e86f517442de448c1edcf932783daaffc2a8fa74
[mesa.git] / src / gallium / drivers / panfrost / midgard / midgard.h
1 /* Author(s):
2 * Connor Abbott
3 * Alyssa Rosenzweig
4 *
5 * Copyright (c) 2013 Connor Abbott (connor@abbott.cx)
6 * Copyright (c) 2018 Alyssa Rosenzweig (alyssa@rosenzweig.io)
7 *
8 * Permission is hereby granted, free of charge, to any person obtaining a copy
9 * of this software and associated documentation files (the "Software"), to deal
10 * in the Software without restriction, including without limitation the rights
11 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
12 * copies of the Software, and to permit persons to whom the Software is
13 * furnished to do so, subject to the following conditions:
14 *
15 * The above copyright notice and this permission notice shall be included in
16 * all copies or substantial portions of the Software.
17 *
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
19 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
20 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
21 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
22 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
23 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
24 * THE SOFTWARE.
25 */
26
27 #ifndef __midgard_h__
28 #define __midgard_h__
29
30 #include <stdint.h>
31 #include <stdbool.h>
32
33 #define MIDGARD_DBG_MSGS 0x0001
34 #define MIDGARD_DBG_SHADERS 0x0002
35
36 extern int midgard_debug;
37
38 typedef enum {
39 midgard_word_type_alu,
40 midgard_word_type_load_store,
41 midgard_word_type_texture,
42 midgard_word_type_unknown
43 } midgard_word_type;
44
45 typedef enum {
46 midgard_alu_vmul,
47 midgard_alu_sadd,
48 midgard_alu_smul,
49 midgard_alu_vadd,
50 midgard_alu_lut
51 } midgard_alu;
52
53 /*
54 * ALU words
55 */
56
57 typedef enum {
58 midgard_alu_op_fadd = 0x10,
59 midgard_alu_op_fmul = 0x14,
60
61 midgard_alu_op_fmin = 0x28,
62 midgard_alu_op_fmax = 0x2C,
63
64 midgard_alu_op_fmov = 0x30,
65 midgard_alu_op_froundeven = 0x34,
66 midgard_alu_op_ftrunc = 0x35,
67 midgard_alu_op_ffloor = 0x36,
68 midgard_alu_op_fceil = 0x37,
69 midgard_alu_op_ffma = 0x38,
70 midgard_alu_op_fdot3 = 0x3C,
71 midgard_alu_op_fdot3r = 0x3D,
72 midgard_alu_op_fdot4 = 0x3E,
73 midgard_alu_op_freduce = 0x3F,
74
75 midgard_alu_op_iadd = 0x40,
76 midgard_alu_op_ishladd = 0x41,
77 midgard_alu_op_isub = 0x46,
78
79 midgard_alu_op_imul = 0x58,
80
81 midgard_alu_op_imin = 0x60,
82 midgard_alu_op_umin = 0x61,
83 midgard_alu_op_imax = 0x62,
84 midgard_alu_op_umax = 0x63,
85 midgard_alu_op_iasr = 0x68,
86 midgard_alu_op_ilsr = 0x69,
87 midgard_alu_op_ishl = 0x6E,
88
89 midgard_alu_op_iand = 0x70,
90 midgard_alu_op_ior = 0x71,
91 midgard_alu_op_inand = 0x72, /* ~(a & b), for inot let a = b */
92 midgard_alu_op_inor = 0x73, /* ~(a | b) */
93 midgard_alu_op_iandnot = 0x74, /* (a & ~b), used for not/b2f */
94 midgard_alu_op_iornot = 0x75, /* (a | ~b) */
95 midgard_alu_op_ixor = 0x76,
96 midgard_alu_op_inxor = 0x77, /* ~(a & b) */
