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 #include <string.h>
32 #include <assert.h>
33
34 #ifdef __cplusplus
35 extern "C" {
36 #endif
37
38 #define BIFROST_DBG_MSGS 0x0001
39 #define BIFROST_DBG_SHADERS 0x0002
40 #define BIFROST_DBG_SHADERDB 0x0004
41 #define BIFROST_DBG_VERBOSE 0x0008
42 #define BIFROST_DBG_INTERNAL 0x0010
43 #define BIFROST_DBG_NOSCHED 0x0020
44 #define BIFROST_DBG_INORDER 0x0040
45 #define BIFROST_DBG_NOVALIDATE 0x0080
46 #define BIFROST_DBG_NOOPT 0x0100
47 #define BIFROST_DBG_NOIDVS 0x0200
48 #define BIFROST_DBG_NOSB 0x0400
49 #define BIFROST_DBG_NOPRELOAD 0x0800
50 #define BIFROST_DBG_SPILL 0x1000
51 #define BIFROST_DBG_NOPSCHED 0x2000
52
53 extern int bifrost_debug;
54
55 enum bifrost_message_type {
56 BIFROST_MESSAGE_NONE = 0,
57 BIFROST_MESSAGE_VARYING = 1,
58 BIFROST_MESSAGE_ATTRIBUTE = 2,
59 BIFROST_MESSAGE_TEX = 3,
60 BIFROST_MESSAGE_VARTEX = 4,
61 BIFROST_MESSAGE_LOAD = 5,
62 BIFROST_MESSAGE_STORE = 6,
63 BIFROST_MESSAGE_ATOMIC = 7,
64 BIFROST_MESSAGE_BARRIER = 8,
65 BIFROST_MESSAGE_BLEND = 9,
66 BIFROST_MESSAGE_TILE = 10,
67 /* type 11 reserved */
68 BIFROST_MESSAGE_Z_STENCIL = 12,
69 BIFROST_MESSAGE_ATEST = 13,
70 BIFROST_MESSAGE_JOB = 14,
71 BIFROST_MESSAGE_64BIT = 15
72 };
73
74 enum bifrost_ftz {
75 BIFROST_FTZ_DISABLE = 0,
76 BIFROST_FTZ_DX11 = 1,
77 BIFROST_FTZ_ALWAYS = 2,
78 BIFROST_FTZ_ABRUPT = 3
79 };
80
81 enum bifrost_exceptions {
82 BIFROST_EXCEPTIONS_ENABLED = 0,
83 BIFROST_EXCEPTIONS_DISABLED = 1,
84 BIFROST_EXCEPTIONS_PRECISE_DIVISION = 2,
85 BIFROST_EXCEPTIONS_PRECISE_SQRT = 3,
86 };
87
88 /* Describes clause flow control, with respect to control flow and branch
89 * reconvergence.
90 *
91 * Control flow may be considered back-to-back (execute clauses back-to-back),
92 * non-back-to-back (switch warps after clause before the next clause), write
93 * elision (back-to-back and elide register slot #3 write from the clause), or
94 * end of shader.
95 *
96 * Branch reconvergence may be disabled, enabled unconditionally, or enabled
97 * based on the program counter. A clause requires reconvergence if it has a
98 * successor that can be executed without first executing the clause itself.
99 * Separate iterations of a loop are treated separately here, so it is also the
100 * case for a loop exit where the iteration count is not warp-invariant.
