pan/bi: Pack st_vary
[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
101 struct bifrost_fma_inst {
102 unsigned src0 : 3;
103 unsigned op : 20;
104 } __attribute__((packed));
105
106 #define BIFROST_ADD_OP_BLEND (0x1952c)
107
108 struct bifrost_add_inst {
109 unsigned src0 : 3;
110 unsigned op : 17;
111 } __attribute__((packed));
112
113 #define BIFROST_ADD_OP_LD_UBO_1 (0x0c1a0 >> 3)
114 #define BIFROST_ADD_OP_LD_UBO_2 (0x0c1e0 >> 3)
115 #define BIFROST_ADD_OP_LD_UBO_3 (0x0caa0 >> 3)
116 #define BIFROST_ADD_OP_LD_UBO_4 (0x0c220 >> 3)
117
118 struct bifrost_add_2src {
119 unsigned src0 : 3;
120 unsigned src1 : 3;
121 unsigned op : 14;
122 } __attribute__((packed));
123
124 #define BIFROST_ADD_OP_ST_VAR (0x19300 >> 8)
125
126 struct bifrost_st_vary {
127 unsigned src0 : 3;
128 unsigned src1 : 3;
129 unsigned src2 : 3;
130 unsigned channels : 2;
131 unsigned op : 9;
132 } __attribute__((packed));
133
134 #define BIFROST_ADD_OP_ATEST (0xc8f)
135
136 struct bifrost_add_atest {
137 /* gl_SampleMask (R60) */
138 unsigned src0 : 3;
139
140 /* Alpha value */
141 unsigned src1 : 3;
142
143 /* If half, X/Y select. If !half, always set */
144 unsigned component : 1;
145 unsigned half : 1;
146
147 unsigned op : 12;
148 } __attribute__((packed));
149
150 enum bifrost_outmod {
151 BIFROST_NONE = 0x0,
152 BIFROST_POS = 0x1,
153 BIFROST_SAT_SIGNED = 0x2,
154 BIFROST_SAT = 0x3,
155 };
156
157 enum bifrost_roundmode {
158 BIFROST_RTE = 0x0, /* round to even */
159 BIFROST_RTP = 0x1, /* round to positive */
160 BIFROST_RTN = 0x2, /* round to negative */
161 BIFROST_RTZ = 0x3 /* round to zero */
162 };
163
164 /* NONE: Same as fmax() and fmin() -- return the other
165 * number if any number is NaN. Also always return +0 if
166 * one argument is +0 and the other is -0.
167 *
168 * NAN_WINS: Instead of never returning a NaN, always return
169 * one. The "greater"/"lesser" NaN is always returned, first
170 * by checking the sign and then the mantissa bits.
171 *
172 * SRC1_WINS: For max, implement src0 > src1 ? src0 : src1.
173 * For min, implement src0 < src1 ? src0 : src1. This
174 * includes handling NaN's and signedness of 0 differently
175 * from above, since +0 and -0 compare equal and comparisons
176 * always return false for NaN's. As a result, this mode is
177 * *not* commutative.
178 *
179 * SRC0_WINS: For max, implement src0 < src1 ? src1 : src0
180 * For min, implement src0 > src1 ? src1 : src0
181 */
182
183
184 enum bifrost_minmax_mode {
185 BIFROST_MINMAX_NONE = 0x0,
186 BIFROST_NAN_WINS = 0x1,
187 BIFROST_SRC1_WINS = 0x2,
188 BIFROST_SRC0_WINS = 0x3,
189 };
190
191 #define BIFROST_FMA_OP_FADD32 (0x58 >> 2)
192
193 struct bifrost_fma_add {
194 unsigned src0 : 3;
195 unsigned src1 : 3;
196 unsigned src1_abs : 1;
197 unsigned src0_neg : 1;
198 unsigned src1_neg : 1;
199 unsigned unk : 3;
200 unsigned src0_abs : 1;
201 enum bifrost_outmod outmod : 2;
202 enum bifrost_roundmode roundmode : 2;
203 unsigned op : 6;
204 } __attribute__((packed));
205
206 #define BIFROST_FMA_OP_FMA (0x00)
207
208 struct bifrost_fma_fma {
209 unsigned src0 : 3;
210 unsigned src1 : 3;
211 unsigned src2 : 3;
212 unsigned src_expand : 3;
213 unsigned src0_abs : 1;
214 unsigned unk : 4;
215 unsigned src0_neg : 1;