97 midgard_alu_op_ilzcnt = 0x78, /* Number of zeroes on left. 31 - ilzcnt(x) = findMSB(x) */
98 midgard_alu_op_ibitcount8 = 0x7A, /* Counts bits in 8-bit increments */
99 midgard_alu_op_imov = 0x7B,
100 midgard_alu_op_iabs = 0x7C,
101
102 midgard_alu_op_feq = 0x80,
103 midgard_alu_op_fne = 0x81,
104 midgard_alu_op_flt = 0x82,
105 midgard_alu_op_fle = 0x83,
106 midgard_alu_op_fball_eq = 0x88,
107 midgard_alu_op_bball_eq = 0x89,
108 midgard_alu_op_fball_lt = 0x8A, /* all(lessThan(.., ..)) */
109 midgard_alu_op_fball_lte = 0x8B, /* all(lessThanEqual(.., ..)) */
110
111 midgard_alu_op_bbany_neq = 0x90, /* used for bvec4(1) */
112 midgard_alu_op_fbany_neq = 0x91, /* bvec4(0) also */
113 midgard_alu_op_fbany_lt = 0x92, /* any(lessThan(.., ..)) */
114 midgard_alu_op_fbany_lte = 0x93, /* any(lessThanEqual(.., ..)) */
115 midgard_alu_op_f2i = 0x99,
116 midgard_alu_op_f2u8 = 0x9C,
117 midgard_alu_op_f2u = 0x9D,
118
119 midgard_alu_op_ieq = 0xA0,
120 midgard_alu_op_ine = 0xA1,
121 midgard_alu_op_ult = 0xA2,
122 midgard_alu_op_ule = 0xA3,
123 midgard_alu_op_ilt = 0xA4,
124 midgard_alu_op_ile = 0xA5,
125 midgard_alu_op_iball_eq = 0xA8,
126 midgard_alu_op_iball_neq = 0xA9,
127 midgard_alu_op_uball_lt = 0xAA,
128 midgard_alu_op_uball_lte = 0xAB,
129 midgard_alu_op_iball_lt = 0xAC,
130 midgard_alu_op_iball_lte = 0xAD,
131
132 midgard_alu_op_ibany_eq = 0xB0,
133 midgard_alu_op_ibany_neq = 0xB1,
134 midgard_alu_op_ubany_lt = 0xB2,
135 midgard_alu_op_ubany_lte = 0xB3,
136 midgard_alu_op_ibany_lt = 0xB4, /* any(lessThan(.., ..)) */
137 midgard_alu_op_ibany_lte = 0xB5, /* any(lessThanEqual(.., ..)) */
138 midgard_alu_op_i2f = 0xB8,
139 midgard_alu_op_u2f = 0xBC,
140
141 midgard_alu_op_icsel = 0xC1,
142 midgard_alu_op_fcsel_i = 0xC4,
143 midgard_alu_op_fcsel = 0xC5,
144 midgard_alu_op_fround = 0xC6,
145
146 midgard_alu_op_fatan_pt2 = 0xE8,
147 midgard_alu_op_fpow_pt1 = 0xEC,
148
149 midgard_alu_op_frcp = 0xF0,
150 midgard_alu_op_frsqrt = 0xF2,
151 midgard_alu_op_fsqrt = 0xF3,
152 midgard_alu_op_fexp2 = 0xF4,
153 midgard_alu_op_flog2 = 0xF5,
154 midgard_alu_op_fsin = 0xF6,
155 midgard_alu_op_fcos = 0xF7,
156 midgard_alu_op_fatan2_pt1 = 0xF9,
157 } midgard_alu_op;
158
159 typedef enum {
160 midgard_outmod_none = 0,
161 midgard_outmod_pos = 1,
162 midgard_outmod_int = 2,
163 midgard_outmod_sat = 3
164 } midgard_outmod;
165
166 typedef enum {
167 midgard_reg_mode_8 = 0,
168 midgard_reg_mode_16 = 1,
169 midgard_reg_mode_32 = 2,
170 midgard_reg_mode_64 = 3 /* TODO: verify */
171 } midgard_reg_mode;
172
173 typedef enum {
174 midgard_dest_override_lower = 0,
175 midgard_dest_override_upper = 1,
176 midgard_dest_override_none = 2
177 } midgard_dest_override;
178
179 typedef enum {
180 midgard_int_sign_extend = 0,
181 midgard_int_zero_extend = 1,
182 midgard_int_normal = 2,
183 midgard_int_reserved = 3
184 } midgard_int_mod;
185
186 #define MIDGARD_FLOAT_MOD_ABS (1 << 0)
187 #define MIDGARD_FLOAT_MOD_NEG (1 << 1)
188
189 typedef struct
190 __attribute__((__packed__))
191 {
192 /* Either midgard_int_mod or from midgard_float_mod_*, depending on the