101 *
102 */
103
104 enum bifrost_flow {
105 /* End-of-shader */
106 BIFROST_FLOW_END = 0,
107
108 /* Non back-to-back, PC-encoded reconvergence */
109 BIFROST_FLOW_NBTB_PC = 1,
110
111 /* Non back-to-back, unconditional reconvergence */
112 BIFROST_FLOW_NBTB_UNCONDITIONAL = 2,
113
114 /* Non back-to-back, no reconvergence */
115 BIFROST_FLOW_NBTB = 3,
116
117 /* Back-to-back, unconditional reconvergence */
118 BIFROST_FLOW_BTB_UNCONDITIONAL = 4,
119
120 /* Back-to-back, no reconvergence */
121 BIFROST_FLOW_BTB_NONE = 5,
122
123 /* Write elision, unconditional reconvergence */
124 BIFROST_FLOW_WE_UNCONDITIONAL = 6,
125
126 /* Write elision, no reconvergence */
127 BIFROST_FLOW_WE = 7,
128 };
129
130 enum bifrost_slot {
131 /* 0-5 are general purpose */
132 BIFROST_SLOT_ELDEST_DEPTH = 6,
133 BIFROST_SLOT_ELDEST_COLOUR = 7,
134 };
135
136 struct bifrost_header {
137 /* Reserved */
138 unsigned zero1 : 5;
139
140 /* Flush-to-zero mode, leave zero for GL */
141 enum bifrost_ftz flush_to_zero : 2;
142
143 /* Convert any infinite result of any floating-point operation to the
144 * biggest representable number */
145 unsigned suppress_inf: 1;
146
147 /* Convert NaN to +0.0 */
148 unsigned suppress_nan : 1;
149
150 /* Floating-point excception handling mode */
151 enum bifrost_exceptions float_exceptions : 2;
152
153 /* Enum describing the flow control, which matters for handling
154 * divergence and reconvergence efficiently */
155 enum bifrost_flow flow_control : 3;
156
157 /* Reserved */
158 unsigned zero2 : 1;
159
160 /* Terminate discarded threads, rather than continuing execution. Set
161 * for fragment shaders for standard GL behaviour of DISCARD. Also in a
162 * fragment shader, this disables helper invocations, so cannot be used
163 * in a shader that requires derivatives or texture LOD computation */
164 unsigned terminate_discarded_threads : 1;
165
166 /* If set, the hardware may prefetch the next clause. If false, the
167 * hardware may not. Clear for unconditional branches. */
168 unsigned next_clause_prefetch : 1;
169
170 /* If set, a barrier will be inserted after the clause waiting for all
171 * message passing instructions to read their staging registers, such
172 * that it is safe for the next clause to write them. */
173 unsigned staging_barrier: 1;
174 unsigned staging_register : 6;
175
176 /* Slots to wait on and slot to be used for message passing
177 * instructions respectively */
178 unsigned dependency_wait : 8;
179 unsigned dependency_slot : 3;
180
181 enum bifrost_message_type message_type : 5;
182 enum bifrost_message_type next_message_type : 5;
183 } __attribute__((packed));
184
185 enum bifrost_packed_src {
186 BIFROST_SRC_PORT0 = 0,
187 BIFROST_SRC_PORT1 = 1,
188 BIFROST_SRC_PORT2 = 2,
189 BIFROST_SRC_STAGE = 3,
190 BIFROST_SRC_FAU_LO = 4,
191 BIFROST_SRC_FAU_HI = 5,
192 BIFROST_SRC_PASS_FMA = 6,
193 BIFROST_SRC_PASS_ADD = 7,
194 };
195
196 struct bifrost_fma_inst {
197 unsigned src0 : 3;
198 unsigned op : 20;
199 } __attribute__((packed));
200
201 struct bifrost_add_inst {
202 unsigned src0 : 3;
203 unsigned op : 17;
204 } __attribute__((packed));
205
206 enum branch_bit_size {
207 BR_SIZE_32 = 0,
208 BR_SIZE_16XX = 1,
209 BR_SIZE_16YY = 2,
210 // For the above combinations of bitsize and location, an extra bit is
211 // encoded via comparing the sources. The only possible source of ambiguity
212 // would be if the sources were the same, but then the branch condition
213 // would be always true or always false anyways, so we can ignore it. But
214 // this no longer works when comparing the y component to the x component,
215 // since it's valid to compare the y component of a source against its own
216 // x component. Instead, the extra bit is encoded via an extra bitsize.
217 BR_SIZE_16YX0 = 3,
218 BR_SIZE_16YX1 = 4,
219 BR_SIZE_32_AND_16X = 5,
220 BR_SIZE_32_AND_16Y = 6,
221 // Used for comparisons with zero and always-true, see below. I think this
222 // only works for integer comparisons.
