) \
375 ((is_unsigned & 1) | ((component & 1) << 1) | ((to_float & 1) << 2) | \
376 ((0x3) << 3) | ((4) << 5) | 0x100)
377
378 /* f16 to f32 */
379 #define BIFROST_CONVERT_5(component) \
380 ((component & 1) | ((1) << 1) | ((5) << 5) | 0x100)
381
382 /* Other conversions */
383 #define BIFROST_CONVERT(is_unsigned, roundmode, swizzle, mode) \
384 ((is_unsigned & 1) | ((roundmode & 3) << 1) | ((swizzle & 3) << 3) | ((mode & 7) << 5))
385
386 #define BIFROST_FMA_CONVERT (0xe0000)
387 #define BIFROST_ADD_CONVERT (0x07800)
388
389 enum bifrost_ldst_type {
390 BIFROST_LDST_F16 = 0,
391 BIFROST_LDST_F32 = 1,
392 BIFROST_LDST_I32 = 2,
393 BIFROST_LDST_U32 = 3
394 };
395
396 #define BIFROST_ADD_OP_LD_VAR_ADDR (0x18000 >> 10)
397
398 struct bifrost_ld_var_addr {
399 unsigned src0 : 3;
400 unsigned src1 : 3;
401 unsigned location : 5;
402 enum bifrost_ldst_type type : 2;
403 unsigned op : 7;
404 } __attribute__((packed));
405
406 #define BIFROST_ADD_OP_LD_ATTR (0x08000 >> 12)
407
408 struct bifrost_ld_attr {
409 unsigned src0 : 3;
410 unsigned src1 : 3;
411 unsigned location : 5;
412 unsigned channels : 2; /* MALI_POSITIVE */
413 enum bifrost_ldst_type type : 2;
414 unsigned op : 5;
415 } __attribute__((packed));
416
417 enum bifrost_interp_mode {
418 BIFROST_INTERP_PER_FRAG = 0x0,
419 BIFROST_INTERP_CENTROID = 0x1,
420 BIFROST_INTERP_DEFAULT = 0x2,
421 BIFROST_INTERP_EXPLICIT = 0x3
422 };
423
424 #define BIFROST_ADD_OP_LD_VAR_16 (0x1a << 1)
425 #define BIFROST_ADD_OP_LD_VAR_32 (0x0a << 1)
426
427 struct bifrost_ld_var {
428 unsigned src0 : 3;
429
430 /* If top two bits set, indirect with src in bottom three */
431 unsigned addr : 5;
432
433 unsigned channels : 2; /* MALI_POSITIVE */
434 enum bifrost_interp_mode interp_mode : 2;
435 unsigned reuse : 1;
436 unsigned flat : 1;
437 unsigned op : 6;
438 } __attribute__((packed));
439
440 struct bifrost_tex_ctrl {
441 unsigned sampler_index : 4; // also used to signal indirects
442 unsigned tex_index : 7;
443 bool no_merge_index : 1; // whether to merge (direct) sampler & texture indices
444 bool filter : 1; // use the usual filtering pipeline (0 for texelFetch & textureGather)
445 unsigned unk0 : 2;
446 bool texel_offset : 1; // *Offset()
447 bool is_shadow : 1;
448 bool is_array : 1;
449 unsigned tex_type : 2; // 2D, 3D, Cube, Buffer
450 bool compute_lod : 1; // 0 for *Lod()
451 bool not_supply_lod : 1; // 0 for *Lod() or when a bias is applied
452 bool calc_gradients : 1; // 0 for *Grad()
453 unsigned unk1 : 1;
454 unsigned result_type : 4; // integer, unsigned, float TODO: why is this 4 bits?
455 unsigned unk2 : 4;
456 } __attribute__((packed));
457
458 struct bifrost_dual_tex_ctrl {
459 unsigned sampler_index0 : 2;
460 unsigned unk0 : 2;
461 unsigned tex_index0 : 2;
462 unsigned sampler_index1 : 2;
463 unsigned tex_index1 : 2;
464 unsigned unk1 : 22;
465 } __attribute__((packed));
466
467 enum branch_bit_size {
468 BR_SIZE_32 = 0,
469 BR_SIZE_16XX = 1,
470 BR_SIZE_16YY = 2,
471 // For the above combinations of bitsize and location, an extra bit is
472 // encoded via comparing the sources. The only possible source of ambiguity
473 // would be if the sources were the same, but then the branch condition
474 // would be always true or always false anyways, so we can ignore it. But
475 // this no longer works when comparing the y component to the x component,
476 // since it's valid to compare the y component of a source against its own
477 // x component. Instead, the extra bit is encoded via an extra bitsize.