193 * type of op */
194 unsigned mod : 2;
195
196 /* replicate lower half if dest = half, or low/high half selection if
197 * dest = full
198 */
199 bool rep_low : 1;
200 bool rep_high : 1; /* unused if dest = full */
201 bool half : 1; /* only matters if dest = full */
202 unsigned swizzle : 8;
203 }
204 midgard_vector_alu_src;
205
206 typedef struct
207 __attribute__((__packed__))
208 {
209 midgard_alu_op op : 8;
210 midgard_reg_mode reg_mode : 2;
211 unsigned src1 : 13;
212 unsigned src2 : 13;
213 midgard_dest_override dest_override : 2;
214 midgard_outmod outmod : 2;
215 unsigned mask : 8;
216 }
217 midgard_vector_alu;
218
219 typedef struct
220 __attribute__((__packed__))
221 {
222 bool abs : 1;
223 bool negate : 1;
224 bool full : 1; /* 0 = half, 1 = full */
225 unsigned component : 3;
226 }
227 midgard_scalar_alu_src;
228
229 typedef struct
230 __attribute__((__packed__))
231 {
232 midgard_alu_op op : 8;
233 unsigned src1 : 6;
234 unsigned src2 : 11;
235 unsigned unknown : 1;
236 midgard_outmod outmod : 2;
237 bool output_full : 1;
238 unsigned output_component : 3;
239 }
240 midgard_scalar_alu;
241
242 typedef struct
243 __attribute__((__packed__))
244 {
245 unsigned src1_reg : 5;
246 unsigned src2_reg : 5;
247 unsigned out_reg : 5;
248 bool src2_imm : 1;
249 }
250 midgard_reg_info;
251
252 /* In addition to conditional branches and jumps (unconditional branches),
253 * Midgard implements a bit of fixed function functionality used in fragment
254 * shaders via specially crafted branches. These have special branch opcodes,
255 * which perform a fixed-function operation and/or use the results of a
256 * fixed-function operation as the branch condition. */
257
258 typedef enum {
259 /* Regular branches */
260 midgard_jmp_writeout_op_branch_uncond = 1,
261 midgard_jmp_writeout_op_branch_cond = 2,
262
263 /* In a fragment shader, execute a discard_if instruction, with the
264 * corresponding condition code. Terminates the shader, so generally
265 * set the branch target to out of the shader */
266 midgard_jmp_writeout_op_discard = 4,
267
268 /* Branch if the tilebuffer is not yet ready. At the beginning of a
269 * fragment shader that reads from the tile buffer, for instance via
270 * ARM_shader_framebuffer_fetch or EXT_pixel_local_storage, this branch
271 * operation should be used as a loop. An instruction like
272 * "br.tilebuffer.always -1" does the trick, corresponding to
273 * "while(!is_tilebuffer_ready) */
274 midgard_jmp_writeout_op_tilebuffer_pending = 6,
275
276 /* In a fragment shader, try to write out the value pushed to r0 to the
277 * tilebuffer, subject to unknown state in r1.z and r1.w. If this
278 * succeeds, the shader terminates. If it fails, it branches to the
279 * specified branch target. Generally, this should be used in a loop to
280 * itself, acting as "do { write(r0); } while(!write_successful);" */
281 midgard_jmp_writeout_op_writeout = 7,
282 } midgard_jmp_writeout_op;
283
284 typedef enum {
285 midgard_condition_write0 = 0,