223 BR_SIZE_ZERO = 7,
224 };
225
226 struct bifrost_regs {
227 unsigned fau_idx : 8;
228 unsigned reg3 : 6;
229 unsigned reg2 : 6;
230 unsigned reg0 : 5;
231 unsigned reg1 : 6;
232 unsigned ctrl : 4;
233 } __attribute__((packed));
234
235 #define BIFROST_FMTC_CONSTANTS 0b0011
236 #define BIFROST_FMTC_FINAL 0b0111
237
238 struct bifrost_fmt_constant {
239 unsigned pos : 4;
240 unsigned tag : 4;
241 uint64_t imm_1 : 60;
242 uint64_t imm_2 : 60;
243 } __attribute__((packed));
244
245 /* Clause formats, encoded in a table */
246
247 enum bi_clause_subword {
248 /* Literal 3-bit values */
249 BI_CLAUSE_SUBWORD_LITERAL_0 = 0,
250 /* etc */
251 BI_CLAUSE_SUBWORD_LITERAL_7 = 7,
252
253 /* The value of the corresponding tuple in the corresponding bits */
254 BI_CLAUSE_SUBWORD_TUPLE_0 = 8,
255 /* etc */
256 BI_CLAUSE_SUBWORD_TUPLE_7 = 15,
257
258 /* Clause header */
259 BI_CLAUSE_SUBWORD_HEADER = 16,
260
261 /* Leave zero, but semantically distinct from literal 0 */
262 BI_CLAUSE_SUBWORD_RESERVED = 17,
263
264 /* Embedded constant 0 */
265 BI_CLAUSE_SUBWORD_CONSTANT = 18,
266
267 /* M bits controlling modifier for the constant */
268 BI_CLAUSE_SUBWORD_M = 19,
269
270 /* Z bit: 1 to begin encoding constants, 0 to terminate the clause */
271 BI_CLAUSE_SUBWORD_Z = 20,
272
273 /* Upper 3-bits of a given tuple and zero extended */
274 BI_CLAUSE_SUBWORD_UPPER_0 = 32,
275 /* etc */
276 BI_CLAUSE_SUBWORD_UPPER_7 = BI_CLAUSE_SUBWORD_UPPER_0 + 7,
277
278 /* Upper 3-bits of two tuples, concatenated and zero-extended */
279 BI_CLAUSE_SUBWORD_UPPER_23 = BI_CLAUSE_SUBWORD_UPPER_0 + 23,
280 BI_CLAUSE_SUBWORD_UPPER_56 = BI_CLAUSE_SUBWORD_UPPER_0 + 56,
281 };
282
283 #define L(x) ((enum bi_clause_subword)(BI_CLAUSE_SUBWORD_LITERAL_0 + x))
284 #define U(x) ((enum bi_clause_subword)(BI_CLAUSE_SUBWORD_UPPER_0 + x))
285 #define T(x) ((enum bi_clause_subword)(BI_CLAUSE_SUBWORD_TUPLE_0 + x))
286 #define EC BI_CLAUSE_SUBWORD_CONSTANT
287 #define M BI_CLAUSE_SUBWORD_M
288 #define Z BI_CLAUSE_SUBWORD_Z
289 #define H BI_CLAUSE_SUBWORD_HEADER
290 #define R BI_CLAUSE_SUBWORD_RESERVED
291
292 struct bi_clause_format {
293 unsigned format; /* format number */
294 unsigned pos; /* index in the clause */
295 enum bi_clause_subword tag_1; /* 2-bits */
296 enum bi_clause_subword tag_2; /* 3-bits */
297 enum bi_clause_subword tag_3; /* 3-bits */
298 enum bi_clause_subword s0_s3; /* 60 bits */
299 enum bi_clause_subword s4; /* 15 bits */
300 enum bi_clause_subword s5_s6; /* 30 bits */
301 enum bi_clause_subword s7; /* 15 bits */
302 };
303
304 static const struct bi_clause_format bi_clause_formats[] = {
305 { 0, 0, L(0), L(5), U(0), T(0), T(0), H, H },
306 { 0, 0, Z, L(1), U(0), T(0), T(0), H, H },
307 { 1, 1, Z, L(0), L(3), T(1), T(1), R, U(1) },
