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