286
287 /* These condition codes denote a conditional branch on FALSE and on
288 * TRUE respectively */
289 midgard_condition_false = 1,
290 midgard_condition_true = 2,
291
292 /* This condition code always branches. For a pure branch, the
293 * unconditional branch coding should be used instead, but for
294 * fixed-function branch opcodes, this is still useful */
295 midgard_condition_always = 3,
296 } midgard_condition;
297
298 typedef struct
299 __attribute__((__packed__))
300 {
301 midgard_jmp_writeout_op op : 3; /* == branch_uncond */
302 unsigned dest_tag : 4; /* tag of branch destination */
303 unsigned unknown : 2;
304 int offset : 7;
305 }
306 midgard_branch_uncond;
307
308 typedef struct
309 __attribute__((__packed__))
310 {
311 midgard_jmp_writeout_op op : 3; /* == branch_cond */
312 unsigned dest_tag : 4; /* tag of branch destination */
313 int offset : 7;
314 midgard_condition cond : 2;
315 }
316 midgard_branch_cond;
317
318 typedef struct
319 __attribute__((__packed__))
320 {
321 midgard_jmp_writeout_op op : 3; /* == branch_cond */
322 unsigned dest_tag : 4; /* tag of branch destination */
323 unsigned unknown : 2;
324 signed offset : 23;
325 unsigned cond : 16;
326 }
327 midgard_branch_extended;
328
329 typedef struct
330 __attribute__((__packed__))
331 {
332 midgard_jmp_writeout_op op : 3; /* == writeout */
333 unsigned unknown : 13;
334 }
335 midgard_writeout;
336
337 /*
338 * Load/store words
339 */
340
341 typedef enum {
342 midgard_op_ld_st_noop = 0x03,
343
344 /* Unclear why this is on the L/S unit, but (with an address of 0,
345 * appropriate swizzle, magic constant 0x24, and xy mask?) moves fp32 cube
346 * map coordinates in r27 to its cube map texture coordinate
347 * destination (e.g r29). 0x4 magic for loading from fp16 instead */
348
349 midgard_op_store_cubemap_coords = 0x0E,
350
351 midgard_op_load_attr_16 = 0x95,
352 midgard_op_load_attr_32 = 0x94,
353 midgard_op_load_vary_16 = 0x99,
354 midgard_op_load_vary_32 = 0x98,
355 midgard_op_load_color_buffer_16 = 0x9D,
356 midgard_op_load_color_buffer_8 = 0xBA,
357 midgard_op_load_uniform_16 = 0xAC,
358 midgard_op_load_uniform_32 = 0xB0,
359 midgard_op_store_vary_16 = 0xD5,
360 midgard_op_store_vary_32 = 0xD4
361 } midgard_load_store_op;
362
363 typedef enum {
364 midgard_interp_centroid = 1,
365 midgard_interp_default = 2
366 } midgard_interpolation;
367
368 typedef struct
369 __attribute__((__packed__))
370 {
371 unsigned zero1 : 4; /* Always zero */
372
373 /* Varying qualifiers, zero if not a varying */
374 unsigned flat : 1;
375 unsigned is_varying : 1; /* Always one for varying, but maybe something else? */
376 midgard_interpolation interpolation : 2;
377
378 unsigned zero2 : 2; /* Always zero */
379 }
380 midgard_varying_parameter;
381
382 typedef struct
383 __attribute__((__packed__))
384 {
385 midgard_load_store_op op : 8;
386 unsigned reg : 5;
387 unsigned mask : 4;
388 unsigned swizzle : 8;
389 unsigned unknown : 16;
390
391 unsigned varying_parameters : 10;
392
393 unsigned address : 9;