308 { 2, 1, L(0), L(4), U(1), T(1), T(1), T(2), T(2) },
309 { 3, 2, Z, L(0), L(4), EC, M, T(2), U(2) },
310 { 4, 2, L(0), L(0), L(1), T(3), T(3), T(2), U(23) },
311 { 4, 2, Z, L(0), L(5), T(3), T(3), T(2), U(23) },
312 { 5, 2, L(2), U(3), U(2), T(3), T(3), T(2), EC },
313 { 6, 3, Z, L(2), U(4), T(4), T(4), EC, EC },
314 { 7, 3, L(1), L(4), U(4), T(4), T(4), T(5), T(5) },
315 { 8, 4, Z, L(0), L(6), EC, M, T(5), U(5) },
316 { 9, 4, Z, L(0), L(7), T(6), T(6), T(5), U(56) },
317 { 10, 4, L(3), U(6), U(5), T(6), T(6), T(5), EC },
318 { 11, 5, Z, L(3), U(7), T(7), T(7), EC, EC },
319 };
320
321 #undef L
322 #undef U
323 #undef T
324 #undef EC
325 #undef M
326 #undef Z
327 #undef H
328 #undef R
329
330 /* 32-bit modes for slots 2/3, as encoded in the register block. Other values
331 * are reserved. First part specifies behaviour of slot 2 (Idle, Read, Write
332 * Full, Write Low, Write High), second part behaviour of slot 3, and the last
333 * part specifies the source for the write (FMA, ADD, or MIX for FMA/ADD).
334 *
335 * IDLE is a special mode disabling both slots, except for the first
336 * instruction in the clause which uses IDLE_1 for the same purpose.
337 *
338 * All fields 0 used as sentinel for reserved encoding, so IDLE(_1) have FMA
339 * set (and ignored) as a placeholder to differentiate from reserved.
340 */
341 enum bifrost_reg_mode {
342 BIFROST_R_WL_FMA = 1,
343 BIFROST_R_WH_FMA = 2,
344 BIFROST_R_W_FMA = 3,
345 BIFROST_R_WL_ADD = 4,
346 BIFROST_R_WH_ADD = 5,
347 BIFROST_R_W_ADD = 6,
348 BIFROST_WL_WL_ADD = 7,
349 BIFROST_WL_WH_ADD = 8,
350 BIFROST_WL_W_ADD = 9,
351 BIFROST_WH_WL_ADD = 10,
352 BIFROST_WH_WH_ADD = 11,
353 BIFROST_WH_W_ADD = 12,
354 BIFROST_W_WL_ADD = 13,
355 BIFROST_W_WH_ADD = 14,
356 BIFROST_W_W_ADD = 15,
357 BIFROST_IDLE_1 = 16,
358 BIFROST_I_W_FMA = 17,
359 BIFROST_I_WL_FMA = 18,
360 BIFROST_I_WH_FMA = 19,
361 BIFROST_R_I = 20,
362 BIFROST_I_W_ADD = 21,
363 BIFROST_I_WL_ADD = 22,
364 BIFROST_I_WH_ADD = 23,
365 BIFROST_WL_WH_MIX = 24,
366 BIFROST_WH_WL_MIX = 26,
367 BIFROST_IDLE = 27,
368 };
369
370 enum bifrost_reg_op {
371 BIFROST_OP_IDLE = 0,
372 BIFROST_OP_READ = 1,
373 BIFROST_OP_WRITE = 2,
374 BIFROST_OP_WRITE_LO = 3,
375 BIFROST_OP_WRITE_HI = 4,
376 };
377
378 struct bifrost_reg_ctrl_23 {
379 enum bifrost_reg_op slot2;
380 enum bifrost_reg_op slot3;
381 bool slot3_fma;
382 };
383
384 #ifndef __cplusplus
385 static const struct bifrost_reg_ctrl_23 bifrost_reg_ctrl_lut[32] = {
386 [BIFROST_R_WL_FMA] = { BIFROST_OP_READ, BIFROST_OP_WRITE_LO, true },
387 [BIFROST_R_WH_FMA] = { BIFROST_OP_READ, BIFROST_OP_WRITE_HI, true },
388 [BIFROST_R_W_FMA] = { BIFROST_OP_READ, BIFROST_OP_WRITE, true },
389 [BIFROST_R_WL_ADD] = { BIFROST_OP_READ, BIFROST_OP_WRITE_LO, false },