216 unsigned src2_neg : 1;
217 unsigned src1_abs : 1;
218 unsigned src2_abs : 1;
219 unsigned op : 5;
220 } __attribute__((packed));
221
222 enum bifrost_csel_cond {
223 BIFROST_FEQ_F = 0x0,
224 BIFROST_FGT_F = 0x1,
225 BIFROST_FGE_F = 0x2,
226 BIFROST_IEQ_F = 0x3,
227 BIFROST_IGT_I = 0x4,
228 BIFROST_IGE_I = 0x5,
229 BIFROST_UGT_I = 0x6,
230 BIFROST_UGE_I = 0x7
231 };
232
233 struct bifrost_csel4 {
234 unsigned src0 : 3;
235 unsigned src1 : 3;
236 unsigned src2 : 3;
237 unsigned src3 : 3;
238 enum bifrost_csel_cond cond : 3;
239 unsigned op : 8;
240 } __attribute__((packed));
241
242 struct bifrost_shift_fma {
243 unsigned src0 : 3;
244 unsigned src1 : 3;
245 unsigned src2 : 3;
246 unsigned half : 3; /* 000 for i32, 100 for i8, 111 for v2i16 */
247 unsigned unk : 1; /* always set? */
248 unsigned invert_1 : 1; /* Inverts sources to combining op */
249 /* For XOR, switches RSHIFT to LSHIFT since only one invert needed */
250 unsigned invert_2 : 1;
251 unsigned op : 8;
252 } __attribute__((packed));
253
254 struct bifrost_shift_add {
255 unsigned src0 : 3;
256 unsigned src1 : 3;
257 unsigned src2 : 3;
258 unsigned zero : 2;
259
260 unsigned invert_1 : 1;
261 unsigned invert_2 : 1;
262
263 unsigned op : 7;
264 } __attribute__((packed));
265
266 enum bifrost_ldst_type {
267 BIFROST_LDST_F16 = 0,
268 BIFROST_LDST_F32 = 1,
269 BIFROST_LDST_I32 = 2,
270 BIFROST_LDST_U32 = 3
271 };
272
273 #define BIFROST_ADD_OP_LD_VAR_ADDR (0x18000 >> 10)
274
275 struct bifrost_ld_var_addr {
276 unsigned src0 : 3;
277 unsigned src1 : 3;
278 unsigned location : 5;
279 enum bifrost_ldst_type type : 2;
280 unsigned op : 7;
281 } __attribute__((packed));
282
283 struct bifrost_ld_attr {
284 unsigned src0 : 3;
285 unsigned src1 : 3;
286 unsigned location : 5;
287 unsigned channels : 2; /* MALI_POSITIVE */
288 enum bifrost_ldst_type type : 2;
289 unsigned op : 5;
290 } __attribute__((packed));
291
292 enum bifrost_interp_mode {
293 BIFROST_INTERP_PER_FRAG = 0x0,
294 BIFROST_INTERP_CENTROID = 0x1,
295 BIFROST_INTERP_DEFAULT = 0x2,
296 BIFROST_INTERP_EXPLICIT = 0x3
297 };
298
299 #define BIFROST_ADD_OP_LD_VAR_16 (0x1a << 1)
300 #define BIFROST_ADD_OP_LD_VAR_32 (0x0a << 1)
301
302 struct bifrost_ld_var {
303 unsigned src0 : 3;
304
305 /* If top two bits set, indirect with src in bottom three */
306 unsigned addr : 5;
307
308 unsigned channels : 2; /* MALI_POSITIVE */
309 enum bifrost_interp_mode interp_mode : 2;
310 unsigned reuse : 1;
311 unsigned flat : 1;
312 unsigned op : 6;
313 } __attribute__((packed));
314
315 struct bifrost_tex_ctrl {
316 unsigned sampler_index : 4; // also used to signal indirects
317 unsigned tex_index : 7;
318 bool no_merge_index : 1; // whether to merge (direct) sampler & texture indices
319 bool filter : 1; // use the usual filtering pipeline (0 for texelFetch & textureGather)
320 unsigned unk0 : 2;
321 bool texel_offset : 1; // *Offset()
322 bool is_shadow : 1;
323 bool is_array : 1;
324 unsigned tex_type : 2; // 2D, 3D, Cube, Buffer
325 bool compute_lod : 1; // 0 for *Lod()
326 bool not_supply_lod : 1; // 0 for *Lod() or when a bias is applied
327 bool calc_gradients : 1; // 0 for *Grad()
328 unsigned unk1 : 1;
329 unsigned result_type : 4; // integer, unsigned, float TODO: why is this 4 bits?