394 }
395 midgard_load_store_word;
396
397 typedef struct
398 __attribute__((__packed__))
399 {
400 unsigned type : 4;
401 unsigned next_type : 4;
402 uint64_t word1 : 60;
403 uint64_t word2 : 60;
404 }
405 midgard_load_store;
406
407 /* Texture pipeline results are in r28-r29 */
408 #define REG_TEX_BASE 28
409
410 /* Texture opcodes... maybe? */
411 #define TEXTURE_OP_NORMAL 0x11
412 #define TEXTURE_OP_TEXEL_FETCH 0x14
413
414 /* Texture format types, found in format */
415 #define TEXTURE_CUBE 0x00
416 #define TEXTURE_2D 0x02
417 #define TEXTURE_3D 0x03
418
419 typedef struct
420 __attribute__((__packed__))
421 {
422 unsigned type : 4;
423 unsigned next_type : 4;
424
425 unsigned op : 6;
426 unsigned shadow : 1;
427 unsigned unknown3 : 1;
428
429 /* A little obscure, but last is set for the last texture operation in
430 * a shader. cont appears to just be last's opposite (?). Yeah, I know,
431 * kind of funky.. BiOpen thinks it could do with memory hinting, or
432 * tile locking? */
433
434 unsigned cont : 1;
435 unsigned last : 1;
436
437 unsigned format : 5;
438 unsigned has_offset : 1;
439
440 /* Like in Bifrost */
441 unsigned filter : 1;
442
443 unsigned in_reg_select : 1;
444 unsigned in_reg_upper : 1;
445
446 unsigned in_reg_swizzle_left : 2;
447 unsigned in_reg_swizzle_right : 2;
448
449 unsigned unknown1 : 2;
450
451 unsigned unknown8 : 4;
452
453 unsigned out_full : 1;
454
455 /* Always 1 afaict... */
456 unsigned unknown7 : 2;
457
458 unsigned out_reg_select : 1;
459 unsigned out_upper : 1;
460
461 unsigned mask : 4;
462
463 unsigned unknown2 : 2;
464
465 unsigned swizzle : 8;
466 unsigned unknown4 : 8;
467
468 unsigned unknownA : 4;
469
470 unsigned offset_unknown1 : 1;
471 unsigned offset_reg_select : 1;
472 unsigned offset_reg_upper : 1;
473 unsigned offset_unknown4 : 1;
474 unsigned offset_unknown5 : 1;
475 unsigned offset_unknown6 : 1;
476 unsigned offset_unknown7 : 1;
477 unsigned offset_unknown8 : 1;
478 unsigned offset_unknown9 : 1;
479
480 unsigned unknownB : 3;
481
482 /* Texture bias or LOD, depending on whether it is executed in a
483 * fragment/vertex shader respectively. Compute as int(2^8 * biasf).
484 *
485 * For texel fetch, this is the LOD as is. */
486 unsigned bias : 8;
487
488 unsigned unknown9 : 8;
489
490 unsigned texture_handle : 16;
491 unsigned sampler_handle : 16;
492 }
493 midgard_texture_word;
494
495 static char *load_store_opcode_names[256] = {
496 [midgard_op_store_cubemap_coords] = "st_cubemap_coords",
497 [midgard_op_load_attr_16] = "ld_attr_16",
498 [midgard_op_load_attr_32] = "ld_attr_32",
499 [midgard_op_load_vary_16] = "ld_vary_16",
500 [midgard_op_load_vary_32] = "ld_vary_32",
501 [midgard_op_load_uniform_16] = "ld_uniform_16",
502 [midgard_op_load_uniform_32] = "ld_uniform_32",
503 [midgard_op_load_color_buffer_8] = "ld_color_buffer_8",
504 [midgard_op_load_color_buffer_16] = "ld_color_buffer_16",
505 [midgard_op_store_vary_16] = "st_vary_16",
506 [midgard_op_store_vary_32] = "st_vary_32"
507 };
508
509 #endif