390 [BIFROST_R_WH_ADD] = { BIFROST_OP_READ, BIFROST_OP_WRITE_HI, false },
391 [BIFROST_R_W_ADD] = { BIFROST_OP_READ, BIFROST_OP_WRITE, false },
392 [BIFROST_WL_WL_ADD] = { BIFROST_OP_WRITE_LO, BIFROST_OP_WRITE_LO, false },
393 [BIFROST_WL_WH_ADD] = { BIFROST_OP_WRITE_LO, BIFROST_OP_WRITE_HI, false },
394 [BIFROST_WL_W_ADD] = { BIFROST_OP_WRITE_LO, BIFROST_OP_WRITE, false },
395 [BIFROST_WH_WL_ADD] = { BIFROST_OP_WRITE_HI, BIFROST_OP_WRITE_LO, false },
396 [BIFROST_WH_WH_ADD] = { BIFROST_OP_WRITE_HI, BIFROST_OP_WRITE_HI, false },
397 [BIFROST_WH_W_ADD] = { BIFROST_OP_WRITE_HI, BIFROST_OP_WRITE, false },
398 [BIFROST_W_WL_ADD] = { BIFROST_OP_WRITE, BIFROST_OP_WRITE_LO, false },
399 [BIFROST_W_WH_ADD] = { BIFROST_OP_WRITE, BIFROST_OP_WRITE_HI, false },
400 [BIFROST_W_W_ADD] = { BIFROST_OP_WRITE, BIFROST_OP_WRITE, false },
401 [BIFROST_IDLE_1] = { BIFROST_OP_IDLE, BIFROST_OP_IDLE, true },
402 [BIFROST_I_W_FMA] = { BIFROST_OP_IDLE, BIFROST_OP_WRITE, true },
403 [BIFROST_I_WL_FMA] = { BIFROST_OP_IDLE, BIFROST_OP_WRITE_LO, true },
404 [BIFROST_I_WH_FMA] = { BIFROST_OP_IDLE, BIFROST_OP_WRITE_HI, true },
405 [BIFROST_R_I] = { BIFROST_OP_READ, BIFROST_OP_IDLE, false },
406 [BIFROST_I_W_ADD] = { BIFROST_OP_IDLE, BIFROST_OP_WRITE, false },
407 [BIFROST_I_WL_ADD] = { BIFROST_OP_IDLE, BIFROST_OP_WRITE_LO, false },
408 [BIFROST_I_WH_ADD] = { BIFROST_OP_IDLE, BIFROST_OP_WRITE_HI, false },
409 [BIFROST_WL_WH_MIX] = { BIFROST_OP_WRITE_LO, BIFROST_OP_WRITE_HI, false },
410 [BIFROST_WH_WL_MIX] = { BIFROST_OP_WRITE_HI, BIFROST_OP_WRITE_LO, false },
411 [BIFROST_IDLE] = { BIFROST_OP_IDLE, BIFROST_OP_IDLE, true },
412 };
413 #endif
414
415 /* Texture operator descriptors in various states. Usually packed in the
416 * compiler and stored as a constant */
417
418 enum bifrost_texture_operation_mode {
419 /* Dual texturing */
420 BIFROST_TEXTURE_OPERATION_DUAL = 1,
421
422 /* Single texturing */
423 BIFROST_TEXTURE_OPERATION_SINGLE = 3,
424 };
425
426 enum bifrost_index {
427 /* Both texture/sampler index immediate */
428 BIFROST_INDEX_IMMEDIATE_SHARED = 0,
429
430 /* Sampler index immediate, texture index from staging */
431 BIFROST_INDEX_IMMEDIATE_SAMPLER = 1,
432
433 /* Texture index immediate, sampler index from staging */
434 BIFROST_INDEX_IMMEDIATE_TEXTURE = 2,
435
436 /* Both indices from (separate) staging registers */
437 BIFROST_INDEX_REGISTER = 3,
438 };
439
440 enum bifrost_tex_op {
441 /* Given explicit derivatives, compute a gradient descriptor */
442 BIFROST_TEX_OP_GRDESC_DER = 4,
443
444 /* Given implicit derivatives (texture coordinates in a fragment
445 * shader), compute a gradient descriptor */
446 BIFROST_TEX_OP_GRDESC = 5,
447
448 /* Fetch a texel. Takes a staging register with LOD level / face index