478 BR_SIZE_16YX0 = 3,
479 BR_SIZE_16YX1 = 4,
480 BR_SIZE_32_AND_16X = 5,
481 BR_SIZE_32_AND_16Y = 6,
482 // Used for comparisons with zero and always-true, see below. I think this
483 // only works for integer comparisons.
484 BR_SIZE_ZERO = 7,
485 };
486
487 enum bifrost_reg_write_unit {
488 REG_WRITE_NONE = 0, // don't write
489 REG_WRITE_TWO, // write using reg2
490 REG_WRITE_THREE, // write using reg3
491 };
492
493 struct bifrost_regs {
494 unsigned uniform_const : 8;
495 unsigned reg2 : 6;
496 unsigned reg3 : 6;
497 unsigned reg0 : 5;
498 unsigned reg1 : 6;
499 unsigned ctrl : 4;
500 } __attribute__((packed));
501
502 enum bifrost_branch_cond {
503 BR_COND_LT = 0,
504 BR_COND_LE = 1,
505 BR_COND_GE = 2,
506 BR_COND_GT = 3,
507 // Equal vs. not-equal determined by src0/src1 comparison
508 BR_COND_EQ = 4,
509 // floating-point comparisons
510 // Becomes UNE when you flip the arguments
511 BR_COND_OEQ = 5,
512 // TODO what happens when you flip the arguments?
513 BR_COND_OGT = 6,
514 BR_COND_OLT = 7,
515 };
516
517 enum bifrost_branch_code {
518 BR_ALWAYS = 63,
519 };
520
521 struct bifrost_branch {
522 unsigned src0 : 3;
523
524 /* For BR_SIZE_ZERO, upper two bits become ctrl */
525 unsigned src1 : 3;
526
527 /* Offset source -- always uniform/const but
528 * theoretically could support indirect jumps? */
529 unsigned src2 : 3;
530
531 enum bifrost_branch_cond cond : 3;
532 enum branch_bit_size size : 3;
533
534 unsigned op : 5;
535 };
536
537 /* Clause packing */
538
539 #define BIFROST_FMA_NOP (0x701960 | BIFROST_SRC_STAGE)
540 #define BIFROST_ADD_NOP (0x3D960 | BIFROST_SRC_STAGE)
541
542 struct bifrost_fmt1 {
543 unsigned ins_0 : 3;
544 unsigned tag : 5;
545 uint64_t ins_1 : 64;
546 unsigned ins_2 : 11;
547 uint64_t header : 45;
548 } __attribute__((packed));
549
550 #define BIFROST_FMT1_INSTRUCTIONS 0b00101
551 #define BIFROST_FMT1_FINAL 0b01001
552 #define BIFROST_FMT1_CONSTANTS 0b00001
553
554 #define BIFROST_FMTC_CONSTANTS 0b0011
555 #define BIFROST_FMTC_FINAL 0b0111
556
557 struct bifrost_fmt_constant {
558 unsigned pos : 4;
559 unsigned tag : 4;
560 uint64_t imm_1 : 60;
561 uint64_t imm_2 : 60;
562 } __attribute__((packed));
563
564 enum bifrost_reg_control {
565 BIFROST_WRITE_FMA_P2 = 1,
566 BIFROST_WRITE_FMA_P2_READ_P3 = 2,
567 BIFROST_FIRST_WRITE_FMA_P2_READ_P3 = 3,
568 BIFROST_READ_P3 = 4,
569 BIFROST_WRITE_ADD_P2 = 5,
570 BIFROST_WRITE_ADD_P2_READ_P3 = 6,
571 BIFROST_WRITE_ADD_P2_FMA_P3 = 7,
572
573 BIFROST_FIRST_NONE = 8,
574 BIFROST_FIRST_WRITE_FMA_P2 = 9,
575 /* INSTR_INVALID_ENC */
576 BIFROST_REG_NONE = 11,
577 BIFROST_FIRST_READ_P3 = 12,
578 BIFROST_FIRST_WRITE_ADD_P2 = 13,
579 BIFROST_FIRST_WRITE_ADD_P2_READ_P3 = 14,
580 BIFROST_FIRST_WRITE_ADD_P2_FMA_P3 = 15
581 };
582
583 #endif