330 unsigned unk2 : 4;
331 } __attribute__((packed));
332
333 struct bifrost_dual_tex_ctrl {
334 unsigned sampler_index0 : 2;
335 unsigned unk0 : 2;
336 unsigned tex_index0 : 2;
337 unsigned sampler_index1 : 2;
338 unsigned tex_index1 : 2;
339 unsigned unk1 : 22;
340 } __attribute__((packed));
341
342 enum branch_bit_size {
343 BR_SIZE_32 = 0,
344 BR_SIZE_16XX = 1,
345 BR_SIZE_16YY = 2,
346 // For the above combinations of bitsize and location, an extra bit is
347 // encoded via comparing the sources. The only possible source of ambiguity
348 // would be if the sources were the same, but then the branch condition
349 // would be always true or always false anyways, so we can ignore it. But
350 // this no longer works when comparing the y component to the x component,
351 // since it's valid to compare the y component of a source against its own
352 // x component. Instead, the extra bit is encoded via an extra bitsize.
353 BR_SIZE_16YX0 = 3,
354 BR_SIZE_16YX1 = 4,
355 BR_SIZE_32_AND_16X = 5,
356 BR_SIZE_32_AND_16Y = 6,
357 // Used for comparisons with zero and always-true, see below. I think this
358 // only works for integer comparisons.
359 BR_SIZE_ZERO = 7,
360 };
361
362 enum bifrost_reg_write_unit {
363 REG_WRITE_NONE = 0, // don't write
364 REG_WRITE_TWO, // write using reg2
365 REG_WRITE_THREE, // write using reg3
366 };
367
368 struct bifrost_regs {
369 unsigned uniform_const : 8;
370 unsigned reg2 : 6;
371 unsigned reg3 : 6;
372 unsigned reg0 : 5;
373 unsigned reg1 : 6;
374 unsigned ctrl : 4;
375 } __attribute__((packed));
376
377 enum bifrost_branch_cond {
378 BR_COND_LT = 0,
379 BR_COND_LE = 1,
380 BR_COND_GE = 2,
381 BR_COND_GT = 3,
382 // Equal vs. not-equal determined by src0/src1 comparison
383 BR_COND_EQ = 4,
384 // floating-point comparisons
385 // Becomes UNE when you flip the arguments
386 BR_COND_OEQ = 5,
387 // TODO what happens when you flip the arguments?
388 BR_COND_OGT = 6,
389 BR_COND_OLT = 7,
390 };
391
392 enum bifrost_branch_code {
393 BR_ALWAYS = 63,
394 };
395
396 struct bifrost_branch {
397 unsigned src0 : 3;
398
399 /* For BR_SIZE_ZERO, upper two bits become ctrl */
400 unsigned src1 : 3;
401
402 /* Offset source -- always uniform/const but
403 * theoretically could support indirect jumps? */
404 unsigned src2 : 3;
405
406 enum bifrost_branch_cond cond : 3;
407 enum branch_bit_size size : 3;
408
409 unsigned op : 5;
410 };
411
412 /* Clause packing */
413
414 #define BIFROST_FMA_NOP (0x701960)
415 #define BIFROST_ADD_NOP (0x3D960)
416
417 struct bifrost_fmt1 {
418 unsigned ins_0 : 3;
419 unsigned tag : 5;
420 uint64_t ins_1 : 64;
421 unsigned ins_2 : 11;
422 uint64_t header : 45;
423 } __attribute__((packed));
424
425 #define BIFROST_FMT1_INSTRUCTIONS 0b00101
426 #define BIFROST_FMT1_FINAL 0b01001
427 #define BIFROST_FMT1_CONSTANTS 0b00001
428
429 #define BIFROST_FMTC_CONSTANTS 0b0011
430 #define BIFROST_FMTC_FINAL 0b0111
431
432 struct bifrost_fmt_constant {
433 unsigned pos : 4;
434 unsigned tag : 4;
435 uint64_t imm_1 : 60;
436 uint64_t imm_2 : 60;
437 } __attribute__((packed));
438
439 enum bifrost_reg_control {
440 BIFROST_WRITE_FMA_P2 = 1,
441 BIFROST_WRITE_FMA_P2_READ_P3 = 2,
442 BIFROST_WRITE_FMA_P2_READ_P3_ALT = 3,
443 BIFROST_READ_P3 = 4,
444 BIFROST_WRITE_ADD_P2 = 5,
445 BIFROST_WRITE_ADD_P2_READ_P3 = 6,
446 BIFROST_WRITE_ADD_P2_FMA_P3 = 7,
447
448 BIFROST_FIRST_NONE = 8,
449 BIFROST_FIRST_WRITE_FMA_P2 = 9,
450 BIFROST_REG_NONE = 11,
451 BIFROST_FIRST_READ_P3 = 12,
452 BIFROST_FIRST_WRITE_ADD_P2 = 13,
453 BIFROST_FIRST_WRITE_ADD_P2_READ_P3 = 14,
454 BIFROST_FIRST_WRITE_ADD_P2_FMA_P3 = 15
455 };
456
457 #endif