449 * packed 16:16 */
450 BIFROST_TEX_OP_FETCH = 6,
451
452 /* Filtered texture */
453 BIFROST_TEX_OP_TEX = 7,
454 };
455
456 enum bifrost_lod_mode {
457 /* Takes two staging registers forming a 64-bit gradient descriptor
458 * (computed by a previous GRDESC or GRDESC_DER operation) */
459 BIFROST_LOD_MODE_GRDESC = 3,
460
461 /* Take a staging register with 8:8 fixed-point in bottom 16-bits
462 * specifying an explicit LOD */
463 BIFROST_LOD_MODE_EXPLICIT = 4,
464
465 /* Takes a staging register with bottom 16-bits as 8:8 fixed-point LOD
466 * bias and top 16-bit as 8:8 fixed-point lower bound (generally left
467 * zero), added and clamped to a computed LOD */
468 BIFROST_LOD_MODE_BIAS = 5,
469
470 /* Set LOD to zero */
471 BIFROST_LOD_MODE_ZERO = 6,
472
473 /* Compute LOD */
474 BIFROST_LOD_MODE_COMPUTE = 7,
475 };
476
477 enum bifrost_texture_format {
478 /* 16-bit floating point, with optional clamping */
479 BIFROST_TEXTURE_FORMAT_F16 = 0,
480 BIFROST_TEXTURE_FORMAT_F16_POS = 1,
481 BIFROST_TEXTURE_FORMAT_F16_PM1 = 2,
482 BIFROST_TEXTURE_FORMAT_F16_1 = 3,
483
484 /* 32-bit floating point, with optional clamping */
485 BIFROST_TEXTURE_FORMAT_F32 = 4,
486 BIFROST_TEXTURE_FORMAT_F32_POS = 5,
487 BIFROST_TEXTURE_FORMAT_F32_PM1 = 6,
488 BIFROST_TEXTURE_FORMAT_F32_1 = 7,
489 };
490
491 enum bifrost_texture_format_full {
492 /* Transclude bifrost_texture_format from above */
493
494 /* Integers, unclamped */
495 BIFROST_TEXTURE_FORMAT_U16 = 12,
496 BIFROST_TEXTURE_FORMAT_S16 = 13,
497 BIFROST_TEXTURE_FORMAT_U32 = 14,
498 BIFROST_TEXTURE_FORMAT_S32 = 15,
499 };
500
501 enum bifrost_texture_fetch {
502 /* Default texelFetch */
503 BIFROST_TEXTURE_FETCH_TEXEL = 1,
504
505 /* Deprecated, fetches 4x U32 of a U8 x 4 texture. Do not use. */
506 BIFROST_TEXTURE_FETCH_GATHER4_RGBA = 3,
507
508 /* Gathers */
509 BIFROST_TEXTURE_FETCH_GATHER4_R = 4,
510 BIFROST_TEXTURE_FETCH_GATHER4_G = 5,
511 BIFROST_TEXTURE_FETCH_GATHER4_B = 6,
512 BIFROST_TEXTURE_FETCH_GATHER4_A = 7
513 };
514
515 struct bifrost_texture_operation {
516 /* If immediate_indices is set:
517 * - immediate sampler index
518 * - index used as texture index
519 * Otherwise:
520 * - bifrost_single_index in lower 2 bits
521 * - 0x3 in upper 2 bits (single-texturing)
522 */
523 unsigned sampler_index_or_mode : 4;
524 unsigned index : 7;
525 bool immediate_indices : 1;
526 enum bifrost_tex_op op : 3;
527
528 /* If set for TEX/FETCH, loads texel offsets and multisample index from
529 * a staging register containing offset_x:offset_y:offset_z:ms_index
530 * packed 8:8:8:8. Offsets must be in [-31, +31]. If set for
531 * GRDESC(_DER), disable LOD bias. */
532 bool offset_or_bias_disable : 1;
533
534 /* If set for TEX/FETCH, loads fp32 shadow comparison value from a
535 * staging register. Implies fetch_component = gather4_r. If set for
536 * GRDESC(_DER), disables LOD clamping. */
537 bool shadow_or_clamp_disable : 1;
538
539 /* If set, loads an uint32 array index from a staging register. */
540 bool array : 1;
541
542 /* Texture dimension, or 0 for a cubemap */
543 unsigned dimension : 2;
544
545 /* Method to compute LOD value or for a FETCH, the
546 * bifrost_texture_fetch component specification */
547 enum bifrost_lod_mode lod_or_fetch : 3;
548
549 /* Reserved */
550 unsigned zero : 1;
551
552 /* Register format for the result */
553 enum bifrost_texture_format_full format : 4;
554
555 /* Write mask for the result */
556 unsigned mask : 4;
557 } __attribute__((packed));
558
559 struct bifrost_dual_texture_operation {
560 unsigned primary_sampler_index : 2;
561 unsigned mode : 2; /* 0x1 for dual */
562 unsigned primary_texture_index : 2;
563 unsigned secondary_sampler_index : 2;
564 unsigned secondary_texture_index : 2;
565
566 /* Leave zero for dual texturing */
567 unsigned reserved : 1;
568 unsigned index_mode_zero : 1;
569
570 /* Base staging register to write the secondary results to */
571 unsigned secondary_register : 6;
572
573 /* Format/mask for each texture */
574 enum bifrost_texture_format secondary_format : 3;
575 unsigned secondary_mask : 4;
576
577 enum bifrost_texture_format primary_format : 3;
578 unsigned primary_mask : 4;
579 } __attribute__((packed));
580
581 static inline uint32_t
bi_dual_tex_as_u32(struct bifrost_dual_texture_operation desc)582 bi_dual_tex_as_u32(struct bifrost_dual_texture_operation desc)
583 {
584 uint32_t desc_u;
585 memcpy(&desc_u, &desc, sizeof(desc));
586
587 return desc_u;
588 }
589
590 #define BIFROST_MEGA_SAMPLE 128
591 #define BIFROST_ALL_SAMPLES 255
592 #define BIFROST_CURRENT_PIXEL 255
593
594 struct bifrost_pixel_indices {
595 unsigned sample : 8;
596 unsigned rt : 8;
597 unsigned x : 8;
598 unsigned y : 8;
599 } __attribute__((packed));
600
601 enum bi_constmod {
602 BI_CONSTMOD_NONE,
603 BI_CONSTMOD_PC_LO,
604 BI_CONSTMOD_PC_HI,
605 BI_CONSTMOD_PC_LO_HI
606 };
607
608 struct bi_constants {
609 /* Raw constant values */
610 uint64_t raw[6];
611
612 /* Associated modifier derived from M values */
613 enum bi_constmod mods[6];
614 };
615
616 /* FAU selectors for constants are out-of-order, construct the top bits
617 * here given a embedded constant index in a clause */
618
619 static inline unsigned
bi_constant_field(unsigned idx)620 bi_constant_field(unsigned idx)
621 {
622 const unsigned values[] = {
623 4, 5, 6, 7, 2, 3
624 };
625
626 assert(idx <= 5);
627 return values[idx] << 4;
628 }
629
630 #ifdef __cplusplus
631 } /* extern C */
632 #endif
633
634 #endif
635